AI-driven assessment of forest dynamics and ecosystem services in response to land use change: current progress and future perspectives

1 Introduction

The capacity of forests to regulate Earth's climate, sustain biodiversity, and support human livelihoods depends fundamentally on their spatial extent, structural integrity, and functional connectivity across landscapes. Covering approximately 4.06 billion hectares, or 31% of the global land surface (FAO, 2020), forests store roughly 45% of all terrestrial carbon and generate around half of terrestrial net primary productivity (Bonan, 2008). They regulate hydrological processes through canopy interception, soil water retention, and evapotranspiration, sustaining water supply and reducing flood risks at the watershed scale (Bonan, 2008). As habitats for the majority of terrestrial species, forests also underpin biodiversity at every trophic level; mixed-species forests, for instance, produce on average 26.5% more biomass than the mean of their component monocultures, illustrating the functional significance of tree diversity for ecosystem productivity (Brockerhoff et al., 2017). The total annual value of global ecosystem services has been estimated at 125 trillion USD (2007 dollars), with forests and wetlands contributing a substantial share (Costanza et al., 2014, 1997).

Yet these services are being eroded at a pace that outstrips our capacity to monitor and quantify the damage. The Food and Agriculture Organization reports that 420 million hectares of forest have been cleared since 1990, with deforestation continuing at approximately 10 million hectares per year during 2015 to 2020 (FAO, 2020). Between 2001 and 2015, 27% of all global forest loss was attributable to commodity-driven permanent deforestation for beef, soy, palm oil, and wood fiber, at a stable rate of roughly 5 million hectares per year that has shown no decline despite corporate zero-deforestation pledges (Curtis et al., 2018). The resulting economic consequences are severe: land use change alone caused an estimated annual loss of 4.3 to 20.2 trillion USD in ecosystem service value between 1997 and 2011, driven primarily by the conversion of tropical forests and wetlands (Costanza et al., 2014). Beyond outright clearing, urbanization, agricultural intensification, and policy-driven land conversion further reshape forested landscapes through fragmentation, edge proliferation, and habitat isolation (Haddad et al., 2015). At the same time, these global patterns resist simple narratives. Satellite observations spanning 1982 to 2016 reveal a net increase of 2.24 million km² in global tree canopy cover (+7.1%), as temperate and boreal gains in China, Europe, and Russia have more than offset tropical losses in Brazil, Argentina, and Southeast Asia (Song et al., 2018). Deforestation, reforestation, and degradation occur simultaneously across different regions and climatic zones, creating a spatially heterogeneous mosaic that demands analytical tools capable of resolving land use transitions at fine scales and tracing their cascading effects on ecosystem service provision.

Traditional approaches to forest ecosystem service assessment have generated foundational knowledge but cannot fully meet this demand. Field-based inventories are constrained by spatial coverage and revisit frequency, making continuous large-scale monitoring impractical. Process-based simulation platforms such as InVEST have enabled spatially explicit modeling of multiple services including water quality regulation, coastal protection, and crop pollination, and have projected that up to 5 billion people could face heightened risks from ecosystem service decline by 2050, particularly in Africa and South Asia (Chaplin-Kramer et al., 2019; Nelson et al., 2009). However, these models require extensive parameterization, simplify the spatial heterogeneity of real-world landscapes, and remain constrained by the resolution and accuracy of input land cover data (Chaplin-Kramer et al., 2019). Three specific limitations stand out: the inability to process the volume and variety of satellite data now available at the speed required for operational monitoring, the difficulty of capturing nonlinear and scale-dependent relationships between land use configurations and service outcomes, and the lack of automated, repeatable workflows suitable for large-area assessment (Yuan et al., 2020; Ma et al., 2019).

These limitations align closely with the capabilities that artificial intelligence now offers. The rapid development of deep learning, together with the expanding archive of multi-source satellite observations, has created a new generation of tools that address each of these bottlenecks. Convolutional neural networks have become the dominant architecture for spatial feature extraction from remote sensing imagery (Zhu et al., 2017; Ma et al., 2019), with specialized variants such as U-Net (Ronneberger et al., 2015) and DeepLab (Chen et al., 2018) enabling pixel-level semantic segmentation at scales from individual tree crowns to continental land cover maps. Recurrent architectures such as Long Short-Term Memory networks capture temporal dependencies in phenological and disturbance trajectories (Yuan et al., 2020). More recently, Vision Transformers have introduced the ability to process global contextual information across entire image scenes, further expanding the methodological repertoire available for remote sensing analysis. A meta-analysis of 176 deep learning studies in remote sensing found that these models achieve median classification accuracies exceeding 91% in land use and land cover mapping, consistently surpassing conventional machine learning classifiers (Ma et al., 2019). Concurrently, the proliferation of satellite missions, from Landsat's five-decade archive to Sentinel-2's five-day revisit cycle and GEDI's spaceborne lidar measurements of canopy structure, has built a multi-modal observational infrastructure that, combined with cloud computing platforms such as Google Earth Engine, supports petabyte-scale geospatial analysis. The convergence of these methodological and observational advances has opened the possibility of monitoring forest dynamics and quantifying ecosystem services continuously, at high resolution, and across entire biomes.

Despite this rapid progress, the literature at the intersection of AI, remote sensing, and forest ecology remains fragmented along disciplinary lines. Reviews have addressed deep learning architectures for remote sensing image processing (Zhu et al., 2017; Ma et al., 2019; Yuan et al., 2020), machine learning applications in sustainable forest management including fire detection, pest monitoring, and carbon estimation (Espíndola et al., 2025), and AI-powered forestry tools for species identification and climate adaptation (Xu and Jiang, 2025). Each of these contributions illuminates one facet of the field, but none follows the full causal chain from land use change drivers, through forest structural and functional responses, to the resulting shifts in ecosystem service supply. Nor do they evaluate how AI and geospatial technologies perform at each link of this chain or how the outputs of one stage feed into the next. This gap matters for practice: approximately 76% of existing AI-forestry studies optimize individual technical components without addressing system-level integration (Xu and Jiang, 2025), leaving land managers and policymakers without a coherent picture of how these tools can be assembled into operational decision-support systems.

This review aims to address this gap at the conceptual and organizational level rather than the technical level: it does not build or computationally test an integrated system, but instead synthesizes the literature into an original four-tier framework that traces the causal chain from land use change drivers through forest structural and functional responses to ecosystem service outcomes, making explicit how the outputs of one tier, such as land cover change maps, function as inputs to the next, such as ecosystem service quantification, a linkage that existing reviews addressing individual tiers in isolation do not make explicit. We comprehensively review research published between 2016 and 2026 on AI and geospatial approaches to forest dynamics and ecosystem services in the context of land use change. To identify relevant studies, we searched Web of Science, Scopus, and Google Scholar using terms combining AI and machine learning methods, forest ecosystems, and ecosystem services or land use change. Title and abstract screening against the eligibility criteria below yielded an initial pool of 65 candidate core studies; a subsequent full-text re-verification, conducted as part of the quality appraisal reported in Section 2.3, confirmed that 27 peer-reviewed articles published between 2016 and 2026 directly report a quantified AI or machine learning method performance metric relevant to forest monitoring, assessment, prediction, or optimization, and these form the core evidentiary basis of this synthesis. An additional 70 references, including foundational methodological works published before 2016, institutional reports, systematic reviews and surveys of adjacent literatures, preprints, and satellite mission documentation, are cited to provide theoretical context, satellite platform specifications, and emerging perspectives but were not part of the core systematic sample; the complete list of these 27 core studies is provided in Section 2.3. We characterize this as a narrative review with an explicit and reproducible search strategy for study identification, rather than a systematic review in the PRISMA sense or a scoping review, because the objective is to synthesize evidence into an original conceptual framework and to identify research gaps and future directions, rather than to exhaustively catalog all studies meeting formal eligibility criteria or to quantitatively pool effect sizes across studies. The review protocol was not pre-registered on a platform such as PROSPERO or the Open Science Framework; we note this transparently as a limitation, since pre-registration would have further reduced the potential for selection bias in study identification and interpretation.

The synthesis is organized around an original four-tier analytical framework that reflects increasing levels of complexity and progressively closer engagement with management and policy: monitoring (detecting what has changed), assessment (quantifying the ecological and economic consequences), prediction (projecting future trajectories under alternative scenarios), and optimization (identifying management strategies that maximize desired outcomes). Beyond synthesizing current progress, we identify critical knowledge gaps and propose a phased research agenda with near-term, medium-term, and long-term priorities.

1.1 Aim, objectives, and research questions

The overall aim of this review is to synthesize the fragmented literature at the intersection of artificial intelligence, geospatial technologies, and forest ecosystem services into an original four-tier analytical framework that traces the causal chain from land use change drivers through forest structural and functional responses to ecosystem service outcomes, characterizing how AI and geospatial methods perform at each tier and how the outputs of one tier feed into the next. Three specific objectives follow from this aim: (1) to catalog and compare the reported performance, boundary conditions, and validation practices of AI and machine learning methods applied at each of the four tiers, monitoring, assessment, prediction, and optimization; (2) to identify the methodological limitations, particularly in model explainability, uncertainty quantification, and cross-domain transferability, that currently constrain the operational and policy relevance of these methods; and (3) to propose a phased research agenda, spanning near-term, medium-term, and long-term horizons, that addresses these limitations. These objectives are addressed through three research questions: RQ1, how does the performance of AI and machine learning methods vary across the monitoring, assessment, prediction, and optimization tiers, and under what conditions of data availability, geographic transfer, and validation design do reported advantages over conventional methods hold; RQ2, what methodological gaps most limit the reliability and policy applicability of AI-derived forest ecosystem service assessments; and RQ3, what research priorities across near-term, medium-term, and long-term horizons would address these gaps and advance integrated forest ecosystem intelligence.

The conceptual framework guiding this review is illustrated in Figure 1.

The remainder of this review is organized as follows. Section 2 details the search strategy and eligibility criteria underlying study identification. Section 3 examines multi-source remote sensing data systems for forest and land use monitoring. Section 4, the core of the review, evaluates AI methods across the monitoring and assessment tiers, covering land cover classification, ecosystem service quantification, and forest fragmentation analysis. Section 5 addresses scenario simulation and predictive modeling under different land use and climate regimes. Section 6 discusses the role of GIS, geovisualization, and cloud computing platforms. Section 7 identifies key challenges at the data, method, and application levels. Section 8 outlines future directions and a structured research agenda. Section 9 concludes with a synthesis of findings and implications. The synthesis is motivated by the monitoring and reporting demands of global sustainability frameworks, including SDG 13 (climate action), SDG 15 (life on land), the Kunming-Montreal Global Biodiversity Framework, and the Nationally Determined Contributions under the Paris Agreement, all of which call for spatially explicit, temporally dynamic, and scientifically robust assessments of forest ecosystem services.

