Editorial: Advancing spatial prediction of soil properties using remotely sensed data and geospatial artificial intelligence (GeoAI): challenges, opportunities, and future directions

Soils underpin food production, water regulation, biodiversity, and the terrestrial carbon cycle, yet their physical and chemical properties vary substantially across space and depth. Conventional soil surveys and laboratory analyses remain indispensable, but their costs and limited spatial density constrain frequent updating and fine-resolution decision support. Digital soil mapping addresses this limitation by relating soil observations to spatially explicit environmental covariates through quantitative prediction models (1). The rapid expansion of Earth observation, proximal sensing, cloud computing, and machine learning has extended this framework toward geospatial artificial intelligence (GeoAI), enabling nonlinear feature extraction, multisource data integration, and scalable prediction. Nevertheless, stronger algorithms do not remove the need for representative sampling, spatially appropriate validation, uncertainty quantification, reproducibility, and pedological interpretation (2–4).

This Research Topic brings together six studies addressing these interconnected requirements across humid, arid, semi-arid, mountainous, and coastal environments. The contributions encompass automated sampling design, transfer learning, ensemble prediction, laboratory spectroscopy, and satellite hyperspectral mapping. Collectively, they demonstrate that progress in soil-property prediction depends less on identifying a universally superior model than on establishing coherent workflows in which sampling, predictor selection, model structure, validation, and uncertainty assessment are appropriately matched to the target soil property and environmental setting.

Rodrigues et al., in “AutoRA: an innovative algorithm for automatic delineation of reference areas in support of smart soil sampling and digital soil twins”, address the foundation of every predictive soil-mapping workflow: the quality, representativeness, and efficiency of field sampling. Their autoRA algorithm employs Gower’s dissimilarity index to delineate environmentally representative reference areas automatically. By evaluating different reference-area sizes and spatial resolutions using simulated soilscapes for Florida, United States, and Rio de Janeiro, Brazil, the study provides a reproducible framework for examining trade-offs among sampling extent, cost, environmental representation, and prediction accuracy. This contribution is particularly relevant to the development of digital soil twins, for which continuously updated spatial representations require efficient strategies for acquiring new field observations.

Han et al., in “A soil organic carbon mapping method based on transfer learning without the use of exogenous data”, address another persistent limitation: the scarcity of labeled soil observations needed to train deep-learning models. Their approach exploits relationships among soil depths by first pretraining a convolutional neural network using observations from all available depth intervals and subsequently fine-tuning the network for a specific target layer. The results demonstrate that information already contained within a soil-profile dataset can support transfer learning without requiring an unrelated external dataset. More broadly, the study illustrates how pedologically meaningful relationships among soil layers can be incorporated into model development to improve data efficiency.

Two contributions investigate ensemble learning in environmentally heterogeneous Peruvian landscapes. Carbajal-Llosa et al., in “Ensemble machine learning for digital mapping of soil pH and electrical conductivity in the Andean agroecosystem of Peru”, integrated support vector machines, artificial neural networks, random forest, and XGBoost through simple and performance-weighted averaging. The weighted ensemble achieved R² values exceeding 0.93 and reduced root mean square error by approximately 72%, while elevation emerged as the most influential predictor of both soil properties. Importantly, prediction uncertainty was lower for pH than for electrical conductivity, demonstrating that map reliability remains property-specific even when the same observations, environmental predictors, and modeling framework are used.

Salazar-Coronel et al., in “Soil organic carbon content mapping along the coast of northern Peru: an ensemble machine learning approach”, integrated topographic, climatic, edaphic, and Sentinel-2-derived vegetation variables to map soil organic carbon in the Motupe River Basin. XGBoost produced the strongest individual performance, with an R² of 0.83. The study also employed spatial block cross-validation, bootstrapping, and 90% prediction intervals to evaluate spatial generalization and uncertainty. This represents an important methodological strength because random cross-validation can produce overly optimistic accuracy estimates when spatially neighboring and environmentally similar observations are assigned to both model-training and model-testing subsets (5).

The remaining studies illustrate the expanding role of hyperspectral sensing. Mustafa et al., in “Hyperspectral-derived spectrotransfer functions for soil properties estimation in arid lands”, used laboratory visible–near-infrared spectroscopy on 188 soil samples from Qena Governorate, Egypt, to estimate sand, silt, clay, and plant-available nitrogen, phosphorus, and potassium. Partial least-squares regression generally outperformed the simplified spectrotransfer functions developed from selected wavelengths, although the latter retained useful predictive ability for several attributes. The results support laboratory spectroscopy as a rapid, non-destructive complement to conventional soil analysis while also demonstrating that spectral predictability differs considerably among soil properties.

Mosaid et al., in “Soil Texture Classification and Mapping in Semi-Arid Regions Using Machine Learning and EnMAP Hyperspectral Data”, extend hyperspectral soil assessment from laboratory measurements to satellite observations. They compared random forest, XGBoost, support vector machines, and k-nearest neighbors under scenarios incorporating EnMAP hyperspectral bands, spectral indices, and terrain attributes. Support vector machines achieved the strongest results, reaching an overall accuracy of 75% and a precision of 79.2% when hyperspectral bands alone were used. The finding that additional indices and terrain variables did not consistently improve performance is instructive: increasing predictor numbers does not necessarily improve a model when variables are redundant, spatial scales are mismatched, or training samples are limited.

Four broad priorities emerge from this Research Topic. First, sampling design and model development should be treated as components of a single optimization problem because environmental coverage and sample density define the domain within which predictions can be trusted. Second, spatially structured validation, independent geographic testing, and explicit evaluation of model extrapolation should become standard practices (5, 6). Third, uncertainty and area-of-applicability maps should accompany soil predictions so that users can distinguish well-supported estimates from locations requiring additional observations. Fourth, future GeoAI systems should prioritize interpretability, transferability, and reproducibility alongside predictive accuracy. Promising directions include pedology-informed learning, domain adaptation across sensors and regions, multiscale fusion of laboratory, proximal, UAV, and satellite observations, and open benchmark datasets supported by harmonized validation protocols and, most critically, moving beyond purely statistical spectral–property correlations toward physics-informed hybrid frameworks that couple particulate-scale radiative transfer with machine learning, with predictions constrained by known mineralogical and dielectric bounds.

These studies demonstrate that remotely sensed data and GeoAI can substantially improve efficiency, spatial resolution, and repeatability of soil information. However, their broader message is that operational soil mapping requires more than algorithmic sophistication. Reliable products emerge when representative observations, appropriate covariates, transparent models, spatially defensible validation, quantified uncertainty and openly shared data are assembled into a coherent workflow. Advancing these elements jointly will be essential for translating research-grade predictions into dependable tools for precision agriculture, soil conservation, carbon management, and climate-resilient land-use planning.

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.

Footnotes

Edited and reviewed by: Jeroen Meersmans, University of Liège, Belgium