Cross-sensor urban land-cover characterization using hyperspectral and multispectral satellite imagery

This study establishes a cross-sensor urban remote sensing baseline for future urban digital-intelligence applications by integrating separately classified and validated multispectral (Landsat, Sentinel-2) and hyperspectral (PRISMA) imagery using a common sampling and evaluation framework. Eight land-cover classes — Dense Urban Area, Medium Urban Area, Sparse Urban Area, Tree Cover, Grass, Cropland, Bare Area, and Water — were mapped using Random Forest, Support Vector Machine (SVM), Gradient Boosting Machine, K-Nearest Neighbor, and Classification and Regression Tree algorithms. Performance was evaluated across all three sensors using overall accuracy, Cohen’s kappa, macro-averaged and class-wise F1-scores, class-area composition, spectral profile, Fisher scores, pairwise separability metrics, and principal component analysis (PCA). Results demonstrated clear sensor-dependent differences in urban representation and classification accuracy. Sentinel-2 improved spatial characterization of heterogeneous urban surfaces through enhanced spatial resolution and red-edge information. PRISMA produced the highest classification performance, with SVM achieving an overall accuracy of 91.26% and a macro-averaged F1-score of 89.72%. However, this advantage was class-specific rather than uniform across all urban-density classes. Spectral-profile and PCA analysis showed hyperspectral observations substantially improved separability between Medium and Sparse Urban classes (spectral angle: 0.54° Landsat, 1.58° Sentinel-2, 4.11° PRISMA). PRISMA classifications also allocated substantially less area to Sparse Urban Area and more to Bare Area than the multispectral classifications. This redistribution may reflect improved discrimination of mixed peri-urban surfaces but cannot be considered more accurate without an independent concurrent area-wide reference map. The findings do not establish universal sensor superiority; rather classification outcomes depend jointly on sensor properties, acquisition timing, preprocessing, class definitions, and classifier choice, supporting a complementary interpretation of multispectral and hyperspectral observations.

Urban systems are experiencing unprecedented complexity driven by rapid urbanization, climate change impacts, fragmented land-use transitions, peri-urban expansion, and accelerating infrastructure growth. These pressures will be exacerbated by a projected 30% increase (from 8 billion to 10.4 billion) in global population by the end of this century (United Nations, 2024; FAO, 2022). Therefore, the continued expansion of built-up land and population agglomeration will adversely affect urban infrastructure, public health, education, safety, and ecosystem services (Yang et al., 2023). Together, these physical, environmental, and social stressors reshape contemporary urban landscapes and create substantial challenges for planning, monitoring, and sustainable governance (Blaschke et al., 2014). Accordingly, reliable urban land-cover information is essential for understanding urban growth dynamics, impervious surface expansion, urban heat island (UHI) development, vegetation fragmentation, environmental degradation, and infrastructure pressures (Arshad et al., 2022; Akiner and Ghasri, 2025; Barcellos-Paula et al., 2025). The conversion of green spaces into built-up areas reduces tree cover and cropland, contributing to intensified UHI effects and increased cooling demands, while the loss of permeable ground elevates runoff and flood risk during storms (Arshad et al., 2022; Padmanaban et al., 2019).

Accurate characterization of urban-density gradients and transitional peri-urban environments has therefore become increasingly important for resilient urban planning and climate-responsive governance, particularly in support of inclusive, safe, resilient, and sustainable cities (Meerow et al., 2016; Dodman et al., 2022; Barcellos-Paula et al., 2025). Urban remote sensing has evolved from moderate-resolution multispectral observations toward high-resolution multispectral and hyperspectral sensing, substantially improving the characterization of complex urban land-cover patterns. Landsat, with its long-term archive since 1972, 30 m spatial resolution, and 16-day revisit cycle, has provided the foundation for monitoring urban growth and land cover and land use (LCLU) change over multiple decades (Chowdhury, 2024; Ouma et al., 2023; Zohaib et al., 2024). Sentinel-2 has improved this capability through finer spatial resolution, 10–20 m bands, higher temporal frequency, and 13 spectral bands, including red-edge and short-wave infrared (SWIR) bands that enhance the detection of vegetation, built-up surfaces, exposed soil, and mixed urban features (Ghayour et al., 2021; Jombo and Adelabu, 2023; Zohaib et al., 2025).

Hyperspectral remote sensing has further expanded urban LCLU characterization by providing high spectral detail through narrow and adjacent spectral bands that enable the identification of subtle differences among urban materials, vegetation types, bare surfaces, and mixed pixels (Moharram and Sundaram, 2023; Lou et al., 2025). For example, PRISMA (PRecursore IperSpettrale della Missione Applicativa), a recently launched hyperspectral sensor, provides 239 spectral bands across the visible and near-infrared (VNIR) and short-wave infrared (SWIR) regions (Schiavon et al., 2023). Other recent and forthcoming hyperspectral missions, including the EnMAP (Environmental Mapping and Analysis Program), CHIME (Copernicus Hyperspectral Imaging Mission for the Environment), and NASA’s SBG (Surface Biology and Geology) missions, further indicate the growing importance of fine spectral details for precise mapping and monitoring applications across diverse landscapes (Valentini et al., 2023; Schiavon et al., 2023; Chabrillat et al., 2024; Nikolaou-Alavanou et al., 2026). Several studies have demonstrated the potential of PRISMA for LCLU mapping, wildfire assessment, geological investigation, cryosphere studies, methane monitoring, and crop discrimination (Delogu et al., 2023; Vavassori et al., 2023; Oxoli et al., 2023; Perretta et al., 2024; Gaur et al., 2025; Tuzcu Kokal et al., 2022; Rogers et al., 2024; Amici and Piscini, 2021; Cavalli, 2023; Guanter et al., 2021; Kokhanovsky et al., 2022; Mirzaei et al., 2024; Shaik et al., 2022). However, assessments of its applicability for urban land-cover characterization and extraction of urban surface materials from spectral information remain limited, particularly in studies linked to resilience-oriented urban planning (Vavassori et al., 2023; De Luca et al., 2025). A small but growing body of work has begun to address this gap: hyperspectral band-selection approaches have been used to characterize urban roof materials (Nimbalkar et al., 2018), and a recent state-of-the-practice review highlights both the promise and the current scarcity of urban-specific hyperspectral reference data and classification studies (Salcido and Laefer, 2024). Building on this emerging but still limited literature, the novelty of the present study is therefore not the use of hyperspectral imagery in urban areas per se, but a systematic cross-sensor evaluation of urban-density gradients under a harmonized classification framework. Different EO systems contribute distinct but complementary sensing dimensions that collectively improve urban monitoring and evidence-based planning.