2 Materials and methods

2.1 Search strategy

Searches were executed in early February 2026, several months prior to manuscript submission, across three databases (Web of Science Core Collection, Scopus, and Google Scholar) using Boolean logic combining three concept categories: an AI/methods category, a forest category, and an ecosystem service/land use change category, with OR logic within each category and AND logic between categories. In Web of Science, the search string was TS = [(“artificial intelligence” OR “machine learning” OR “deep learning” OR “convolutional neural network” OR “random forest” OR “reinforcement learning”) AND (“forest” OR “land cover” OR “land use”) AND (“ecosystem service” OR “land use change” OR “biomass” OR “biodiversity” OR “water yield”)], restricted to publication years 2016 to 2026. In Scopus, the equivalent string was TITLE-ABS-KEY[(“artificial intelligence” OR “machine learning” OR “deep learning” OR “convolutional neural network” OR “random forest” OR “reinforcement learning”) AND (“forest” OR “land cover” OR “land use”) AND (“ecosystem service” OR “land use change” OR “biomass” OR “biodiversity” OR “water yield”)] AND PUBYEAR > 2015 AND PUBYEAR < 2027. In Google Scholar, the query was (“machine learning” OR “deep learning”) (“forest” OR “land cover”) (“ecosystem service” OR “land use change” OR “biomass” OR “biodiversity”), restricted to 2016 to 2026. Given the substantial overlap expected across the three databases and the absence of structured export and deduplication tools for Google Scholar at the scale returned, no attempt was made to deduplicate and screen the combined output of these three searches on a record-by-record PRISMA basis; this is consistent with this review's stated identity as a narrative review with an explicit and reproducible search strategy rather than a systematic review in the PRISMA sense (Section 1). Instead, records were identified and screened by title and abstract against the eligibility criteria specified in Section 2.2, followed by full-text re-verification as part of the quality appraisal reported in Section 2.3, yielding the final core sample of 27 studies discussed therein.

2.2 Eligibility criteria

Records were included in the core evidentiary sample if they satisfied all of the following: (1) published in English; (2) published between 2016 and 2026; (3) applied an artificial intelligence or machine learning method (including but not limited to deep learning, random forest, support vector machines, or reinforcement learning) as an original empirical contribution, rather than reviewing, surveying, or conceptually discussing such methods; (4) the forest-related, water-related, or biodiversity-related system under study is directly relevant to forest ecosystem monitoring, assessment, prediction, or optimization as defined by this review's four-tier framework; and (5) reported at least one directly quantified performance metric for the AI/ML method (e.g., overall accuracy, F1-score, R², AUC, Kappa, or a comparable statistic) rather than discussing the method only in general or theoretical terms. Two of the 65 candidate core studies described below, Chen et al. (2015) and Phillips et al. (2006), were identified through backward citation tracing during title-and-abstract screening rather than surfaced directly by the search strings in Section 2.1, which were restricted to publication years 2016–2026; both studies predate this window and were retained through full-text re-verification as an explicit exception to criterion (2), because they report the original quantified performance metrics for methods, the POK classification approach and the MaxEnt algorithm respectively, that are used as direct baseline comparators for several core studies published within the review window (Sections 3.2, 4.2.3); excluding their own reported metrics would have left these baseline comparisons without a primary source in the core evidentiary sample, so both are counted among the final core sample of 27 studies reported below. Peer-reviewed journal articles constituted the majority of included records; a small number of conference proceedings and preprints were also included where they represented original empirical work or benchmarks of substantive relevance not yet available in peer-reviewed form, consistent with the treatment of several such sources cited elsewhere in this review. Records were excluded if they were not in English, fell outside the 2016 to 2026 window, were themselves systematic reviews, surveys, or meta-analyses rather than original empirical studies, did not apply an AI/ML method as an original contribution, or discussed forest-relevant content without reporting a directly quantified AI/ML performance metric. An initial title-and-abstract screening pass identified 65 candidate core studies; full-text re-verification against criteria (3) and (5) above, conducted as part of the quality appraisal reported in Section 2.3, found that 38 of these initially screened records were, on full-text inspection, systematic reviews of adjacent literatures, studies in which forest-relevant content appeared only as contextual framing rather than as a directly quantified result, or studies whose primary quantitative contribution lay outside forest-specific AI/ML performance. These 38 records were reclassified as contextual citations, yielding a final core sample of 27 studies. The complete set of 70 contextual references cited throughout this review comprises these 38 reclassified records together with a further set of references that were never part of the 65-study candidate pool, including foundational methodological works published before 2016, institutional reports, systematic reviews and surveys of adjacent literatures, preprints, and satellite mission documentation; several of the latter, discussed in Section 3.3, were included specifically because they reflect satellite missions launched after the original search window described in Section 2.1.

2.3 Quality appraisal

To assess methodological reporting transparency across the 27 core studies, we developed a five-criterion Methodological Reporting Transparency Checklist, informed by established quality appraisal frameworks such as the Mixed Methods Appraisal Tool and ROBIS but adapted to the reporting conventions of applied AI/ML studies in remote sensing and forest ecology, for which existing appraisal tools were not designed. The five criteria were: (1) Validation Strategy (VS), whether the study used an independent test set, random cross-validation, or spatial/spatiotemporal blocking, or did not state its validation approach; (2) Sample Size (SS), whether the training sample size was explicitly reported; (3) Code/Data availability (C/D), whether code or data underlying the reported results were made publicly accessible; (4) Scope, whether validation was conducted at a single site or across multiple sites; and (5) Uncertainty Quantification (UQ), whether the study reported any formal measure of prediction uncertainty. Each of the 27 core studies was independently scored against these five criteria by cross-referencing the original publications; the complete scored results are reported in Table 1. Across the sample, 18.5% (5/27) employed an independent test set and 25.9% (7/27) employed spatial or spatiotemporal blocking (44.4% combined), 48.1% reported a specific training sample size, 44.4% provided publicly accessible code or data, 48.1% validated across multiple sites, and only 3.7% (1/27) reported any measure of prediction uncertainty. This checklist does not constitute a formal risk-of-bias assessment in the systematic-review sense, since criteria such as validation strategy and uncertainty quantification reflect reporting transparency rather than an integrated bias-risk judgment; it instead provides readers with an explicit, verifiable basis for gauging the methodological rigor represented across the core evidentiary sample underlying this narrative synthesis.

Table 1. Methodological reporting transparency checklist for the 27 core studies.

VS, Validation Strategy (I = independent test set, R = random cross-validation, S = spatial/spatiotemporal blocking, N = not stated); SS, training Sample Size reported; C/D, Code/Data availability; Scope, Single- or Multi-site; UQ, Uncertainty Quantification reported. Across the 27 core studies, 18.5% (5/27) employed an independent test set and 25.9% (7/27) employed spatial or spatiotemporal blocking (44.4% combined), 48.1% reported a specific training sample size, 44.4% provided publicly accessible code or data, 48.1% validated across multiple sites, and only 3.7% (1/27) reported any measure of prediction uncertainty.

3 Multi-source remote sensing data systems for forest and land use monitoring

The capacity to monitor forest dynamics and land use transitions at scale depends fundamentally on the availability, continuity, and complementarity of remote sensing data. Over the past two decades, the observational infrastructure for forest research has expanded from a small number of moderate-resolution optical sensors to a multi-modal constellation spanning optical, radar, and lidar domains, with spatial resolutions from sub-meter to kilometer scales and temporal archives extending back nearly five decades. This section examines the principal data sources used in AI-driven forest ecosystem service research, organized along three axes: from coarse to fine spatial resolution, from two-dimensional spectral measurements to three-dimensional structural characterization, and from established platforms to emerging technologies.

3.1 Optical remote sensing: Landsat, sentinel-2, and MODIS

The Landsat program provides the longest continuous moderate-resolution satellite record of the Earth's land surface, spanning over five decades since the launch of Landsat-1 in 1972. The archive has evolved through successive sensor generations, from the 60 m, 4-band Multispectral Scanner (MSS) to the 30 m, 12-bit Operational Land Imager (OLI) aboard Landsat-8, which achieves signal-to-noise ratios up to eight times higher than its predecessors (Wulder et al., 2019). The adoption of a free and open data policy in 2008 transformed the utility of this archive, with monthly downloads exceeding one million scenes as of 2019. Landsat's 30 m resolution and 16-day revisit cycle have enabled foundational global assessments of forest cover change, most notably the dataset by Hansen et al. (2013), which documented 2.3 million km² of global forest loss between 2000 and 2012 at 30 m resolution using Google cloud computing infrastructure. The emergence of time series algorithms such as LandTrendr and CCDC has further shifted the analytical paradigm from scene-based classification to pixel-level temporal trajectory analysis (Wulder et al., 2019). However, the 16-day revisit cycle limits observation frequency in persistently cloudy regions, constraining the detection of rapid forest disturbances.

The Sentinel-2 mission, launched by the European Space Agency, addresses this temporal limitation. Its Multi-Spectral Instrument (MSI) acquires 13 spectral bands at 10 m (visible and near-infrared), 20 m (red edge and shortwave infrared), and 60 m (atmospheric correction) resolution across a 290 km swath width, achieving a 5-day revisit cycle at the equator with twin satellites (Drusch et al., 2012). The inclusion of three red-edge bands at 705, 740, and 783 nm is particularly valuable for vegetation monitoring, as these wavelengths are sensitive to chlorophyll content and canopy structure. When combined with Landsat, Sentinel-2 forms a virtual constellation that achieves an average global revisit of approximately 2.9 days (Wulder et al., 2019), creating dense time series that are well suited to phenological analysis and near-real-time forest change detection.

The Moderate Resolution Imaging Spectroradiometer (MODIS) complements these medium-resolution sensors by providing daily global coverage at 250 m to 1 km resolution. Among its product suite, the MOD17A3HGF annual net primary productivity (NPP) product is of particular relevance to forest ecosystem service research. The underlying algorithm calculates gross primary production (GPP) as the product of light use efficiency, the fraction of absorbed photosynthetically active radiation (FPAR), and incident PAR, with NPP derived by subtracting autotrophic respiration components (Running et al., 2004). Inputs include MODIS FPAR/LAI composites at 1 km resolution and global meteorological reanalysis data. The product estimated global annual NPP at 57.7 Pg C in 2001 and 55.5 Pg C in 2002, with water limitation controlling productivity across 40% of global land area, temperature across 33%, and radiation across 27% (Running et al., 2004). While the coarse resolution of meteorological inputs (1° × 1.25°) introduces uncertainty at site level, particularly through underestimation of vapor pressure deficit in summer, the MOD17A3HGF product remains the most widely used satellite-derived NPP dataset for large-scale forest carbon assessments. Recent applications have demonstrated its value for quantifying NPP spatiotemporal dynamics at the watershed scale in relation to climate variability and urbanization (Li et al., 2026).

3.2 Land cover products

Satellite-derived land cover products provide the spatial framework within which forest dynamics and ecosystem service changes are analyzed. At the global scale, GlobeLand30 offers 30 m land cover maps for 2000, 2010, and 2020, classified into 10 land cover types using a pixel-object-knowledge (POK) approach applied to over 10,000 Landsat scenes. The most recent version (V2020) achieved an overall accuracy of approximately 85.72% (Kappa = 0.82). In regional comparisons, the POK method outperformed fully automated classifiers, achieving higher accuracy than SVM (73.4%) and Random Forest (71.6%) on the same test areas (Chen et al., 2015). At the national scale, the China Land Cover Dataset (CLCD) provides annual 30 m land cover maps from 1990 to 2019, derived from 335,709 Landsat scenes processed on Google Earth Engine using Random Forest classification with 36 input features spanning spectral, phenological, topographic, and geographic dimensions (Yang and Huang, 2021). CLCD's original authors reported a self-validated overall accuracy of 79.31%; independent validation using the Global Land Cover Validation Sample Set (GLCVSS), a more conservative benchmark not tied to the product's own training and validation samples, yielded an overall accuracy of 65.64%, which nonetheless still outperformed GlobeLand30 (63.12%) and other global products under the same independent framework. CLCD documented a 4.34% increase in forest area across China over the study period, alongside a 148.71% expansion of impervious surfaces, providing a quantitative basis for analyzing land use change impacts on forest ecosystem services at the national scale.

3.3 Active remote sensing: spaceborne lidar and SAR

Optical sensors capture the horizontal extent of forests but cannot directly measure their vertical structure, a critical determinant of carbon storage, habitat complexity, and ecosystem function. The Global Ecosystem Dynamics Investigation (GEDI), a spaceborne waveform lidar operating from the International Space Station, was designed specifically to fill this gap. GEDI employs three lasers to produce eight parallel ground tracks of 25 m footprints spaced 60 m along-track, acquiring over 10 billion waveforms across latitudes from 51.6°S to 51.6°N (Dubayah et al., 2020). Its product suite includes footprint-level canopy height metrics (L2A), canopy cover and leaf area index (L2B), and gridded aboveground biomass density at 1 km resolution (L4B), calibrated against 5,646 field plots across 208 sites in 27 countries. The complementary ICESat-2 mission extends lidar coverage to 88° latitude using a photon-counting approach at 532 nm, with a smaller footprint (17 m) and much finer along-track sampling (0.7 m) than GEDI, though its primary design target is the cryosphere rather than vegetation (Markus et al., 2017). Together, GEDI and ICESat-2 provide complementary three-dimensional measurements of forest structure: GEDI offers detailed waveform-derived biomass estimates across tropical and temperate zones, while ICESat-2 extends canopy height coverage to boreal forests at high latitudes.

Synthetic aperture radar (SAR) provides a further dimension of all-weather, day-and-night observation capability. The Sentinel-1 C-band SAR mission acquires dual-polarization (VV+VH or HH+HV) imagery in interferometric wide-swath mode at 20 m by 5 m resolution across a 250 km swath, with a 6-day revisit cycle using twin satellites (Torres et al., 2012). Forest and non-forest classification from Sentinel-1 alone has achieved balanced accuracies ranging from approximately 80% to 93% across six study sites spanning boreal, temperate, and tropical biomes, with multi-temporal statistics outperforming single-scene classification (Hansen et al., 2020), and average overall accuracies of 88.2% at 10 m resolution when validated against national-scale forest maps from twelve European countries (Dostálová et al., 2021). The all-weather, day-and-night observation capability of C-band SAR makes Sentinel-1 particularly valuable in regions where persistent cloud cover limits the utility of optical sensors, such as tropical forest zones.

The saturation of shorter-wavelength SAR at high biomass levels, discussed further in Section 4.2.1, motivated the development of longer-wavelength spaceborne radar. ESA's Biomass mission, the seventh Earth Explorer and the first spaceborne P-band (435 MHz) synthetic aperture radar, launched on 29 April 2025 (European Space Agency, 2026a). Following in-orbit commissioning, ESA formally announced the transition to scientific operations in January 2026, at which point open Level-1 (and subsequently Level-2) data products were released (European Space Agency, 2026b).

Operating from a 666 km sun-synchronous orbit with a fully polarimetric 12 m deployable mesh reflector, Biomass is designed to penetrate more deeply into forest canopies and exhibit enhanced sensitivity to woody structural elements such as trunks and large branches relative to shorter C-band and optical wavelengths, though radar backscatter at P-band remains additionally influenced by soil moisture, incidence angle, and polarization rather than providing a direct measurement of trunk structure alone (European Space Agency, 2026c). The mission's standard geophysical products target global forest aboveground biomass and height maps at approximately 200 m resolution and severe forest disturbance maps at approximately 50 m resolution, generated through an initial tomographic global coverage phase followed by repeated interferometric global coverage cycles for the remainder of the mission (European Space Agency, 2026b). Coverage between 75°N and 56°S is additionally constrained by radio-frequency coordination requirements that exclude operation within line-of-sight of United States Department of Defense space tracking radars, mainly affecting parts of Europe and North America; ESA has indicated that this restriction is not expected to affect the mission's core objectives, as the primary target forest regions lie outside these excluded areas (European Space Agency, 2026c).

The NASA-ISRO Synthetic Aperture Radar (NISAR) mission, launched on 30 July 2025, transitioned from commissioning to routine science operations around the turn of 2025 to 2026, as the first dual-frequency (L-band, 24 cm; S-band, 10 cm) spaceborne SAR, using a SweepSAR wide-swath acquisition mode with an exact 12-day repeat cycle (National Aeronautics and Space Administration, 2026). The mission's L-band radar provides global systematic land and vegetation coverage, whereas its S-band radar is directed primarily at India and a set of global calibration and validation sites rather than achieving worldwide coverage (National Aeronautics and Space Administration, 2025). While NISAR's raw SAR imagery achieves approximately 3 to 10 m resolution depending on acquisition mode, its official Level-2/3 ecosystem science products are specified at the hectare scale rather than at this native imaging resolution: NASA's baseline science requirement, for a minimum three-year mission (with a two-year threshold requirement), calls for global annual mapping of aboveground woody biomass disturbance and recovery at the hectare scale, with an accuracy of 20 Mg/ha or better specifically for areas of woody biomass below 100 Mg/ha (National Aeronautics and Space Administration, 2025). This specification indicates that NISAR's biomass mandate is oriented toward lower- to moderate-biomass forests rather than specifically toward resolving saturation at the highest biomass densities.

Biomass saturation itself should not be treated as a single, universal physical threshold: reported saturation points vary substantially by sensor, wavelength, forest type, moisture conditions, and retrieval algorithm, and the approximately 80 Mg/ha figure discussed in Section 4.2.1 reflects saturation behavior reported in specific study contexts rather than a fixed ceiling common to all remote sensing modalities. Given this, Biomass's P-band instrument is more accurately described as purpose-built for high-biomass forest structural retrieval, whereas NISAR's L-band and S-band radars address forest disturbance, structure, and lower-to-moderate biomass estimation, complementing rather than duplicating this high-biomass capability. Given the 2025 launch dates of both missions and their transition to routine science operations only around the turn of 2025 to 2026, forest-specific AI/ML applications using either mission's data had not yet appeared in the peer-reviewed literature captured by this review's search window as of early February 2026 (Section 2.1).

Despite these advances, no single sensor system can simultaneously provide the spatial resolution, temporal frequency, spectral coverage, and three-dimensional structural information needed for comprehensive forest ecosystem service assessment. This limitation motivates the integration of multiple data sources and the development of fusion methodologies.

The key characteristics of the principal remote sensing data sources discussed above are summarized in Table 2.

Table 2. Comparison of principal remote sensing data sources for forest dynamics and ecosystem service monitoring.

OA, Overall Accuracy; GLCVSS, Global Land Cover Validation Sample Set. The two accuracy figures reported for CLCD refer to different evaluation frameworks: 79.31% is the self-validated OA reported by the original authors, whereas 65.64% is the independent OA assessed against the GLCVSS global reference samples.

3.4 Emerging data sources and multi-source data fusion

Unmanned aerial vehicles (UAVs) have emerged as a complementary platform for fine-scale forest research, carrying sensors ranging from RGB and multispectral cameras to lightweight lidar systems and thermal imagers (Colomina and Molina, 2014). UAV platforms deliver centimeter-level spatial resolution suitable for individual tree crown delineation, canopy gap analysis, and local biodiversity assessment, though their utility is limited to small areas by short flight endurance, airspace regulations, and wind sensitivity.

The proliferation of these diverse data sources creates both opportunities and challenges for integrated analysis. Multi-source data fusion methods aim to combine the complementary strengths of different sensors, and have been systematically classified into pixel-level, feature-level, and decision-level approaches (Samadzadegan et al., 2025). Among spatiotemporal fusion algorithms, STARFM and its enhanced variant ESTARFM generate synthetic high-resolution time series by blending Landsat's spatial detail with MODIS's temporal frequency, though their performance degrades in heterogeneous landscapes where the assumption of uniform reflectance change within coarse pixels breaks down (Zhu et al., 2018). More recent methods such as FSDAF combine spectral unmixing with spatial interpolation to handle abrupt land cover transitions. A critical constraint identified across fusion approaches is the ground sampling distance ratio between input sensors: standard fusion performs reliably at ratios of 1:2 to 1:5, while multi-sensor fusion remains acceptable at 1:10 to 1:20, with larger ratios introducing spectral distortion and spatial detail loss (Samadzadegan et al., 2025).

The computational demands of processing these increasingly voluminous and heterogeneous datasets, combined with the limitations of physics-based fusion algorithms in heterogeneous landscapes, point to a growing role for AI-driven approaches to data integration. Deep learning methods for spatiotemporal fusion, though still emerging when reviewed by Zhu et al. (2018), have since matured considerably, as discussed in Section 4.

4 AI methods for forest ecosystem service assessment

This section evaluates the application of AI and machine learning methods to forest ecosystem service research, organized by the first two tiers of the analytical framework introduced in Section 1: monitoring (Section 4.1) and assessment (Sections 4.2, 4.3). For each task domain, we examine how different methods perform, identify the conditions under which deep learning offers genuine advantages over conventional machine learning, and highlight persistent failure modes. Section 4.4 synthesizes these findings into a practical decision framework for method selection, and Section 4.5 traces the temporal evolution of methods across the review period.

4.1 Monitoring tier: forest cover change and land use transition detection

The detection of forest cover change from satellite imagery is the foundational task in AI-driven forest research, and the one with the longest methodological history. Conventional machine learning classifiers, particularly Random Forest (RF) and Support Vector Machine (SVM), have been the workhorses of land cover classification for over a decade. RF constructs ensembles of decision trees trained on random feature subsets, achieving theoretical guarantees against overfitting as the number of trees increases (Breiman, 2001). SVM finds optimal separating hyperplanes in high-dimensional feature space and performs well even with limited training samples (Maxwell et al., 2018). Systematic comparisons across multiple datasets show that neither method is universally superior: SVM achieved 89.1% overall accuracy (OA) vs. RF's 87.1% on the Indian Pines hyperspectral benchmark, while RF outperformed SVM (81.5% vs. 76.3%) on object-based classification tasks (Maxwell et al., 2018). A study using Sentinel-2 imagery for land cover mapping found SVM consistently outperforming RF and k-nearest neighbor (kNN), particularly under small sample conditions (93.96% vs. 90.71% vs. 89.80%), though all three classifiers converged above 93.85% when training samples exceeded approximately 750 pixels per class (Thanh Noi and Kappas, 2018). RF's practical advantages lie in its faster computation, fewer hyperparameters (typically only the number of trees and features per split), built-in variable importance measures, and greater robustness to label noise (Belgiu and Drăguţ, 2016; Breiman, 2001).