Although multispectral and hyperspectral EO datasets are increasingly used for urban applications, limited research has comparatively and systematically evaluated how Landsat, Sentinel-2, and PRISMA independently and collectively characterize urban land-cover patterns. Existing studies typically focus on single-sensor assessments or binary sensor comparisons (e.g., Landsat and Sentinel-2; Jombo and Adelabu, 2023; Ghayour et al., 2021), without incorporating the emerging hyperspectral dimension. The application of new-generation spaceborne hyperspectral imagery for detailed urban land-cover characterization remains scarce, especially in studies explicitly supporting resilience-oriented planning (Delogu et al., 2023; Vangi et al., 2021). Furthermore, limited research has framed cross-sensor urban classification within the broader concept of urban digital intelligence, where multiple EO sensing dimensions contribute complementary evidence for resilient urban planning, climate adaptation, and environmental decision-making aligned with SDGs. In this study, urban digital intelligence refers to the use of harmonized cross-sensor LCLU products as baseline geospatial inputs for future urban monitoring and decision-support applications.

Moderate-resolution multispectral datasets may generalize heterogeneous urban patterns and introduce mixed-pixel effects, particularly in fragmented peri-urban environments. Conversely, hyperspectral satellite imagery provides finer spectral resolution but faces limitations related to spatial detail, temporal frequency, and operational complexities due to data volume, spatial coverage, and computational demands (Abbas et al., 2021; Schiavon et al., 2023; De Luca et al., 2024; Lou et al., 2025; Gaur et al., 2025; Vaddi et al., 2025). As a result, urban land-cover patterns may be differently represented across sensors depending on sensor characteristics and processing methods. Therefore, understanding these sensor-dependent differences and information extraction approaches is essential for developing robust monitoring frameworks.

Therefore, this study conceptualizes cross-sensor urban land-cover characterization as a foundational component of multidimensional urban intelligence rather than solely a single-sensor classification task. It conducts a systematic three-way comparison of Landsat, Sentinel-2, and PRISMA for urban land-cover characterization within a unified framework designed to support resilient urban planning. Specifically, the study aims to: (i) compare the classification performance of Random Forest, Support Vector Machine, Gradient Boosting Machine, K-Nearest Neighbor, and Classification and Regression Tree classifiers across Landsat-8, Sentinel-2, and PRISMA imagery; (ii) quantify the spectral separability of Dense Urban, Medium Urban, and Sparse Urban classes across the three sensor systems using spectral-angle, Euclidean-distance, and centroid-distance metrics; and (iii) assess how sensor characteristics and classifier choice influence the discrimination of transitional urban-density classes. The resulting cross-sensor LCLU products are intended to establish a baseline for future urban digital-intelligence applications when integrated with additional environmental, socioeconomic, infrastructure, and planning information.

Description of the study area

The study area focuses on a highly urbanized and rapidly transitioning area within Lahore District, one of South Asia’s fastest-growing metropolitan regions (Figure 1) (Zafar et al., 2024). The landscape includes a dense urban core, expanding peri-urban zones, agricultural, vegetated, bare-surface, and riverine environments. Together, these features create a distinct urban-to-peri-urban gradient, making it a suitable testing ground for analyzing cross-sensor urban land-cover representation. The available PRISMA coverage was selected because it captured these contrasting land-cover conditions and urban-density transitions within the same spatial extent (Figure 1). The study area is centered at approximately 31.52° N, 74.36° E, covering approximately 1,789 km 2 and encompassing most of the urban extent of District Lahore. The region of interest has undergone profound transformation toward a rapidly expanding metropolis, with significant conversion from agricultural and natural land to residential and industrial zones (Zafar et al., 2024). The conversion of green spaces into built-up areas has reduced tree cover and cropland (Arshad et al., 2022), thereby contributing to intensified urban heat island (UHI) effects. The urban landscape exhibits substantial spatial heterogeneity associated with residential expansion, infrastructure development, peri-urban transition zones, fragmented vegetation patches, and mixed LCLU patterns. Such characteristics produce significant spectral and spatial complexity, especially for medium and sparse urban-density classes where built-up surfaces are intermixed with vegetation, exposed soil, and agricultural fields, thereby making the study area suitable for evaluating cross-sensor representation of urban land-cover.

A brief overview of methodology

This study employed a harmonized cross-sensor evaluation framework (Figure 2) to evaluate urban land-cover characterization across Landsat-8, Sentinel-2, and PRISMA imagery over the rapidly urbanizing metropolitan area (Table 1; Figure 1). Following sensor-specific preprocessing and a unified eight-class LCLU sampling design (Table 2), five machine-learning classifiers were independently trained for each sensor dataset. We evaluated classification outputs using accuracy metrics, cross-sensor class-area estimation, and spectral separability analysis to interpret how sensor characteristics influence urban landscape representation (Figure 2).

Satellite data and preprocessing

This study used Landsat-8 OLI, Sentinel-2 MSI, and PRISMA imagery for cross-sensor urban LCLU classification (Table 1). All datasets were downloaded as surface reflectance products and harmonized using common sampling sites to support spatial correspondence and cross-sensor comparability before classification. To reduce spatial mismatch across sensors, Landsat-8 OLI was used as the reference image for co-registration with Sentinel-2 MSI and PRISMA images. Geometric alignment accuracy was evaluated using independent ground-reference checkpoints distributed across the study area; the post-registration assessment yielded an overall planimetric RMSE of approximately 5 m, equivalent to 0.17 pixels for Landsat-8 and PRISMA at 30 m spatial resolution and 0.50 pixels for Sentinel-2 at 10 m spatial resolution, with no systematic directional displacement observed between sensors. The main preprocessing steps included cloud screening, mosaicking, scaling or converting to reflectance, geometric alignment, band selection, and, for PRISMA, removal of noisy bands with subsequent spectral smoothing. Landsat-8 and Sentinel-2 surface reflectance imagery acquired during February-March 2020 were selected (Table 1). The most suitable PRISMA hyperspectral scene covering the urban extent was available in August 2020. These images represented the closest cloud-free multispectral observations available for the study area within the same year. Because the primary objective of this study was the cross-sensor characterization of relatively stable urban structural patterns rather than seasonal change detection, images were selected to minimize phenological differences where possible by targeting peak growing seasons; nevertheless, the six-month acquisition gap between the multispectral (February/March 2020) and PRISMA (August 2020) scenes could not be fully eliminated given hyperspectral data availability. This within-year acquisition mismatch is treated here as a major limitation rather than a minor caveat: seasonal differences in vegetation condition, soil moisture, and exposed-surface reflectance may influence class-wise area estimates for Cropland, Grass, Bare Area, and Sparse Urban Area, and this constraint is revisited when interpreting cross-sensor area comparisons in the Results and Discussion.