Alongside these classification-based approaches, land use transfer matrix analysis remains a widely used method for quantifying the magnitude and direction of transitions between land cover types across two or more time periods. While not an AI method per se, transfer matrices depend critically on the accuracy of the underlying land cover classifications, and the adoption of ML and DL classifiers has substantially improved the reliability of the transition estimates that feed into them. When combined with AI-enhanced multi-temporal classification, transfer matrices provide a structured accounting framework that is particularly useful for policy-relevant analyses of forest-to-agriculture or forest-to-urban conversion rates (Yang and Huang, 2021).

Deep learning methods have progressively entered the forest change detection domain, led by convolutional neural network architectures originally developed for computer vision. U-Net, with its symmetric encoder-decoder structure and skip connections that preserve fine-grained spatial information across resolution levels (Ronneberger et al., 2015), has become the most widely adopted architecture, appearing in 45% of deep learning studies on deforestation detection (Md Jelas et al., 2024). DeepLabv3+, which employs atrous spatial pyramid pooling (ASPP) to extract multi-scale features without reducing spatial resolution (Chen et al., 2018), is the second most common choice. In a direct comparison on Sentinel-2 forest change detection in Hunan Province, U-Net++ achieved the highest F1-score (0.771), followed by U-Net (0.765), LinkNet (0.752), DeepLabv3+ (0.744), and STANet (0.671), though DeepLabv3+ required substantially longer training time per epoch (12.35 min vs. 0.89 min for U-Net++) (Xiang et al., 2023). At higher spatial resolution, U-Net applied to PlanetScope imagery (4.77 m) across the entire state of California achieved 95.5% accuracy on independent validation, detecting forest disturbances that the 30 m Global Forest Change product missed entirely (Carter et al., 2024). This figure reflects the model's binary forest/non-forest classification accuracy rather than its performance at detecting forest loss specifically: when validated against CALFIRE wildfire and logging perimeters, the same study reports a substantially lower F1-score of 54.5% for forest loss detection, a gap the original authors partly attribute to known imprecision in the CALFIRE perimeter data (e.g., unburned areas retained within a fire's outer boundary) rather than to model error alone. Across all reviewed studies, deep learning models consistently outperformed conventional machine learning baselines for deforestation detection (Md Jelas et al., 2024).

However, the superiority of deep learning is conditional rather than absolute. A comparison of U-Net and RF on Landsat imagery in southeastern Australia found a 6-percentage-point advantage for U-Net (OA 79% vs. 73%), but also revealed that the CNN model could achieve higher accuracy than its own training labels, suggesting that label quality may matter more than architectural choice (Boston et al., 2022). When applied to multi-band time series, a deep learning temporal segmentation model outperformed conventional algorithms such as BFAST and CCDC, but underperformed them when only a single spectral band was used (He et al., 2024). Models trained in one climatic region can lose 10 to 30 percentage points in accuracy when transferred to another, with temperate-to-tropical transfer being particularly problematic (Amoah-Nuamah et al., 2025). Cloud contamination in tropical regions further degrades optical-based detection, with persistent cloud cover preventing change identification in affected pixels (Amoah-Nuamah et al., 2025).

From a theoretical perspective, the effectiveness of encoder-decoder architectures in landscape-scale classification can be understood through the lens of spatial scale theory in landscape ecology. Turner et al. (1989) demonstrated that landscape pattern metrics respond variably, and often nonlinearly, to changes in spatial grain, with rare land cover types disappearing first as resolution coarsens. The hierarchical feature extraction in U-Net and similar architectures mirrors this multi-scale structure: encoding layers progressively abstract local patterns into regional context, while decoding layers recover fine-grained spatial detail through skip connections. This architectural property aligns with the multi-scale nature of forest fragmentation and land use mosaics, providing a theoretical basis for why these models outperform pixel-level classifiers that lack spatial context.

Figure 2 synthesizes the accuracy data compiled across the studies reviewed in this section. In Figure 2a, traditional ML classifiers exhibit a wider accuracy range (73% to 96%) than deep learning models (78% to 81%), reflecting the diversity of datasets, resolutions, and sample sizes across ML studies rather than a genuine performance advantage. The narrow range of DL data points results from the smaller number of studies reporting overall accuracy for large-area forest classification using deep learning; most DL studies in this domain report F1-scores for change detection tasks instead (Figure 2b). Figure 2b reveals that within the deep learning paradigm, U-Net family architectures achieve F1-scores of 0.75 to 0.77 across diverse forest change detection applications. These panels are intended as a qualitative synthesis of reported performance across the reviewed literature rather than a formal meta-analysis: sample sizes, indicated for each group in the figure, range from a single study (the DeepLab and Other-DL groups in Figure 2b, following the exclusion of systematic-review-derived values described in the Figure 2 caption) to 13 (the traditional ML group in Figure 2a), and Figure 2c pools R² values across methodologically heterogeneous approaches, including CNN-LSTM and HRNet+CASA, that differ in input data sources and validation protocols; no formal significance testing was conducted. Groups based on very few studies should therefore be read as indicative rather than statistically representative.

4.2 Assessment tier: ecosystem service quantification and mapping

4.2.1 Carbon storage and net primary productivity

The estimation of forest carbon stocks and NPP represents the most mature application of AI in ecosystem service quantification. Machine learning approaches, led by RF and gradient boosting methods, are used to model the relationship between remotely sensed features and field-measured biomass. In a systematic review covering 25 studies, RF appeared as the baseline method in 88% of forest aboveground biomass (AGB) estimation research, though XGBoost outperformed RF in 75% of direct comparisons where both were tested (Nguyen and Saha, 2025). A multi-source data fusion study in Hangzhou demonstrated that CNN-LSTM achieved the highest accuracy (R² = 0.741, RMSE = 26.43 Mg/ha) when integrating six remote sensing sources (three SAR and three optical), outperforming RF on the same data (Tian et al., 2024). The improvement from multi-source fusion was substantial (5 to 10 percentage points in R²), though multi-SAR fusion alone yielded more modest gains, likely due to limited contrast among C-band and L-band backscatter features (Tian et al., 2024). A comparative study in tropical forests found that deep learning approaches achieved mean absolute percentage errors (MAPE) of 4.3% to 6.3% for simultaneous above- and belowground biomass prediction, outperforming conventional statistical methods on the same field datasets (Huy et al., 2025).

From a theoretical standpoint, the effectiveness of recurrent architectures such as LSTM in carbon flux and NPP time series modeling can be understood through the ecological concept of legacy effects. Forest carbon dynamics exhibit temporal dependencies in which current productivity is influenced not only by present conditions but also by antecedent climate, disturbance history, and nutrient cycling over preceding months or years. The gating mechanism of LSTM networks, which selectively retains or discards information across time steps, is architecturally well suited to capture these lagged ecological responses, providing a temporal analog to the spatial multi-scale reasoning that U-Net provides in the monitoring tier.

A persistent challenge across all biomass estimation studies is the saturation effect: models systematically underestimate high biomass values (above approximately 80 Mg/ha) and overestimate low values, a pattern observed in both conventional ML and deep learning approaches (Tian et al., 2024; Nguyen and Saha, 2025). We suggest that spaceborne lidar data from GEDI and ICESat-2, by providing direct measurements of canopy vertical structure rather than relying on passive spectral signals, represent a promising avenue for mitigating this saturation effect, though this inference draws on the general capabilities of lidar-derived structural metrics rather than a direct empirical test within the biomass estimation studies reviewed here. The overall performance advantage of deep learning over traditional ML for biomass estimation is visualized in Figure 2c, where CNN-LSTM models achieve consistently higher R² values than RF, particularly when multiple data sources are fused.

The coupling of AI classification with process-based models offers an alternative pathway for NPP estimation. Wang et al. (2024) demonstrated an HRNet-CASA framework in which a deep learning model first classified forest types from very high resolution imagery (OA = 80.60%), and the CASA model then estimated NPP using species-specific maximum light use efficiency parameters. This approach yielded R² values of 0.64 to 0.75 across forest types, with coniferous forests (R² = 0.75) estimated more accurately than bamboo forests (R² = 0.64), reflecting the greater structural homogeneity of conifer canopies.

Critically, the vast majority of AI-based carbon estimation studies report only point predictions without uncertainty bounds. Among the studies reviewed, only Tasuev et al. (2026) explicitly incorporated uncertainty quantification, using conformal prediction to generate spatially explicit confidence intervals alongside 10 m resolution carbon stock maps. For carbon estimates that may inform emissions trading or national carbon accounting, predictions without quantified uncertainty have limited policy credibility.

4.2.2 Water yield and hydrological regulation

AI methods for water-related ecosystem services typically operate through coupling with established hydrological models rather than replacing them. The InVEST water yield module, based on the Budyko curve framework, has been integrated with machine learning in several configurations. Bojer et al. (2025) used RF, SVM, XGBoost, and LightGBM to model climate inputs for InVEST in the Gilgel Gibe watershed, Ethiopia, finding that ML-optimized calibration reduced water yield estimation bias from 15% to 3.1% compared to default parameterization. The sensitivity of InVEST to its seasonality factor (Z parameter) underscores the value of this approach: default parameterization (Z = 10) yielded a Nash-Sutcliffe efficiency (NSE) of only 0.55, whereas ML-guided calibration (Z = 11.2) raised NSE to 0.70 (Bojer et al., 2025). Xi et al. (2024) coupled the PLUS land use simulation model with InVEST to project water supply and demand under three SSP-RCP scenarios in the Hanjiang River Basin, calibrating the Z factor to 3.08 to match observed runoff coefficients. Their analysis demonstrated that, using the Geodetector single-factor (q-statistic) detector, which independently quantifies each driver's spatial coupling with the dependent variable rather than partitioning variance additively across factors, land use type showed the strongest explanatory power for water supply spatial variation (q = 59%), followed by actual evapotranspiration (q = 47%) and precipitation (q = 40%); because Geodetector q-statistics are computed independently for each factor, they are not expected to sum to 100%. Spatial analysis using Moran's I (global values averaging 0.258 for supply and 0.288 for demand, both significant at p < 0.001) and Getis-Ord Gi* hot spot analysis confirmed significant spatial clustering in both water supply and demand patterns.

A counterintuitive finding from multiple studies is that forest conversion to other land uses increases water yield at the catchment scale, because forests consume more water through transpiration and canopy interception than cropland or built-up areas (Xi et al., 2024). This highlights the importance of evaluating ecosystem services as bundles rather than in isolation, as maximizing water yield would conflict with carbon sequestration and biodiversity objectives. Spatial regression approaches, particularly geographically weighted regression (GWR), have proven effective for disentangling the relative contributions of different drivers to ecosystem service variation, consistently outperforming global regression models due to the spatial non-stationarity of driver-service relationships (Sannigrahi et al., 2020).