Landsat multispectral imagery

Landsat-8 Operational Land Imager (OLI) surface reflectance imagery was obtained from the Google Earth Engine (GEE) platform (Gorelick et al., 2017) using the LANDSAT/LC08/C02/T1_L2 collection. Level 2 Collection 2 Tier 1 atmospherically corrected imagery was used, including seven spectral bands covering the visible (Bands 1–4; 443–655 nm), NIR (Band 5; 865 nm), and SWIR (Bands 6–7; 1609–2201 nm) regions at a spatial resolution of 30 m. The image was acquired on 15 February 2020, with minimal cloud cover (<3%). The study area is located within a single Landsat tile (path 149, row 38). Surface reflectance values were scaled using the standard scale factor of 0.0000275 and an offset of −0.2 (Earth Resources Observation and Science (EROS) Center, 2020).

Sentinel-2 multispectral imagery

Sentinel-2 Multispectral Imager (MSI) imagery was acquired through GEE using the image collection (COPERNICUS/S2_SR_HARMONIZED) (European Space Agency, 2021; Gorelick et al., 2017), which provides atmospherically corrected surface reflectance data. The imagery contains 13 spectral bands ranging from the visible to SWIR regions with spatial resolutions of 10 m, 20 m, and 60 m. Three images, all acquired on 03 March 2020 with less than 3% cloud cover, were mosaicked to cover the study area. Following Zohaib et al., 2025, a harmonized 10-band dataset comprising B2-B8A and B11-B12 was generated for classification. The native 20 m bands were resampled to a 10 m grid using nearest neighbor, while the native 10 m bands retained their original resolution; the 60 m bands were not included. This band-resolution harmonization may have modified local spectral variability and is therefore considered when interpreting Sentinel-2 performance.

PRISMA hyperspectral imagery

PRISMA hyperspectral imagery was accessed through the Italian Space Agency (ASI) portal (http://prisma.asi.it/). The instrument consists of a panchromatic camera (5 m GSD, 400–700 nm) and two hyperspectral spectrometers capturing 239 spectral bands from 400 to 2500 nm (66 VIS-NIR, 173 SWIR; spectral resolution <12 nm; spatial resolution of 30 m across a 30 km swath). We used the L2D product (orthorectified, geocoded surface reflectance) processed through the ASI standard chain.

The ASI employs the SRTM90_CGIAR_4.1 global DEM for geocoding; given the flat alluvial terrain of the study area, terrain-related errors were expected to be minimal (Capolupo, 2021). Two PRISMA tiles acquired on 2 August 2020 with minimal cloud cover (<5%) were mosaicked into a single raster. The HDF5 format was converted to GeoTIFF using the R library ‘prismaread’ (Busetto and Ranghetti, 2020). Geometric rectification of images was improved using Landsat-8 OLI imagery as the reference.

PRISMA preprocessing was implemented in two steps. First, bands affected by spectral noise and water vapor absorption were removed based on spectral reflectance analysis and prior literature (Flynn et al., 2020; Zhang et al., 2018), specifically removing bands between 1290 and 1490 nm, 1700–2000 nm, and 2350–2495 nm, producing a usable set containing the remaining 187 bands from the initial count of 239. Spectral noise was subsequently reduced using a Savitzky–Golay filter while retaining the principal shape of the land-cover reflectance curves. Quantitative assessment showed that the median spectral roughness decreased from 0.00141311 before filtering to 0.000058853 after filtering, corresponding to a median pixel-wise reduction of 95.89%. Figure 3 provides a qualitative illustration of PRISMA preprocessing rather than a formal before-and-after performance assessment.

LCLU classification scheme

To account for the diversity and spatial heterogeneity of urban and peri-urban environments across all three satellite datasets, eight land-cover classes were defined to enable consistent cross-sensor comparison and interpretation (Table 2). The classification scheme was designed to capture gradients of urban intensity as well as major vegetation and surface types relevant to urban environmental monitoring and resilient urban planning. The urban landscape was represented using three intensity-based classes: Dense Urban Area, Medium Urban Area, and Sparse Urban Area. These samples were assigned to their particular classes through visual interpretation of high-resolution Google Earth Pro imagery at scales of 1:500–1:2,000. Classification was based on the continuity and concentration of built-up structures, building spacing, dominance of impervious surfaces, and the relative presence of vegetation, exposed soil, and open spaces.

Dense Urban Area comprised continuous, closely spaced built-up structures with minimal vegetation or open land; Medium Urban Area contained moderately compact development interspersed with fragmented vegetation and open spaces; and Sparse Urban Area consisted of isolated or widely spaced buildings surrounded predominantly by vegetation, exposed soil, cropland, or other open land. Vegetated surfaces were classified into Tree Cover, Grass, and Cropland to distinguish differences in vegetation structure, canopy density, and land use. Additional surface classes included Bare Area, representing exposed soil and sparsely vegetated surfaces, and Water, comprising rivers, ponds, reservoirs, and other permanent water bodies.

The harmonized classification framework was consistently applied across Landsat, Sentinel-2, and PRISMA datasets to facilitate comparative evaluation of multispectral and hyperspectral sensor capabilities for urban land-cover characterization.

Training and validation design

Reference samples were labelled by a single interpreter before model development using high-resolution Google Earth Pro imagery (1:500–1:2,000), spectral profiles, and field data collected in August 2020. Ambiguous, strongly mixed, and class-boundary samples were excluded. Because only one interpreter was involved, inter-interpreter agreement was not assessed. Labels were finalized before the split and were not revised after examining model predictions; validation labels were therefore independent of model outputs, although they were not assigned by a separate interpreter. Identical reference locations were used across all sensors, with spectral values extracted separately. A total of 1,712 reference sample points were distributed across all eight land-cover classes and split into training and validation sets at a 70:30 ratio with a fixed random seed of 42, resulting in 1,198 training samples and 514 validation samples. The same fixed partition was used throughout the PRISMA workflow, and no validation sample contributed to PCA estimation or classifier training. The split was stratified by class, with a minimum 30 m exclusion distance maintained between the two subsets. The same buffered point-level partition was applied across all three sensors to ensure consistent cross-sensor comparison. For cross-sensor comparability, the same set of land-cover classes, accuracy metrics, and validation approaches were applied across the three sensors, with classification performance assessed in terms of overall accuracy, producer’s accuracy, user’s accuracy, kappa statistic, and macro-averaged F1-score. Classifier implementation followed a sensor-aware framework. For PRISMA, hyperparameters were optimized iteratively using approaches from hyperspectral machine-learning investigations (Mountrakis et al., 2011). To maintain methodological consistency and operational comparability, Landsat and Sentinel-2 classifications were implemented using established Google Earth Engine workflows. This sensor-aware implementation was intended to compare existing operational workflows rather than apply the same hyperparameter settings across substantially different feature spaces. Therefore, the observed performance differences are the combined effects of sensor characteristics, preprocessing, feature construction, and classifier implementation. Given the high dimensionality of PRISMA hyperspectral imagery, the Hughes phenomenon, the decline in classification performance caused by excessive feature dimensionality relative to available training samples, was addressed through principal component analysis (PCA)-based dimensionality reduction prior to classification (Alonso et al., 2011).