4.2.3 Biodiversity assessment

Species distribution modeling (SDM) represents a growing frontier for AI in biodiversity-related ecosystem services. MaxEnt, the most widely used SDM algorithm, estimates habitat suitability from presence-only data by finding the maximum entropy distribution consistent with observed environmental constraints (Phillips et al., 2006). Deep learning approaches are now demonstrating measurable advantages over MaxEnt. Gillespie et al. (2024) developed a dual-head architecture (TResNet for remote sensing imagery plus MLP for climate variables) that simultaneously predicted the distributions of 2,221 plant species, achieving a median AUC of 0.9496 compared to 0.8825 for MaxEnt and 0.882 for RF, with the advantage increasing substantially in spatial cross-validation (AUC 0.8682 vs. 0.7339 for MaxEnt). Deneu et al. (2022) showed that CNN-based SDMs using 1 m resolution imagery outperformed environmental RF models on all metrics, with the advantage most pronounced for rare species, where the CNN's MSA30 score was nearly double that of RF (13.2% vs. 6.9%). Importantly, the CNN captured fine-scale habitat features such as roads, coastlines, and canopy gaps that coarser-resolution models could not resolve, breaking the classical precision-generality tradeoff in ecological modeling (Deneu et al., 2022). As illustrated in Figure 2d, CNN-based SDMs achieve more consistently high AUC values than both MaxEnt and RF, with the performance gap widening under spatial cross-validation conditions that test genuine geographic transferability.

4.3 Forest fragmentation and degradation assessment

Forest fragmentation analysis quantifies how the spatial configuration of forested landscapes changes over time, using landscape pattern indices such as mean patch area, patch density, landscape shape index, and aggregation index. AI methods have been integrated into fragmentation assessment primarily through two pathways: automated land cover classification that feeds into landscape metric calculation, and direct modeling of fragmentation drivers.

Lin et al. (2024) developed a composite Forest Fragmentation Change Index (FFCI) integrating five landscape metrics through principal component analysis, and applied RF to identify driving factors of fragmentation dynamics in southeastern China. Their analysis revealed that distance to forest edge and distance to roads were consistently the dominant predictors, and that 7.4% of forest areas experienced simultaneous vegetation improvement and increased fragmentation, demonstrating that greening trends do not necessarily prevent structural degradation. However, these driver-identification results should be interpreted with caution, as the underlying RF models explained only a modest share of variance (R² = 0.207 to 0.232), indicating that additional unmeasured factors likely contribute to fragmentation dynamics beyond the four predictors examined. Matyukira and Mhangara (2023) applied an XGBoost-Random Forest classifier to 30-year Landsat time series in South Africa, achieving OA of 0.90 (Kappa = 0.86), and documented a 44% decline in mean patch size for natural grasslands between 1990 and 2020, alongside an 11% increase in total patch count and a rise in Shannon diversity from 0.65 to 0.82, collectively indicating progressive fragmentation and loss of landscape dominance driven by invasive species encroachment.

These findings carry an important implication for ecosystem service assessment: vegetation indices such as NDVI, which measure greenness, can mask ongoing structural degradation. A landscape may appear healthy by spectral measures while becoming increasingly fragmented, with cascading negative effects on habitat connectivity, edge-sensitive species, and ecosystem service flows.

The key characteristics of the representative studies reviewed in Sections 4.1–4.3 are summarized in Table 3.

Table 3. Representative studies of AI methods for forest dynamics and ecosystem service assessment.

Of the 14 core studies summarized above, 11 (79%) reported code and/or data availability as “Not reported,” and three (Bojer et al., 2025; Gillespie et al., 2024; Deneu et al., 2022) provided open code or data. This near-total absence of open resources across the reviewed corpus substantiates both the case for standardized, openly available cross-domain benchmarks raised in Section 8.1 and the explicit call for open data sharing in the Conclusion (Section 9). The overall accuracies reported above for Thanh Noi and Kappas (2018) reflect the optimal training-sample condition (approximately 750 pixels/class); Section 4.1 additionally reports this study's results under small-sample conditions, which are correspondingly lower.

4.4 Method selection decision framework

The evidence reviewed above suggests that no single AI method is optimal across all forest monitoring and assessment tasks (Kazanskiy et al., 2025). To guide practitioners through this decision space, we synthesize the evidence into a flowchart (Figure 3) organized around five key considerations. This flowchart is intended as a practical starting point for method selection rather than a validated decision rule, since most of its branch points, including the sample-size thresholds discussed below, are drawn from individual studies rather than cross-study consensus. Method selection should be guided by the intersection of several practical considerations. When training data are limited, RF and SVM remain reliable choices: SVM is the least sensitive to training sample size among conventional classifiers, and all three methods (RF, SVM, kNN) converge above 93.85% OA once samples exceed approximately 750 pixels per class in a single study using Sentinel-2 imagery in Vietnam (Thanh Noi and Kappas, 2018); this threshold has not been independently validated across other sensors, biomes, or classification tasks and should be treated as an illustrative benchmark rather than a generalizable rule. RF offers additional practical advantages including faster computation, fewer hyperparameters, and greater robustness to label noise relative to single decision trees and ANNs (Maxwell et al., 2018). When training data are abundant and the task involves pixel-level spatial pattern recognition at high resolution, U-Net and its variants have outperformed pixel-based classifiers such as RF, for example by approximately 6 percentage points in a single Landsat-based comparison in southeastern Australia, though this gap was partly confounded by label-quality differences between the two models rather than architecture alone (Section 4.1) (Boston et al., 2022). Separately, in a single-province comparison of deep-learning architectures for Sentinel-2 forest change detection in Hunan, China, U-Net-family variants achieved the highest F1-scores among the architectures tested, though this comparison did not include a traditional pixel-based classifier as a baseline (Xiang et al., 2023). For species distribution modeling at fine spatial scales, CNN-based approaches now clearly outperform both MaxEnt and RF, particularly for rare species and when high-resolution imagery is available (Gillespie et al., 2024; Deneu et al., 2022).

4.5 Method evolution: 2016 to 2026

The methodological landscape has undergone a clear temporal shift across the review period (Figure 4). In the early years (2016 to 2018), RF dominated forest remote sensing applications, valued for its reliability, interpretability, and low computational requirements (Belgiu and Drăguţ, 2016). From 2019 to 2021, CNN architectures, particularly U-Net and DeepLabv3+, rapidly gained adoption for spatial classification and change detection tasks, driven by the availability of GPU computing and transfer learning from pre-trained models (Md Jelas et al., 2024). Since 2022, Transformer-based architectures and hybrid CNN-Transformer models have begun appearing in the literature, offering global contextual modeling at the cost of higher computational demands. Most recently, state-space models such as Mamba have been proposed as computationally efficient alternatives to Transformers for change detection, achieving linear rather than quadratic complexity in sequence length (Lei et al., 2026). Foundation models pre-trained on large-scale remote sensing datasets represent the latest frontier and are discussed in Section 8.

Beyond this temporal trajectory, Figure 5 maps the maturity landscape of AI applications across ten distinct research directions. Zone assignment reflects a qualitative editorial judgment informed primarily by each direction's publication history length (the x-axis position) among other considerations from the broader synthesis in this review, rather than a reproducible quantitative rule derived solely from the plotted variables. We note two properties of this classification that merit caution in interpretation. First, accuracy does not increase monotonically from mature to emerging tiers: mature ML-based biomass/carbon estimation, for instance, reports a lower median accuracy (R² = 0.41) than its growing-tier deep-learning counterpart (R² = 0.680), reflecting the saturation effect documented in Section 4.2.1. Second, earliest-appearance year on the x-axis correlates with, but does not by itself determine, zone membership, as illustrated by DL-based biomass/NPP estimation, which appears later on the x-axis than several emerging directions yet is classified as growing. Three clear tiers emerge. Mature directions include ML land-cover classification, for which eligible core studies report overall accuracies in the high 80s to mid 90s: 79.31% for RF-based national-scale mapping (Yang and Huang, 2021), 89.80% to 95.32% across RF, SVM, and kNN configurations (Thanh Noi and Kappas, 2018), and 85% (independent test set) to 89% (independent ground-truth validation) for XGBoost-Random Forest classification (Matyukira and Mhangara, 2023); MaxEnt species distribution modeling, with a median AUC of 0.878 (Phillips et al., 2006; Gillespie et al., 2024); and ML biomass estimation, with a median R² of 0.41 (Tian et al., 2024; Tasuev et al., 2026). These mature directions possess the longest publication histories and the largest literature bases. The notably lower accuracy in biomass estimation compared to classification tasks is not a methodological deficiency but rather a consequence of the saturation effect discussed in Section 4.2.1, where signal saturation above approximately 80 Mg/ha imposes a fundamental accuracy ceiling on passive remote sensing approaches (Tian et al., 2024). Growing directions include DL change detection, with median F1 of 0.752 (Xiang et al., 2023); DL biomass estimation, with median R² of 0.680 (Tian et al., 2024; Wang et al., 2024); and InVEST+ML coupling, with median R² of 0.65–0.75 (Bojer et al., 2025), where the literature continues to expand and accuracy levels have begun to support operational deployment. Emerging directions include CNN-based species distribution modeling, with median AUC of 0.921 (Gillespie et al., 2024; Deneu et al., 2022); edge AI/IoT monitoring, with median F1 of 0.930 (Amoah-Nuamah et al., 2025); and AI-driven fragmentation analysis, with reported OA of 0.90 (Matyukira and Mhangara, 2023). Although these directions remain limited in literature volume, they already demonstrate high accuracy. Notably, the emerging CNN-SDM direction reports a higher median accuracy than the mature MaxEnt direction (AUC 0.921 vs. 0.878); this comparison should be read with caution rather than as fully independent evidence, since three of the four CNN/DL values and two of the four MaxEnt values entering these medians are drawn from a single study, Gillespie et al. (2024), which reports both methods as a within-study benchmark rather than as separately replicated results. The direction of the comparison is nonetheless corroborated by evidence independent of that study on each side: Deneu et al. (2022) separately reports a CNN-based AUC of 0.915, and Phillips et al. (2006), the original MaxEnt publication, reports AUC values of 0.873 and 0.986 on taxa unrelated to the Gillespie et al. (2024) dataset. With this caveat, the comparison still illustrates that adoption maturity and performance superiority are two distinct dimensions: when architectural innovation aligns with task requirements, newer methods can match or exceed established ones, though the size of this specific gap should not be treated as a precise, fully independent multi-study effect estimate. Transformer and foundation model approaches remain at the earliest stage, with no forest-specific accuracy benchmarks yet reported (Lei et al., 2026).

5 Scenario simulation and forest dynamics prediction

This section addresses the third tier of the analytical framework: prediction. While Sections 4.1–4.3 examined how AI methods monitor current forest conditions and quantify ecosystem services, this section evaluates how simulation models project future land use trajectories and how AI is beginning to optimize forest management decisions under uncertainty.

5.1 Land use change simulation: from statistical to AI-enhanced models

The simulation of future land use and land cover change is essential for anticipating how forest dynamics and associated ecosystem services will respond to alternative development pathways. Over the past three decades, simulation models have evolved through three distinct generations (Sun et al., 2025). The first generation, exemplified by CA-Markov models, combines cellular automata spatial rules with Markov chain transition probability matrices to project land use change based on historical patterns. Bashir et al. (2022) applied CA-Markov to simulate 2030 land use in the western Himalayas, achieving a Kappa of 0.81 when validating 2020 predictions from 2000 and 2010 inputs. Their analysis projected continued forest decline from 1,823 km² in 2000 to 1,630 km² by 2030, driven by the expansion of horticulture (53.57% increase over 2000 to 2020) and urbanization (31.2% increase). However, CA-Markov models assume stationary transition probabilities and cannot simulate abrupt trend changes or land use types without historical precedent.