Machine-learning classification framework

Five machine-learning classifiers were implemented for each sensor dataset: Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting Machine (GBM), K-Nearest Neighbors (KNN), and Classification and Regression Tree (CART). These classifiers were selected because of their widespread application in urban remote sensing and their ability to model nonlinear and high-dimensional feature spaces (Abdollahi et al., 2023; Zohaib et al., 2024; Zohaib et al., 2025).

For PRISMA, multiple candidate hyperparameter configurations were empirically evaluated using the training-derived PCA features. Each configuration was fitted using the 1,198 training samples and evaluated using the fixed set of 514 validation samples. The final configuration was selected primarily according to validation macro-F1-score, with accuracy considered as a secondary metric. Candidate configurations were compared using internal cross-validated accuracy and macro-F1-score, after which the selected models were assessed using the fixed validation samples. The optimized SVM used an RBF kernel with C = 420 and gamma = 0.19. RF used 1,040 trees, entropy-based splitting, a maximum depth of 16, log 2 feature sampling, a minimum split size of 10, balanced-subsample class weighting, and no bootstrap sampling. GBM used 360 trees, a learning rate of 0.091, a maximum depth of 6, log 2 feature sampling, and a subsampling rate of 0.78. KNN used k = 10, distance weighting, and Manhattan distance. CART used log-loss splitting, a maximum depth of 12, a maximum of 150 leaf nodes, a minimum split size of 23, and a minimum leaf size of 3. For Landsat-8 and Sentinel-2, the same classifier families were implemented in Google Earth Engine using identical training and validation assignments. RF used 300 trees; SVM used a linear kernel; GBM used 200 trees with a shrinkage rate of 0.05; KNN used k = 7; and CART used a maximum of 100 nodes.

Feature-space construction and accuracy assessment

For Landsat, seven spectral bands (visible through SWIR) served as the input feature space. For Sentinel-2, the combined 10-band dataset including VNIR, red-edge, and SWIR bands at a harmonized spatial resolution was employed. For PRISMA, the 187-band hyperspectral cube obtained after bad-band removal and Savitzky–Golay smoothing formed the basis of feature-space construction. The reference samples were divided into fixed training and validation subsets before principal component analysis (PCA). PCA was fitted exclusively to the 1,198 training spectra, and the first 15 principal components, explaining 99.9% of the training-set variance, were retained as classifier inputs, following established practice for managing hyperspectral dimensionality (Kumar and Garg, 2023). The mean vector and component loadings estimated from the training spectra were applied without refitting to the 514 validation spectra and the complete preprocessed PRISMA cube. Therefore, the validation samples and other image pixels did not contribute to estimating the PCA transformation. All five PRISMA classifiers were trained using these 15 components, identical sample assignments, and a fixed random seed of 42. Classification performance was evaluated using overall accuracy (OA), Cohen’s Kappa (κ), Producer’s Accuracy (PA), User’s Accuracy (UA), F1-score per class, and macro-averaged F1-score (Table 3). These metrics were derived from the confusion matrix using True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN), whereas OA and Cohen’s Kappa were calculated from the complete confusion matrix.

Spectral-profile and PCA analysis

Spectral-profile analysis was performed to assess the separability of Dense Urban Area, Medium Urban Area, and Sparse Urban Area observations for Landsat, Sentinel-2, and PRISMA. The pairwise separability was quantified using three complementary metrics: spectral angle was calculated as the angular difference between mean spectral vectors in n-dimensional space; smaller angles correspond to greater spectral similarity (Masoumi et al., 2016; Kruse et al., 1993); Euclidean distance between mean class profiles; and standardized centroid distance, which normalizes inter-class separation with respect to within-class variance (Rossiter et al., 2020). Fisher-score analysis was used to identify the bands or spectral regions most useful for separating urban-density classes, where the Fisher score quantifies the ratio of between-class to within-class variance for each spectral feature and higher scores indicate stronger class-discrimination power. To compare urban-density separability, we applied PCA separately to Dense Urban, Medium Urban, and Sparse Urban spectral samples extracted from each sensor (Ping et al., 2021; Jaiswal et al., 2021). This exploratory analysis was distinct from the 15-component PCA feature space used to classify PRISMA. The class separation was visualized using the first two components, and the total variance explained by PC1 and PC2 was reported for each sensor.

Harmonization and cross-sensor area estimation

An integrated preprocessing and evaluation strategy was adopted to compare Landsat, Sentinel-2, and PRISMA classifications. Common land-cover classes, classification metrics, validation procedures, and interpretation methods were employed across all sensor studies. Before cross-sensor area estimation, all classified rasters were reprojected to WGS 84/UTM Zone 43N (EPSG:32,643), using nearest-neighbor resampling in order to preserve categorical class labels. This post-classification reprojection was performed solely for area estimation and did not alter the 10 m feature resolution used for Sentinel-2 classification. Class-wise areas were quantified using projected pixel counts and pixel areas, and all valid mapped areas were verified for consistency across sensors. The resulting area distributions were compared to evaluate the effects of sensor properties and classifier performance on spatial assignment of LCLU classes, especially for transitional urban-density classes and spectrally ambiguous non-urban surfaces.

The classification of urban land cover across Landsat-8, Sentinel-2, and PRISMA imagery demonstrates clear sensor-dependent differences in classification performance, class-area distributions, and spectral separability among urban-density gradients. PRISMA SVM achieved the highest overall accuracy across all classifiers, with SVM attaining 91.26% overall accuracy and the macro-averaged F1-score of 89.72%, compared to 86.05% under Random Forest for Landsat and 83.14% for Sentinel-2. Improvements were most pronounced for transitional urban-density classes, where Medium Urban Area F1-scores increased from approximately 47% in both multispectral sensors to 86% with PRISMA. Cross-sensor area analysis further revealed that hyperspectral observations redistributed substantial land-cover proportions from Sparse Urban Area toward Bare Area and Cropland relative to multispectral classifications, consistent with more spectrally specific characterization of mixed peri-urban surfaces, although seasonal acquisition differences, preprocessing asymmetry, and classifier behavior may also have contributed to this redistribution. Spectral separability analysis confirmed that the Medium and Sparse Urban classes, which were near-indistinguishable in Landsat with a spectral angle of 0.54°, became meaningfully separable in PRISMA at 4.11°, providing spectral evidence consistent with the observed accuracy gains.