The second generation introduced spatially explicit, demand-driven frameworks. CLUE-S (Verburg et al., 2002) separates non-spatial demand calculation from spatial allocation, using stepwise logistic regression to generate probability surfaces for each land use type. Validation on Sibuyan Island, Philippines, yielded ROC values of 0.82 to 0.94 across four land use types, with forest achieving the highest fit (0.94). CLUE-S introduced two conceptual advances that persist in subsequent models: hierarchical organization, in which macro-level demand constraints guide micro-level spatial allocation, and conversion elasticity parameters that control the resistance of each land use type to change.

The third generation integrates machine learning into the simulation framework. FLUS (Li et al., 2017) replaced logistic regression with artificial neural networks for generating land use probability surfaces and introduced a self-adaptive inertia and competition mechanism to handle multi-type land use transitions simultaneously. Applied at 1 km global resolution across 17 world regions, FLUS achieved Figure of Merit values of 10% to 29% (mean 19%) for 2000 to 2010 validation, comparable to or exceeding previous global models. Critically, the resolution analysis demonstrated that coarsening from 1 km to 10 km caused a 19.77% underestimation of urban area and lost 60% to 97% of simulated urban growth across four SRES scenarios, underscoring the importance of fine-resolution simulation for capturing spatially heterogeneous processes such as forest fragmentation.

The most recent evolution, the PLUS model, employs Random Forest in its Land Expansion Analysis Strategy (LEAS) module to identify drivers of land use expansion, combined with a cellular automata module using random patch seeds for spatial allocation. The shift from logistic regression and Markov chains to tree-based ensemble methods represents a fundamental change in how driver-response relationships are modeled: Random Forest captures nonlinear and interactive effects among driving factors (such as the joint influence of slope, road proximity, and soil quality on deforestation probability) that linear or stationary-probability models cannot represent, which is the primary reason for the consistent accuracy gains observed across successive model generations. Wang et al. (2023) applied PLUS to Yunnan Province under three SSP scenarios (SSP126, SSP245, and SSP585) derived from CMIP6 climate projections, achieving OA of 0.9379 and Kappa of 0.8947. Their results projected declining carbon storage and habitat quality under all scenarios, with SSP585 producing the lowest habitat quality (0.7348 vs. 0.7420 under SSP126) but the highest water yield (1,731.80 mm vs. 1,573.25 mm under SSP126), a pattern consistent with the general forest conversion water yield relationship discussed in Section 4.2.2 and further reinforced under SSP585 by increased precipitation. Spatial autocorrelation analysis using Moran's I confirmed significant clustering of ecosystem service values under all scenarios (0.69 to 0.78), with Getis-Ord Gi* hot spot analysis revealing that SSP585 produced more dispersed hot spots and westward-shifted cold spots compared to SSP126 and SSP245 (Wang et al., 2023).

Li et al. (2025) extended the PLUS framework by incorporating gradient boosting to identify key drivers of ecosystem service variation before simulation. Their analysis of the Yunnan-Guizhou Plateau demonstrated that land use type was the dominant driver of habitat quality and carbon storage (importance score exceeding 1.85), while precipitation dominated water yield (score exceeding 1.26). A comparison of five machine learning models for driver identification found gradient boosting consistently outperforming alternatives, achieving R² approaching 1.0 for carbon storage prediction, while SVM produced negative R² values indicating severe overfitting (Li et al., 2025). Under three scenarios projected to 2035, the ecological priority scenario yielded the highest values across all ecosystem service indicators, confirming that proactive conservation policies can simultaneously enhance carbon storage, habitat quality, water yield, and soil retention.

5.2 Climate scenarios and forest vulnerability prediction

All of the simulation models reviewed above require externally specified climate and socioeconomic trajectories as boundary conditions. The scenario frameworks underpinning these simulations are drawn from the Coupled Model Intercomparison Project Phase 6 (CMIP6), which combines Shared Socioeconomic Pathways (SSPs) describing societal trajectories with Representative Concentration Pathways (RCPs) describing radiative forcing levels (O'Neill et al., 2016). The four Tier 1 scenarios span from SSP1-2.6 (sustainable development, consistent with the 2°C target) to SSP5-8.5 (fossil fuel intensive growth, high-end forcing), with SSP2-4.5 and SSP3-7.0 representing intermediate and regional-rivalry pathways respectively. Land use change trajectories diverge substantially across SSPs: SSP1 projects the largest forest area increase, while SSP3 projects the greatest cropland expansion and forest loss (O'Neill et al., 2016).

Beyond land use simulation, machine learning is also being applied to predict forest ecosystem vulnerability directly from climate projections. Otieno et al. (2025) combined CMIP6 SSP2-4.5 projections from three general circulation models with Random Forest, XGBoost, and SVM to predict vegetation index trajectories on Mount Kenya. Random Forest achieved the best predictive performance for Enhanced Vegetation Index (R² = 0.82, RMSE = 0.07), while SVM performed best for Bare Soil Index (R² = 0.74) but poorly for EVI (R² = 0.57). Historical analysis revealed a 45.4% decline in NDVI and 55.1% increase in bare soil exposure between 2000 and 2020, with future projections indicating a further 49% to 55% decline in EVI by 2040 under SSP2-4.5, concentrated in low-elevation zones where anthropogenic pressure and climate sensitivity intersect (Otieno et al., 2025).

5.3 AI-optimized forest management planning

The fourth tier of the analytical framework, optimization, represents the most recent and least mature application of AI in forest ecosystem service research. Reinforcement learning (RL), in which an agent learns optimal sequential decisions through trial-and-error interaction with a simulated environment, offers a natural framework for long-term forest management planning where current harvesting decisions affect future forest states and service provision. Just as the encoder-decoder architecture of U-Net mirrors the multi-scale spatial structure of landscapes (Section 4.1) and LSTM memory gates mirror the temporal legacy effects in carbon dynamics (Section 4.2.1), the sequential decision framework of RL mirrors the multi-decadal planning horizon of forest management, where actions taken today constrain the option space available decades later.

Ji et al. (2024) applied Q-learning to optimize a 60-year harvesting plan across 353 forest compartments in South Korea, balancing timber production, carbon storage, and water retention. Under four weighting scenarios, the carbon-priority scenario (S3) unexpectedly produced the highest cumulative economic value (54.4 billion KRW, approximately twice that of the timber-priority scenario S2 at 23.9 billion), because prioritizing carbon storage maintained higher standing volumes that generated greater long-term timber value (Ji et al., 2024). This finding provides quantitative evidence that ecosystem service objectives and economic objectives need not be in conflict.

Malo et al. (2021) advanced the methodological frontier substantially by applying hybrid Proximal Policy Optimization (H-PPO), a deep reinforcement learning algorithm, to a stochastic optimal harvesting problem with 44 continuous state variables (4 tree species across 11 diameter classes), mixed discrete-continuous action spaces, and natural disaster risk. The deep RL approach reproduced known optimal solutions for the deterministic problem (bare land value of EUR 2,336 vs. the benchmark EUR 2,335) while reducing computation time to 0.06% of the original method (6 min vs. 170 h). Under stochastic growth conditions, the deterministic-equivalent principle held approximately: the optimal policy derived under certainty performed nearly as well in stochastic environments (BLV EUR 2,355 vs. EUR 2,357 under the fully stochastic policy). A novel finding was that natural disaster risk favored continuous cover forestry over rotation forestry, reversing the conventional expectation, because disaster risk functions similarly to an increased discount rate, which theoretically favors the continuous cover regime.

These reinforcement learning applications remain at the proof-of-concept stage, with neither study incorporating real-time monitoring data or ecosystem services beyond timber and carbon. However, the demonstrated ability to handle high-dimensional state spaces, stochastic dynamics, and multi-objective tradeoffs positions deep RL as a promising framework for integrating the monitoring, assessment, and prediction capabilities reviewed in Sections 4, 5.1 into actionable management recommendations. A recently introduced benchmark, BoreaRL, illustrates both the promise and the practical difficulty of extending these proof-of-concept applications to genuinely multi-objective settings. BoreaRL couples a physically grounded simulator of coupled energy, carbon, and water fluxes with a multi-objective reinforcement learning framework for climate-adaptive boreal forest management, explicitly modeling the tradeoff between carbon sequestration and permafrost preservation (Dsouza et al., 2025). Benchmarking results reveal a fundamental asymmetry in learning difficulty: carbon-focused objectives are substantially easier for current MORL algorithms to optimize than permafrost-preservation objectives, and standard gradient-based preference-conditioned approaches fail to generalize across sites, with a heuristic episode-selection mechanism, built on an untrained, randomly initialized site-scoring function, outperforming these more sophisticated gradient-based alternatives. This finding is instructive for the RL applications reviewed above: Q-learning approached the multi-objective balance of timber, carbon, and water retention through a small number of fixed weighting scenarios rather than a systematic exploration of the trade-off space (Ji et al., 2024), while H-PPO was validated on a single-objective stochastic harvesting simulator optimizing economic timber value under natural disaster risk, without an explicit ecosystem-service reward beyond this economic objective (Malo et al., 2021); neither has yet been tested on a systematically preference-conditioned, physically-grounded benchmark of the kind BoreaRL provides. BoreaRL's results suggest that extending these approaches to explicitly multi-objective settings involving additional, potentially conflicting ecosystem service objectives, such as biodiversity or hydrological regulation, is unlikely to be a straightforward extension, since even a two-objective carbon-permafrost tradeoff already exposes substantial limitations in current MORL methods, underscoring both the simulator-to-real gap and the objective-conflict challenges that concrete benchmarks like BoreaRL are designed to expose.

6 GIS, geovisualization, and decision support systems

The AI and simulation methods reviewed in Sections 4, 5 do not operate in isolation. Their practical value for forest management depends on computational infrastructure for processing large-scale geospatial data, spatial analytical methods for interpreting results, and emerging frameworks for integrating monitoring, modeling, and decision support into unified systems. This section examines three enabling components: spatial statistical methods for ecosystem service pattern analysis (Section 6.1), cloud computing platforms for scalable geospatial processing (Section 6.2), and the emerging concept of forest digital twins as integrative decision support frameworks (Section 6.3).

6.1 Spatial statistical methods for ecosystem service analysis

Spatial autocorrelation analysis has become a standard component of AI-driven ecosystem service assessments, as demonstrated by the Moran's I and Getis-Ord Gi* applications reviewed in Sections 4.2.2, 5.1. The theoretical foundation for these analyses rests on Local Indicators of Spatial Association (LISA), formalized by Anselin (1995). The Local Moran statistic, defined as

Ii=zi∑jwijzj(1)

where z_*i* and z_*j* are the standardized deviations of observed values at locations i and j from the mean, and w_*ij* are elements of the spatial weights matrix, quantifies the degree of significant spatial clustering around each observation (Equation 1), with the sum of all local values proportional to the global Moran's I. The Moran scatter plot provides a complementary interpretation framework: the four quadrants (HH, LL, LH, and HL) distinguish positive spatial autocorrelation (high-value or low-value clusters) from negative autocorrelation (spatial outliers), enabling identification of both hot spots and anomalous locations that may warrant targeted management attention (Anselin, 1995).

A critical methodological caveat is that the null distribution of Local Moran statistics deviates substantially from normality (kurtosis 7.9 vs. the normal value of 3.0), meaning that normal approximations tend to over-reject the null hypothesis. Furthermore, the presence of global spatial autocorrelation alters the mean, standard deviation, and skewness of individual LISA statistics, invalidating theoretical moment calculations. Conditional randomization with permutation-based inference (typically 10,000 permutations) and Bonferroni correction for multiple comparisons represent the reliable approach for significance testing in spatially autocorrelated datasets (Anselin, 1995). These considerations are directly relevant to the ecosystem service mapping applications reviewed earlier, where water supply patterns with Moran's I = 0.258 (Xi et al., 2024) and ecosystem service values with Moran's I = 0.69 to 0.78 (Wang et al., 2023) both exhibited strong spatial dependence.

6.2 Cloud computing platforms for scalable forest analysis

The transition from desktop-based GIS to cloud computing platforms has fundamentally altered the scale at which forest monitoring and ecosystem service assessment can be conducted. Google Earth Engine (GEE), a planetary-scale geospatial analysis platform, integrates a multi-petabyte data catalog with high-performance parallel computing services (Gorelick et al., 2017). The platform's architecture employs lazy evaluation (computing only the output pixels currently requested), information-preserving storage (maintaining original projections and bit depth to avoid resampling degradation), and four parallel processing models: image tiling, spatial aggregation via scatter-gather, streaming collection processing, and subexpression caching. Performance scales near-linearly with CPU allocation, as demonstrated by processing 654,178 Landsat scenes for the Hansen et al. (2013) global forest change map and achieving throughput of approximately 2 minutes per 10,000 km² for crop yield estimation (Gorelick et al., 2017).

A meta-analysis of 349 GEE publications (2010 to 2019) found that 90% used remote sensing data directly, with Landsat comprising 82% of data sources. Random Forest was the most commonly applied classifier (49% of studies), achieving a median overall accuracy of approximately 90%, followed by CART (13%), SVM (11%), and decision trees (12%) (Tamiminia et al., 2020). An updated review of 72 GEE studies for land use classification (2016 to 2025) confirmed the continued dominance of RF (36%) and SVM (19%), with overall accuracies ranging from 80% to 98.68%, though deep learning methods are beginning to appear on the platform (approximately 7% of studies) (Alshehri et al., 2025).

Despite these capabilities, GEE imposes significant constraints that affect AI-driven forest research. Memory limitations cap the size of training datasets for machine learning models; interactive computations time out at 270 seconds; and the platform lacks native support for deep learning, object-based image analysis, and non-parallelizable operations such as watershed delineation (Gorelick et al., 2017; Tamiminia et al., 2020). Although TensorFlow integration has been available since 2019, GPU resource allocation and advanced architectures (such as U-Net or Transformer models) cannot be flexibly deployed on GEE (Alshehri et al., 2025). A practical workflow increasingly adopted in the literature uses GEE for data acquisition, preprocessing, and initial classification, with local GPU environments handling deep learning model training and inference.

6.3 Forest digital twins: toward integrated decision support

The concept of digital twins, which originated in manufacturing, is beginning to be explored as an integrative framework for forest ecosystem monitoring, simulation, and management. A forest digital twin (FDT) can be understood as a virtual representation of a physical forest that is continuously updated through data exchange and supports simulation-based decision making (Buonocore et al., 2022; Döllner et al., 2023).

Buonocore et al. (2022) proposed a nested FDT framework in which individual tree digital twins, each characterized by 15 groups of state variables spanning leaf temperature, radial growth, root dynamics, sap flow, soil moisture, and microclimate conditions, are aggregated into forest-level representations. The framework incorporates process models for seven physiological processes (including photosynthesis, phenology, BVOC emissions, and transpiration) and envisions a blockchain layer for transparent carbon credit verification and trading. Döllner et al. (2023) complemented this conceptual design with a six-layer technical architecture progressing from physical forest through data lake, semantic processing, forest information model, simulation, and visualization to decision support. Their design emphasizes that all raw data must be preserved in the data lake because processing methods will improve and new applications may require re-extraction from historical records.

A critical distinction for assessing the current state of FDT development is the three-level classification distinguishing digital models (no automatic data exchange), digital shadows (one-way physical-to-digital data flow), and true digital twins (bidirectional integration) (Tagarakis et al., 2024). A systematic review of 34 agricultural and forestry digital twin studies found that only 6 achieved full bidirectional integration, with the majority remaining at the digital model (10 studies) or digital shadow (18 studies) stage, and only one system deployed at operational readiness (TRL 7 to 9) (Tagarakis et al., 2024). Forestry-specific digital twin research emerged only in 2022 to 2023, with applications including 3D tree skeleton reconstruction from LiDAR, forest scene rendering, UAV flight planning, and greenhouse gas monitoring, all at concept or prototype stage.

The operational gap between concept and deployment is further illustrated by Jiang et al. (2022), who developed an LSTM-based approach to predict future Landsat imagery of forest landscapes, achieving a correctness rate of 80% for the best-performing LSTM-CGAN architecture. While this represents a digital model rather than a true digital twin, it demonstrates the feasibility of temporal prediction as a component of future FDT systems. The integration of such prediction capabilities with the ALS-derived three-dimensional forest structure data reviewed by White et al. (2016), which achieves sub-meter tree height accuracy and has reached operational status for area-based forest inventory approaches, represents a plausible near-term pathway toward functional forest digital shadows.

7 Key challenges

The preceding sections have documented substantial progress in applying AI to forest ecosystem service research. However, several persistent challenges limit the reliability, transferability, and policy relevance of current approaches. This section organizes these challenges into three layers: data constraints, methodological limitations, and application barriers.

7.1 Data constraints

The most fundamental constraint is the scarcity of high-quality labeled data for forest remote sensing tasks. Training deep learning models for deforestation detection, species classification, or biomass estimation requires spatially explicit ground truth that is expensive and labor-intensive to produce, particularly in remote tropical forests where field access is limited. Fewer than 20% of AI-based forest monitoring studies have focused on tropical ecosystems, and models trained in temperate regions lose 10 to 30 percentage points in accuracy when transferred to tropical contexts (Amoah-Nuamah et al., 2025). The annotation challenge is compounded by the inherent difficulty of forest label creation: even expert interpreters disagree on tree species identification from aerial imagery, with reported error rates of 30% to 60% for visual crown-level species assignment (White et al., 2016).

Cloud contamination further limits data availability in tropical and montane regions, preventing consistent time series construction from optical sensors. The resulting data gaps propagate into change detection algorithms, producing incomplete or biased assessments of forest dynamics (Xiang et al., 2023). While SAR data from Sentinel-1 offers all-weather observation capability, its integration into forest monitoring workflows remains limited by the additional complexity of SAR preprocessing and interpretation.

7.2 Methodological limitations

7.2.1 The explainability deficit

The rapid adoption of deep learning in forest remote sensing has not been accompanied by commensurate investment in model explainability. A systematic review of XAI in remote sensing found an approximately 70-fold gap between the number of machine learning publications and those addressing explainability (Höhl et al., 2024). Among studies that do apply XAI methods, over 90% rely solely on anecdotal evidence, selecting individual examples for visual inspection rather than conducting systematic quantitative evaluation (Höhl et al., 2024). SHAP dominates XAI application in environmental sciences (135 of 575 reviewed studies), with 80% of studies employing only a single explanation method (Schiller et al., 2025). This methodological monoculture, rooted in SHAP's game-theoretic appeal and software accessibility rather than demonstrated superiority for specific tasks, risks creating what has been described as an “illusion of sufficient understanding” (Schiller et al., 2025; Messeri and Crockett, 2024). More broadly, only 1.2% of environmental AI studies have addressed trustworthiness as a core research objective, despite frequent claims that XAI methods inherently increase trust (Schiller et al., 2025).

The challenge is compounded by remote sensing-specific factors that conventional XAI methods, designed for natural images, do not address: the semantic significance of spatial and spectral resolution (the same landscape appears differently across sensors), geographic dependencies arising from unobserved factors such as atmospheric teleconnections, and temporal dependencies in time series that single-image attribution methods cannot capture (Höhl et al., 2024). Furthermore, the common practice of treating Random Forest as an “interpretable” baseline against which deep learning “black boxes” are compared is itself questionable, as social science literature establishes that large ensemble models with thousands of trees are not genuinely interpretable (Gevaert, 2022).

From a regulatory perspective, the EU AI Act requires high-risk AI systems to be “sufficiently transparent” for users to interpret outputs and use them appropriately (Gevaert, 2022). For forest ecosystem service assessments that inform carbon credit markets, conservation planning, or land use policy, this creates a concrete compliance requirement that current practice largely fails to meet.

7.2.2 Uncertainty quantification

Closely related to the explainability deficit is the near-absence of uncertainty quantification in AI-based forest assessments. As noted in Section 4.2.1, only one of the reviewed biomass estimation studies (Tasuev et al., 2026) incorporated formal uncertainty bounds. This gap is problematic because the two fundamental types of uncertainty, aleatoric (irreducible data noise) and epistemic (reducible knowledge gaps), have different management implications: the former sets a floor on achievable accuracy, while the latter can be reduced through additional data collection or model improvement (Abdar et al., 2021). Established methods such as MC Dropout, deep ensembles, and variational inference exist but remain rarely applied in forest remote sensing, despite their direct relevance to applications where prediction errors carry financial or ecological consequences.

7.2.3 Cross-domain transferability

The transferability of trained models across geographic regions, sensor configurations, and temporal periods remains a major barrier to operational deployment. Domain shift in remote sensing arises from differences in ground sampling distance, imaging modality (e.g., RGB vs. infrared composites), landscape context (e.g., urban vs. rural building morphology), and environmental conditions including weather, illumination, and seasonal phenology (Lyu et al., 2025). While unsupervised domain adaptation methods, including distribution alignment, adversarial learning, and self-training with pseudo-labels, have been developed to address these shifts, their application in forest research remains limited. Ahishali et al. (2025) demonstrated that direct cross-site transfer for dead tree segmentation from aerial imagery reduced Dice score from 0.74 (within-domain) to 0.24 (cross-domain), a 50-percentage-point degradation. Domain adaptation recovered performance to 0.44, recovering approximately 40% of the gap but still falling far short of within-domain accuracy. Foundation models such as the Segment Anything Model (SAM) represent a potential pathway to improved transferability, but have not yet demonstrated consistent advantages over task-specific models in remote sensing benchmarks (Lyu et al., 2025).

A related but conceptually distinct source of over-optimism concerns how model accuracy is validated in the first place, rather than how it degrades under transfer. Many of the accuracy figures compiled in this review, particularly for species distribution modeling (Section 4.2.3) and biomass estimation (Section 4.2.1), are reported under cross-validation schemes that, even when point-level spatial buffers are applied to exclude near-duplicate locations from the test set, still draw training and test samples from overlapping broad geographic regions. Because ecological and remotely sensed variables remain spatially autocorrelated at scales larger than typical buffer distances, such schemes can still overstate the accuracy a model would achieve when genuinely extrapolating to previously unseen regions. The spatial cross-validation results reported for species distribution models in this review illustrate the resulting gap directly: AUC for the CNN-based approach of Gillespie et al. (2024) fell from 0.9496 under a within-state random split with point-level buffering to 0.8682 when entire latitudinal bands were held out as genuinely unseen regions, while MaxEnt fell further, from 0.8825 to 0.7339, under the same regional holdout. This indicates that even carefully buffered random splits likely overstate the transferability of both approaches to new geographic areas relative to true regional extrapolation. Spatial or spatiotemporal blocking, in which entire geographic clusters or time periods are held out together rather than individual samples, provides a more realistic estimate of out-of-sample performance and should be treated as a minimum reporting standard for accuracy claims in this domain. Standardized out-of-distribution benchmarking frameworks developed outside forest remote sensing, such as WILDS (Koh et al., 2021), which explicitly includes satellite imagery and poverty mapping tasks with geographic and temporal distribution shifts, offer a template for the kind of standardized, held-out evaluation protocol that forest ecosystem service studies currently lack.