Spatial patterns of cross-sensor LCLU classifications

Classification outputs showed broadly consistent spatial organization of dominant LCLU classes, but with noticeable differences in how individual classifiers partition urban density and spectrally similar non-urban surfaces (Figure 4). Across all three sensors, sparse and medium urban classes covered large proportions of the study area, while Water remained the smallest class. The five classifiers identified similar broad landscape structure but varied in the degree of spatial fragmentation and class allocation, particularly among Sparse Urban Area, Medium Urban Area, Bare Area, Grass, and Cropland. Landsat classifications produced relatively generalized urban structures, Sentinel-2 improved the representation of fragmented urban morphology due to finer resolution, and PRISMA classifications produced more spectrally differentiated allocation of transitional urban surfaces (Figure 5).

Figure 5 presents detailed insets from the best-performing classifier for each sensor (Landsat-8 RF, Sentinel-2 RF, and PRISMA SVM), alongside the complete classification maps in Figure 4.

The three inset locations are shown in the ground-reference overview (Figure 5a) and corresponding close-ups for Dense, Medium, and Sparse Urban Area (Figures 5b–d), with the same locations shown for Landsat-8 RF (Figures 5e–h), Sentinel-2 RF (Figures 5i–l), and PRISMA SVM (Figures 5m–p).

Quantification within the displayed frames revealed clear differences in mapped urban-class allocation among the three sensor-classifier pipelines. Dense Urban Area covered 945.45 ha in Landsat-8, 584.96 ha in Sentinel-2, and 976.14 ha in PRISMA. Medium Urban Area increased from 1,422.99 ha in Landsat-8 and 1,559.12 ha in Sentinel-2 to 2,081.97 ha in PRISMA. Conversely, Sparse Urban Area decreased from 2,182.68 ha in Landsat-8 and 1,643.00 ha in Sentinel-2 to 1,047.87 ha in PRISMA. Thus, within the evaluated frames, the PRISMA SVM output allocated comparatively more area to the Medium Urban class (Figure 5o) and less area to the Sparse Urban class (Figure 5p) than did the two multispectral outputs. These values represent observed mapped-area redistribution under the current sensor–classifier pipelines rather than independently validated estimates of true class areas. The differences may reflect the combined effects of spectral and spatial characteristics, acquisition timing, preprocessing, and classifier behavior.

Area composition across the sensor-classifier combinations

All three sensors produced comparable total valid mapped areas, indicating consistent masking and area computation (Figure 6). The Landsat results indicate that Sparse Urban Area is consistently the largest or among the largest classes, ranging from 469.87 km 2 under GBM to 566.81 km 2 under SVM. Medium Urban Area varies substantially among classifiers, with CART mapping the greatest extent at 435.90 km 2. Dense Urban Area is highest under GBM (351.78 km 2) and lowest under CART (263.24 km 2), indicating that urban-density partitioning is classifier-sensitive in Landsat imagery. In Sentinel-2, KNN yields the largest Sparse Urban Area estimate at 566.98 km 2. Vegetated and agricultural classes show substantial variation: Grass ranges from 163.59 km 2 under KNN to 280.19 km 2 under SVM, while Cropland ranges from 81.08 km 2 under KNN to 236.03 km 2 under CART, indicating sensitivity to mixed urban–vegetation surfaces. The training-only PCA results for PRISMA differ markedly from the multispectral results. Bare Area is the largest class across all PRISMA classifiers, ranging from 596.46 km 2 under SVM to 745.72 km 2 under CART. Most notably, Sparse Urban Area is substantially smaller in PRISMA, ranging from 125.30 km 2 under RF to 161.65 km 2 under KNN, compared with 469.87–566.81 km 2 for Landsat-8 and 425.66–566.98 km 2 for Sentinel-2. Although acquired in different months, all scenes represented periods of active cropland cover, providing broadly comparable cultivated-land conditions. The larger Sparse Urban extent in Landsat-8 and Sentinel-2 may partly reflect mixed pixels containing vegetation and exposed soil that spectrally resemble sparsely built surfaces. PRISMA’s denser spectral sampling may have supported the different allocation of these mixtures, particularly toward Bare Area and Cropland. However, this remains a plausible interpretation under the current pipeline, and the observed differences cannot be attributed solely to sensor characteristics because residual effects of crop phenology, soil moisture, preprocessing, and classifier behavior cannot be excluded.

Overall classification accuracy and classifier ranking

PRISMA SVM was the highest-performing sensor-classifier combination, achieving 91.26% OA with κ = 0.894 and F1-score = 89.72% (Figure 7). This represents a 5.21 percentage-point improvement over the best Landsat result (RF, 86.05%) and an 8.12 percentage-point improvement over the best Sentinel-2 result (RF, 83.14%). Among PRISMA classifiers, SVM outperformed RF by 0.61 points in OA and 0.28 points in macro-averaged F1-score. The accuracy results confirmed three main patterns. First, RF was the most stable classifier for Landsat and Sentinel-2, providing a defensible baseline for multispectral urban mapping. Second, under the adopted preprocessing and classification framework, PRISMA generally exceeded the multispectral benchmarks, particularly for SVM, RF, and GBM, consistent with the spectral advantage of hyperspectral observations for urban-density classification. Third, KNN and CART were the least stable classifiers across sensors, though their performance remained acceptable for specific class-level outcomes.

Class-wise accuracy and urban-density confusion

When the best-performing classifiers, RF for Landsat-8 and Sentinel-2 and SVM for PRISMA, were compared, a pronounced improvement was observed for Medium Urban Area. Its F1-score rose from 46.70% (both Landsat and Sentinel-2 RF) to 86.15% (PRISMA SVM), a gain of 39.45 percentage points (Figure 8). Sparse Urban Area improved from 68.40% (Landsat RF) and 63.70% (Sentinel-2 RF) to 82.14% (PRISMA SVM). PRISMA also substantially improved Cropland classification (F1-score = 89.6% compared with 72.70% for Landsat, 47.10% for Sentinel-2). Sentinel-2 KNN produced the lowest Cropland F1-score at 15.38%, indicating substantial difficulty in separating Cropland under this classifier–sensor combination. Mean F1-scores across all classifiers confirmed that Medium Urban Area (Landsat 46.03%, Sentinel-2 43.38%, PRISMA 81.8%) and Sparse Urban Area (Landsat 64.70%, Sentinel-2 57.99%, PRISMA 78.7%) were the most challenging urban classes for all sensors. Even in PRISMA, these classes remain less separable than Tree, Bare Area, Dense Urban Area, and Water, confirming that urban-density classes represent a gradual surface-composition continuum rather than sharply bounded spectral categories.

Spectral separability of cross-sensor urban-density classes

The spectral-profile, PCA, and separability analyses revealed substantial differences in how Landsat, Sentinel-2, and PRISMA represented peri-urban heterogeneity and transitional urban surfaces (Figures 9, 10; Tables 4, 5). Mean spectral profiles showed that all sensors captured distinct reflectance differences among Dense Urban Area, Medium Urban Area, and Sparse Urban Area, particularly within the visible, NIR, and SWIR regions (Figure 10). However, class separability varied considerably among sensors (Figure 9). Based on the overall accuracy patterns, class-wise separability, confusion matrices, spectral profiles, PCA clustering, and area-composition behavior, Table 4 summarizes the strongest-performing sensor for each land-cover class. Overall, the results demonstrate that temporal continuity, spatial detail, and spectral richness provide complementary but non-equivalent representations of urban systems. Under the adopted framework, Landsat performed strongly for different classes, Sentinel-2 provided a finer representation of spatial heterogeneity, and PRISMA provided the best discrimination for Medium and Sparse Urban classes and several mixed peri-urban surfaces.

The reported advantages in Table 4 represent the highest class-wise F1-scores obtained under the adopted preprocessing, sampling, acquisition, and classifier framework and should not be interpreted as universal sensor rankings.

Pairwise separability metrics

Spectral separability analysis revealed marked differences among Landsat, Sentinel-2, and PRISMA observations (Figure 9; Table 5). The highest-ranked Landsat features included NIR and SWIR bands, whereas the most informative Sentinel-2 features were concentrated in the NIR and red-edge regions; PRISMA exhibited greater separation across its narrow-band hyperspectral feature space. A Fisher-score analysis was conducted to identify the most informative bands and spectral ranges for separating urban-density classes across Landsat, Sentinel-2, and PRISMA; the results are presented in Table 5.

The highest individual PRISMA Fisher scores were concentrated in the narrow NIR range 809–830 nm. Because Fisher scores evaluate each band independently, these rankings identify the strongest univariate contributions but do not isolate the contribution of PRISMA SWIR wavelengths within the PCA-derived feature space. Therefore, any PRISMA SWIR contribution is interpreted as plausible rather than empirically demonstrated.

Landsat showed extremely weak separability between Medium and Sparse Urban Area, and thus strong spectral ambiguity for transitional urban surfaces. Sentinel-2 moderately increased separability due to enhanced spatial and red-edge information, although large overlap persisted. By contrast, PRISMA produced larger spectral angles, Euclidean distance, and centroid distances, especially for the Dense-Sparse Urban Area pair.

The pairwise results indicated that Medium and Sparse Urban Area are practically indistinguishable spectrally in Landsat (spectral angle 0.54°, centroid distance 0.54) and slightly better in Sentinel-2 (1.58°, 0.67). In PRISMA, the separation improved substantially, with a spectral angle of 4.11° and a centroid distance of 6.13. Dense–Sparse separation reached a spectral angle of 7.67° and a centroid distance of 19.91 in PRISMA, indicating that, under the current pipeline, PRISMA exhibited stronger measured separation of the urban-density gradient; however, this result does not establish that SWIR bands were responsible for the improvement.

PCA-based spectral structure of urban clusters

The PCA scatter plots visually support these separability results. Dense Urban Area in Landsat and Sentinel-2 feature spaces is generally distinct, but Medium and Sparse Urban Area occupy overlapping regions (Figure 10). Compared with multispectral observations, PRISMA hyperspectral observations enable better relative separation among the three urban-density classes, although Medium Urban Area and Sparse Urban Area still show partial overlap in PCA space. Greater PC2 variance was observed for PRISMA (21.72%) than for Landsat (13.55%), indicating that a larger proportion of the PRISMA spectral variability was represented along the second component; however, Medium and Sparse Urban samples remained partially overlapping. Abrupt breaks visible in the PRISMA spectral profiles (Figure 10c) correspond to the removal of noisy and water-absorption bands during preprocessing (Section 2.3.3) and do not represent unexplained physical spectral behavior.

Hyperspectral imagery enhances spectral discrimination of transitional urban landscapes

Under the adopted preprocessing and sampling framework, PRISMA produced the highest measured classification accuracy and the greatest spectral separability among the three evaluated sensors, consistent with growing evidence supporting hyperspectral advantages for complex land-cover discrimination (Vaddi et al., 2025; Delogu et al., 2023; Lekka et al., 2024). The overall accuracy improvement appears to be associated with the qualitatively different representation of the urban landscape produced by PRISMA. Rather than simply producing higher urban classification accuracy, PRISMA appeared to represent transitional urban environments with greater spectral specificity. The large redistribution of pixels from Sparse Urban Area in multispectral classifications is consistent with hyperspectral information reducing the spectral ambiguity associated with sparse urban environments; however, seasonal acquisition differences, PRISMA-specific preprocessing and PCA-based feature construction, and classifier-specific behavior may also contribute to this pattern, and the interpretation should be treated as plausible rather than demonstrated pending independent validation. This has potential implications for urban resilience planning, where inaccurate identification of peri-urban surface types can lead to misallocation of green infrastructure investments or underestimation of impervious surface extent (Techapinyawat et al., 2024; Dodman et al., 2022).