7.3 Integration and deployment barriers

7.3.1 The physics-AI integration gap

A persistent structural challenge is the separation between data-driven AI approaches and process-based ecological models. Reichstein et al. (2019) identified five pathways for integrating deep learning with physical models in Earth system science: improving parameterization, replacing weak sub-models, analyzing model-observation biases, constraining sub-model inputs, and surrogate modeling. Despite this clear conceptual framework, most forest ecosystem service studies reviewed here employ AI and process models sequentially (e.g., classification followed by InVEST) rather than in genuinely integrated architectures where physical constraints inform neural network learning. The result is that AI models may produce physically inconsistent or implausible predictions when extrapolated beyond training conditions, particularly when not constrained by conservation principles that process-based models inherently enforce (Reichstein et al., 2019).

7.3.2 Temporal scale mismatch

Finally, a gap exists between the temporal resolution of AI-based monitoring (typically annual to sub-annual) and the timescales over which forest ecosystem processes operate. Habitat fragmentation effects accumulate over decades: species richness declines by more than 20% within 1 year of fragmentation, exceeds 50% after 10 years, and continues increasing in the longest available time series beyond 20 years (Haddad et al., 2015). Ecosystem function debt follows a similar trajectory, with losses reaching 80% in small, isolated patches after a decade. With nearly 20% of global forest situated within 100 m of a forest edge, where microclimate alteration and edge effects are most severe, and over 70% within 1 km, these cumulative effects are pervasive (Haddad et al., 2015). We infer that current AI approaches, which predominantly analyze static snapshots or short time series, may risk underestimating the full extent of fragmentation-driven ecosystem service loss by failing to capture these slow, compounding dynamics, though no AI-based forest monitoring study reviewed here has directly tested this hypothesis over multi-decadal timescales, and confirming it would require longitudinal validation against the kind of long-term ecological datasets used by Haddad et al. (2015).

8 Future directions and research outlook

The challenges identified in Section 7 point toward several research frontiers that are likely to reshape AI-driven forest ecosystem service assessment over the coming decade. This section organizes these directions along three time horizons: near-term methodological improvements that can be implemented with existing tools, medium-term architectural shifts centered on foundation models, and long-term systemic transformations toward integrated forest intelligence.

8.1 Near-term: methodological consolidation (2025 to 2028)

Three priorities can be addressed immediately using available methods and data. First, uncertainty quantification should become a standard reporting requirement rather than an optional addition. Methods such as MC Dropout, deep ensembles, and variational inference are mature and computationally tractable (Abdar et al., 2021), yet remain nearly absent from forest remote sensing applications. Establishing community norms for reporting prediction intervals alongside point estimates would substantially increase the policy credibility of AI-derived carbon stock maps and ecosystem service assessments.

Second, explainability practices need diversification beyond the current SHAP monoculture. The development of XAI methods that account for the spatial, spectral, and temporal structure of remote sensing data represents a tractable research goal, with concept-based explanations (such as TCAV) and counterfactual approaches offering underexplored alternatives to attribution-based methods (Höhl et al., 2024; Gevaert, 2022).

Third, standardized benchmarking protocols for cross-domain transferability should be established. Current studies report accuracy on single sites without systematically testing geographic, sensor, and temporal transfer, making it impossible to assess operational readiness. Benchmark datasets spanning multiple biomes, sensor configurations, and time periods would enable meaningful comparison of domain adaptation strategies (Lyu et al., 2025).

8.2 Medium-term: foundation models for forest research (2028 to 2032)

Remote sensing foundation models (RSFMs), pre-trained on large-scale Earth observation data and adaptable to diverse downstream tasks through fine-tuning, represent the most significant architectural shift on the horizon (Xiao et al., 2024). Current RSFMs are categorized into visual foundation models (VFMs) using self-supervised pretraining strategies such as masked image modeling, vision-language models (VLMs) combining image understanding with text-based reasoning, and emerging architectures including state-space models such as Mamba that offer linear computational complexity for long-sequence processing (Huo et al., 2025).

However, existing RSFMs have not yet achieved the transformative impact observed in natural language processing. On standard benchmarks, no single model dominates across all tasks: SkySense achieves the highest multi-label classification accuracy (mAP 88.7% on BigEarthNet), while different models lead on detection and segmentation tasks (Xiao et al., 2024). More critically, VLMs perform below random chance on attribute reasoning and evaluation reasoning tasks in remote sensing contexts, revealing fundamental gaps in geospatial understanding (Xiao et al., 2024; Huo et al., 2025).

Zhu et al. (2026) proposed a framework of 11 characteristics that an ideal Earth foundation model should satisfy, including geographic location embedding, geographically balanced representation, scale awareness, wavelength embedding, multi-sensor fusion, temporal variables, task agnosticism, carbon emission minimization, uncertainty quantification, physical consistency, and integration with large language models. An evaluation of existing models against these criteria found that no current model satisfies all 11 characteristics, with uncertainty quantification and physical consistency representing the weakest dimensions (Zhu et al., 2026). Among these eleven characteristics, three carry particular weight for forest ecosystem service research: scale awareness, given the multi-scale nature of forest structure from individual tree crowns to landscape-level fragmentation patterns (Section 4.1); uncertainty quantification, given the near-total absence of prediction intervals in current carbon and biomass estimation studies (Section 7.2.2); and geographically balanced representation, given the systematic under-representation of tropical forests in training data (Section 7.1) despite these regions bearing the greatest deforestation pressure. For forest ecosystem service research specifically, the most impactful near-term applications of foundation models are likely to be few-shot classification and geographic search, as demonstrated by Strong et al. (2025), who deployed a foundation model-based system for detecting concentrated animal feeding operations across six US states using only a handful of user-provided examples, with training and inference each requiring only minutes on a MacBook Pro (Strong et al., 2025).

8.3 Long-term: toward integrated forest intelligence (2032 and beyond)

The long-term vision is a convergence of the monitoring, assessment, prediction, and optimization tiers reviewed in this article into unified, continuously learning systems. Forest digital twins (Section 6.3) represent one pathway toward this integration, but their realization requires advances on multiple fronts: bidirectional data exchange between physical forests and virtual representations currently achieved by fewer than 20% of digital twin systems (Tagarakis et al., 2024), physics-informed neural architectures that embed conservation laws into learning (Reichstein et al., 2019; Zhu et al., 2026), and reinforcement learning frameworks that translate monitoring and prediction outputs into actionable management recommendations (Section 5.3).

A critical enabling condition is the resolution of what Zhu et al. (2026) identified as geographically imbalanced representation: with over 1,000 active remote sensing satellites generating nearly 100 PB of open data but labeled datasets covering only approximately 0.1% of this volume, and with training data disproportionately drawn from Europe and North America, current models risk perpetuating geographic biases that underserve the tropical and subtropical forests where biodiversity loss and deforestation pressure are greatest. Newbold et al. (2015) documented that human land use has already reduced local species richness by an average of 13.6% globally, with losses concentrated in economically disadvantaged regions, and projected that under business-as-usual scenarios this decline will reach 41.5% of terrestrial grid cells exceeding the 20% threshold associated with substantial impairment of ecosystem function by 2095. Ensuring that next-generation AI systems serve these high-priority regions equitably is both a technical challenge (domain adaptation, few-shot learning) and an ethical imperative. The staged research agenda is summarized in Figure 6.

9 Conclusion

This review has examined the current state and future trajectory of AI-driven assessment of forest dynamics and ecosystem services in the context of land use change. Organized around a four-tier analytical framework spanning monitoring, assessment, prediction, and optimization, the review synthesized evidence from 27 core empirical studies directly reporting a quantified AI/ML performance metric, supplemented by 70 contextual references.

At the monitoring tier, deep learning architectures, particularly U-Net and its variants, have demonstrated consistent advantages over conventional machine learning classifiers for forest change detection, with F1-scores of 0.75 to 0.77 across diverse applications. However, this advantage is conditional: models lose 10 to 30 percentage points in accuracy when transferred across climatic regions, and performance on single-band inputs can fall below that of established algorithms. At the assessment tier, AI methods have advanced carbon stock estimation (CNN-LSTM R² = 0.741 with multi-source fusion), species distribution modeling (CNN AUC = 0.921 vs. MaxEnt 0.878), and water yield prediction (ML-calibrated InVEST reducing bias from 15% to 3.1%), while revealing a critical gap in uncertainty quantification that limits policy applicability. At the prediction tier, land use simulation models have evolved from stationary Markov chains to machine learning-enhanced frameworks (PLUS achieving OA = 0.94), and reinforcement learning has provided initial evidence that ecosystem service and economic objectives can be simultaneously optimized. Cloud computing platforms, spatial statistical methods, and emerging digital twin concepts provide the computational and conceptual infrastructure linking these tiers.

Three overarching themes emerge from this synthesis. First, the choice of AI method should be driven by data availability and task characteristics rather than architectural novelty: Random Forest remains the optimal choice when training samples are scarce, while deep learning provides measurable gains only when sufficient labeled data and spatial context are available, and the reported magnitude of these gains can itself be confounded by factors such as training label quality rather than architecture alone (Section 4.1). Second, the integration of AI with process-based ecological models remains shallow, with most studies employing sequential rather than genuinely coupled workflows, limiting physical consistency, and extrapolation reliability. Third, the field faces a collective action problem in explainability, uncertainty quantification, and cross-domain evaluation, where individual studies lack incentives to invest in these dimensions despite their importance for cumulative scientific progress and policy trust. The second of these three themes applies reflexively to the present synthesis: the four-tier framework organizing this review traces the causal linkages between tiers conceptually, but it is itself a taxonomic and organizational integration rather than a technically integrated computational system, and closing this gap will require the kind of bidirectionally coupled forest digital twins discussed in Section 6.3 rather than narrative synthesis alone.

Foundation models pre-trained on large-scale Earth observation data represent the most promising pathway for addressing several of these challenges simultaneously, but their application to forest ecosystem services remains nascent, with no current model satisfying the full set of requirements for geographic balance, physical consistency, and uncertainty quantification. Realizing the potential of AI for forest ecosystem stewardship will require not only methodological advances but also institutional commitments to standardized benchmarking, open data sharing, and interdisciplinary collaboration between AI researchers, ecologists, and policy makers.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/ffgc.2026.1933500/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Avram Sorin, University of Craiova, Romania

Reviewed by: Omarsherif Mohammed Jemal, Arsi University, Ethiopia

Gabriel Osei Forkuo, Transilvania University of Braşov, Romania