The class-wise improvements in Medium Urban Area and Sparse Urban Area demonstrate the potential planning value of more reliable transitional urban-density classification. These transitional urban-density classes correspond to peri-urban zones, informal settlements, and developing areas. These transitional zones are the kind of areas where planning applications such as heat-mitigation corridor design, green-infrastructure retrofits, and densification monitoring are typically targeted (Meerow et al., 2016; Sharifi, 2020; Dodman et al., 2022), so their accurate characterization is planning-relevant; however, the present study maps land cover only and does not itself compute heat-mitigation corridors or evaluate green-infrastructure outcomes, and the outputs may support such analyses when integrated with thermal, cadastral, or demographic datasets. PRISMA provides dense spectral coverage across approximately 400–2500 nm, which may support the differentiation of materials and cover types that are confounded in multispectral imagery, particularly within mixed urban pixels (Vangi et al., 2021; Cavalli, 2023; De Luca et al., 2025; Sajadi et al., 2024; Gaur et al., 2025). Because PRISMA has the same 30 m spatial resolution as Landsat and a coarser resolution than Sentinel-2, its observed urban-density discrimination cannot be attributed to finer spatial detail (De Luca et al., 2024; Perretta et al., 2024). Instead, the observed results are consistent with dense spectral sampling providing additional information on mixed pixels containing built-up surfaces, exposed soil, and vegetation. Although the present analysis does not isolate the respective contributions of the VNIR and SWIR regions, some mixed pixels may have been reassigned to spectrally dominant non-urban classes such as Bare Area. Confirming the mechanism would require independent material-level reference data.

Landsat provides long-term urban monitoring baseline

Although Landsat showed comparatively weak spectral separability between Medium and Sparse Urban surfaces, its historical continuity remains important for urban-change analysis and regional urban dynamics. It provides stable regional urban structure representation and consistent urban expansion baselines suitable for long-term urban monitoring. The primary limitation of Landsat for urban land-cover characterization is the spectrally broad band configuration. This broad band configuration limits angular discrimination between Medium and Sparse Urban classes (spectral angle: 0.54°), two planning-relevant transitional urban classes. Despite these limitations, Landsat’s temporal archive remains critical for urban insights. The five-decade longitudinal continuity enables assessment of urbanization pathways, impervious surface patterns, and vegetation loss patterns that no other currently available sensor can provide through a single continuous historical record. For planning resilient cities, this temporal dimension may provide baseline data relevant to SDG 11.3.1 monitoring and to evaluating the long-term impact of resilience interventions, although the present single-date, three-sensor comparison does not compute SDG 11.3.1 indicators or perform multi-temporal urban-expansion analysis.

Sentinel-2 improves spatial heterogeneity and vegetation sensitivity

Sentinel-2 provides finer spatial characterization of fragmented urban morphology, vegetation distributions, and heterogeneous peri-urban patterns through its finer spatial resolution (10–20 m) and additional red-edge bands. Despite these improvements, Medium and Sparse Urban classes remained difficult to separate (spectral angle: 1.58°; standardized centroid distance: 0.67), confirming that spectral ambiguity at the medium-to-sparse urban density transition persists even at 20 m resolution with red-edge information. This large variation in Grass and Cropland area estimates across Sentinel-2 classifiers indicates that higher spatial resolution yields greater detail in some places but also likely increases the model’s sensitivity to mixed urban-vegetation boundary pixels, a well-known problem in urban remote sensing (Blaschke et al., 2014). Sentinel-2’s lower overall accuracy may reflect the combination of harmonizing its native 10 m and 20 m bands (with greater within-class variability at finer spatial resolution), scale differences between reference labels and individual pixels, and standard GEE classifier settings. Therefore, its lower accuracy under the present framework should not be interpreted as an inherent limitation of Sentinel-2 imagery. In urban-resilience applications, this vegetation sensitivity may provide useful baseline information for monitoring urban green spaces, quantifying canopy coverage, evaluating green-infrastructure health, and informing nature-based solutions for UHI mitigation (Ruiz-Valero et al., 2026; Sharifi, 2020).

Cross-sensor sensitivity of machine-learning classifiers for characterizing urban landscapes

The sensitivity of machine-learning classifiers varied across the sensors, confirming that sensor-specific classification strategies are important for operational urban monitoring. Random Forest (RF) achieved the best overall performance for both Landsat and Sentinel-2 classifications, making it the most robust and computationally efficient algorithm for multispectral datasets. Given its ensemble-based structure, computational efficiency, and relatively low sensitivity to hyperparameter optimization, RF is particularly suitable for regional multispectral urban mapping applications. The support vector machine with a radial basis function kernel (SVM-RBF) performed better for PRISMA hyperspectral imagery, where the kernel-based approach handles high-dimensional and non-linearly separable spectral feature spaces. This result implies that classifier choice should reflect sensor characteristics, rather than applying a uniform classification framework across all Earth observation datasets. The enhanced SVM-RBF performance for PRISMA is likely related to its ability to implicitly project PCA-compressed hyperspectral features into higher-dimensional decision spaces where spectrally overlapping urban classes become more separable. However, PRISMA classifiers were specifically parameterized, whereas Landsat-8 and Sentinel-2 used standard GEE settings. Consequently, PRISMA’s measured advantage cannot be separated completely from differences in optimization effort, and equivalent tuning of the multispectral classifiers could potentially narrow the observed performance gap. The support-vector optimization technique facilitates robust generalization to residual spectral noise and mixed urban spectral conditions (Mountrakis et al., 2011; Zohaib et al., 2025; Vaddi et al., 2025). Spectral separability analysis showed that the highest-ranked univariate PRISMA bands for distinguishing the urban-density classes were concentrated within the NIR interval of 809–830 nm (Table 5). Although the broader VNIR–SWIR feature space may plausibly contribute through multiband covariance in the PCA representation, the specific contribution of SWIR bands was not isolated empirically in this study. The PCA representation showed improved relative separation among the urban-density classes, although Medium and Sparse Urban samples remained partially overlapping. These findings are in line with meta-analytic evidence demonstrating that in hyperspectral classification, SVM generally outperforms RF by approximately 3%–4% in mean overall accuracy (Sheykhmousa et al., 2020), and recent EnMAP hyperspectral studies confirmed the superiority of SVM-RBF in complex land-cover discrimination tasks (Lekka et al., 2024; Gaur et al., 2025; Lou et al., 2025; Nikolaou-Alavanou et al., 2026).

Using RF, Sentinel-2 achieved an overall accuracy of 83.14%, which was lower than Landsat-8 (86.05%) within the harmonized cross-sensor framework. However, this outcome should not be viewed as an inherent restriction of Sentinel-2. In the companion Landsat–Sentinel study using a sensor-optimized nine-class scheme, Sentinel-2 achieved substantially higher overall accuracy (93.97%) compared to Landsat-8 (91.78%) (Zohaib et al., 2025). The standardized centroid distance between Dense Urban Area and Medium Urban Area increased from 2.66 for Landsat to 3.72 for Sentinel-2, corroborating the increased discrimination ability of the red-edge region. Sentinel-2 showed considerable sensitivity to vegetation variations, with variation in Grass area estimates (163.59–280.19 km 2) among classifiers, consistent with the sensitivity of red-edge wavelengths to chlorophyll concentration and vegetation structure (Zohaib et al., 2025). RF showed the strongest classification performance for Landsat-8, and previous urban multispectral classification studies reported Landsat-8 accuracies of 84%–92% (Chowdhury, 2024; Ouma et al., 2023; Jombo and Adelabu, 2023). The robustness of RF in Landsat-based urban mapping has been widely attributed to its ensemble averaging and bootstrap aggregation mechanisms, which reduce sensitivity to mixed pixels, spectral noise, and intra-class variability common in medium-resolution urban imagery (Abdollahi et al., 2023; Yasser et al., 2026; Zohaib et al., 2024).

Integrated cross-sensor EO framework for urban monitoring

The findings demonstrate that Landsat, Sentinel-2, and PRISMA provide complementary but non-equivalent urban intelligence dimensions. No single sensor captures all aspects of urban systems. Instead, integrated cross-sensor EO frameworks provide a more comprehensive representation of urban environments than any single sensor could achieve independently (Schiavon et al., 2023; Vaddi et al., 2025; Zohaib and Abbas, 2026). These sensors provide complementary information to represent the urban landscape across temporal, spatial, and spectral dimensions. This complementarity has practical implications for operational urban monitoring through cross-sensor integration. For regional and multi-temporal urban growth analysis, Landsat remains the most accessible and cost-effective solution for establishing historical trajectories and providing historical inputs relevant to future SDG 11.3.1 assessment. For detailed urban morphology characterization and vegetation monitoring, Sentinel-2 provides a useful combination of spatial resolution, spectral information, and data availability. For detailed spectral characterization of urban materials and peri-urban transition zones, hyperspectral data (current or future operational missions such as PRISMA, EnMAP, CHIME, and SBG) represent a significant advance, particularly for discriminating the urban-density gradients most relevant to planning interventions.

Implications for resilient urban planning and climate adaptation

Cross-sensor urban land-cover characterization may provide baseline information for resilient urban planning, urban climate assessment, green infrastructure monitoring, impervious surface mapping, and environmental governance. Improved characterization may support subsequent urban heat, ecological, vegetation, and infrastructure assessments when combined with the required thematic datasets (Arshad et al., 2022; Dodman et al., 2022). The ability to systematically track land-cover patterns using open-access satellite data and machine-learning models enhances the transparency, replicability, and spatial comparability of urban planning metrics, providing reproducible spatial inputs potentially relevant to future SDG 11.3.1 monitoring on land consumption efficiency (Dodman et al., 2022; Sharifi, 2020; Meerow et al., 2016).

For rapidly urbanizing South Asian cities, where peri-urban transition zones, fragmented vegetation, and bare-soil surfaces are interspersed with medium-density settlements, the spectral richness of PRISMA may improve the characterization of spectrally mixed peri-urban surfaces relevant to subsequent environmental-vulnerability assessments. The redistribution illustrates that sensor choice can influence the representation of peri-urban surfaces; however, independent validation is required before these differences are translated into planning decisions. A spatially explicit, multidimensional land-cover analysis may provide baseline information for climate-responsive urban assessment when combined with thermal, demographic, cadastral, and field data (Barcellos-Paula et al., 2025; Al-Raeei, 2025).

Toward urban digital intelligence

These findings also point toward a broader trajectory in which EO systems, cross-sensor integration, and future hyperspectral missions such as ESA’s CHIME and NASA’s SBG could support more advanced, multidimensional urban monitoring as the underlying technologies mature (Schiavon et al., 2023; Al-Raeei, 2025). Realizing this potential will likely require deep-learning architectures capable of exploiting the full spectral and spatial content of multi-sensor data (Vaddi et al., 2025), together with integration into operational cloud-computing and decision-support platforms; these remain directions for future work rather than outcomes demonstrated in the present study.

Limitations and future directions

Although the results are encouraging, several broader limitations remain relevant when considering cross-sensor urban land-cover characterization. Variation in spatial resolution, spectral dimensionality, revisit frequency, and scene coverage affects the comparability and operational scalability of multispectral and hyperspectral datasets. PRISMA demonstrated strong discriminatory capability in distinguishing complex urban-density gradients; however, current hyperspectral missions remain limited by temporal frequency and spatial coverage compared to operational multispectral systems such as Landsat and Sentinel-2. Furthermore, classification performance and urban-density patterns may differ across the city’s diverse morphological, climatic, and socio-environmental features. Differences in sensor-specific preprocessing, feature construction, and classifier optimization may also contribute to observed differences among results, despite a harmonized analysis methodology. In addition, the present study focused on traditional machine-learning methods and did not evaluate deep learning or spatial-contextual architectures that could offer additional advantages for analyzing large-scale datasets for urban feature extraction and cross-sensor representation learning. A further limitation is that the available Landsat/Sentinel-2 and PRISMA scenes were acquired in different months of 2020. Although the study focuses primarily on relatively stable urban-density patterns, seasonal differences may influence vegetation, cropland, bare soil, and sparse urban class proportions. The reference samples were not explicitly evaluated for spatial autocorrelation; therefore, some spatial dependence between training and validation samples cannot be ruled out, and future research should consider spatial block validation. Future research should investigate multi-city transferability analysis, the incorporation of advanced AI and deep learning methods, and the operational deployment of emerging hyperspectral missions to support scalable and efficient urban monitoring models for resilient urban development and climate adaptation.

This study presents a systematic cross-sensor comparison of Landsat-8, Sentinel-2, and PRISMA for urban land-cover characterization in a rapidly urbanizing metropolitan environment. The findings highlight that urban land-cover intelligence is inherently multidimensional, whereby different sensors provide complementary observational strengths rather than competing capabilities. Landsat offers long-term temporal continuity for regional urban monitoring, Sentinel-2 enhances spatial characterization of heterogeneous urban landscapes, and PRISMA demonstrates the potential value of hyperspectral observations for improving urban-density discrimination within the adopted framework, particularly for complex urban-density transitions and mixed peri-urban surfaces. The study also shows that classifier performance depends on the sensor, highlighting the importance of applying machine-learning strategies to multispectral and hyperspectral feature spaces. More generally, the findings highlight the need for resilient urban monitoring to move beyond reliance on individual sensors toward integrated cross-sensor Earth observation frameworks that combine temporal, spatial, and spectral dimensions of urban intelligence. As next-generation hyperspectral missions and AI-driven geospatial analytics continue to develop, integrated EO approaches may play an increasingly important role in supporting evidence-based urban planning, climate adaptation, green infrastructure assessment, and sustainable urban development. The proposed cross-sensor urban land-cover characterization perspective provides a transferable foundation for future multi-city and multi-temporal urban monitoring systems.