Accurate land-cover classification in heterogeneous and fragmented agro-natural landscapes represents a key challenge for environmental monitoring and territorial planning, particularly in complex ecosystems such as native grasslands. In this study, we evaluated the comparative performance of approaches based on multitemporal optical predictors derived from Sentinel-2 and AlphaEarth embeddings for supervised land-cover classification in the Tandilia mountain system, Argentina. We implemented reproducible workflows entirely within Google Earth Engine, using exclusively open-access satellite data. We defined eight predictor configurations that combined optical, phenological, and topographic information with advanced satellite representations, and trained Random Forest models using a spatial block–based design with repeated stratified sampling across land-cover classes. Model performance was evaluated using an external, spatially independent reference dataset, employing global and class-wise performance metrics, as well as analyses of model stability and map-level spatial consistency. Results show that, while Sentinel-2–based approaches achieved high accuracy (OA ≈ 89–93%; Macro-F1 ≈ 88–90%), AlphaEarth-based models exhibited consistently superior performance (OA ≈ 94–96%; Macro-F1 ≈ 94–95%), lower variability across runs, and a marked reduction in confusion for spectrally complex classes, particularly native grasslands and water bodies. These improvements were reflected in both class-wise accuracy metrics and enhanced spatial coherence, with reduced classification noise at the map level. Overall, under the regional conditions evaluated here, our findings highlight the potential of AlphaEarth embeddings to improve land-cover classification in complex and fragmented landscapes, while also emphasising the trade-off between improved classification performance and the limited biophysical interpretability of latent embedding dimensions.
Cage aquaculture is expanding rapidly in Lake Victoria, yet routine monitoring remains constrained by a lack of repeatable, scalable, and cost-effective alternatives to traditional manual field surveillance. We evaluated two remote sensing workflows, object-based image analysis with a random forest classifier (OBIA-RF) and the Segment Anything Model (SAM), to map spatiotemporal cage aquaculture dynamics between 2019 and 2024. We used a two-stage framework that combines Sentinel-2 multispectral imagery for water-zone delineation with high-resolution Google Earth Pro images for cage-scale extraction. We quantified expansion, abandonment, and technological shifts leading to intensification. Results indicate a 253% increase in the horizontal cage footprint with spatially heterogeneous patterns. Bathymetric analyses revealed a nearshore bias and a depth-associated pattern. Abandonment was concentrated in shallower sites (mean depth: 4.9 m) and expansion occurred in deeper waters (mean depth: 9.2 m). 70% of new expansion remained below the 10 m sustainability recommendation set by the Kenya Marine and Fisheries Research Institute (KMFRI). Spatial overlays revealed that cages directly encroached onto designated Sustainable Aquaculture Water Area (SAWA) fish breeding zones, proving economic accessibility currently outweighs regulatory zoning. For the transition to cage aquaculture to be sustainable and equitable, coastal governance must be coupled with targeted credit facilities to help smallholders overcome the seven-fold capital cost increase required for offshore infrastructure, preventing concentration in ecologically vulnerable nearshore zones. This study provides a transferable framework for monitoring aquaculture cages in lakes across Sub-Saharan Africa. It demonstrates that deep-learning models offer a robust solution for monitoring cages in data-scarce environment.
Mountain pine (Pinus mugo) is a critical species in the krummholz zone of alpine ecosystems and vital to understanding treeline dynamics under ongoing climate change. Despite their various ecosystem functions, their spatial distribution remains poorly quantified across complex mountain terrain. In this study, Sentinel-2 multi-temporal optical imagery, along with topographic and textural features, is used to train a random forest model and map the species in southern Germany at 10-meter resolution. The classification achieves an F1 score of 87.5–90.4% under spatially independent validation, and was applied to generate annual mountain pine distribution maps from 2017 to 2025. Applying the approach to higher-resolution UAS data acquired near Zugspitze in Summer 2025 highlights the sensor-scale limitation. This is evidenced by underdetection in the Sentinel-2 maps at higher elevations, particularly in landscapes with smaller, fragmented patches, due to mixed pixels at the Sentinel-2 scale. The annual mountain pine time-series maps highlight patch dynamics, distribution in relation to terrain, and a detectable upslope shift in the upper boundary of the distribution within the German part of the European Alps. To demonstrate the ecological relevance of these maps, an exploratory fixed-effects analysis suggests that greenness is influenced mainly by growing-season and early-winter temperature, pointing to an energy-limited system favoured by warming conditions. Our results highlight the potential of high-resolution species-level mapping using machine learning in data-limited regions. The approach provides scalable workflows for other species in alpine ecosystems, supporting robust monitoring amid ongoing environmental change.
The detection and monitoring of oil spills at sea (OSS) are essential for environmental disaster management and the protection of marine ecosystems. The objective was to integrate scattering coefficient (σ0) data from Sentinel-1 (S1) with multispectral images from Sentinel-2 (S2) for OSS detection, as in the case that occurred on January 15, 2022, at the La Pampilla refinery in Ventanilla, Lima, Peru. To achieve this, the σ0 values in VV polarization from S1-GRDH were analyzed, establishing characteristic threshold ranges of -21 dB to -13 dB for fresh oil (January 25) and -19 dB to -13 dB for weathered oil (February 2). This 3 dB shift in the median quantifies the weathering and emulsification processes of the hydrocarbon. The methodology employed includes the calculation of spectral indices (OSI, VNRI, IVI, Zakzouk) and the evaluation of S2 spectral signatures. The OSI, with a range of 1.1 to 1.5, achieved the highest agreement (κ = 0.808, “Very good”), establishing itself as the most reliable indicator for operational binary detection. Although OSI is optimal for binary detection, net area quantification was performed by spectral unmixing, yielding 6,818 ha (∼68 km2). Validation using the Kappa index and Mann-Whitney-Wilcoxon U test (p<0.05) demonstrated statistical significance and confirmed the spectral homogeneity of the signatures. The findings include this oil spill map, which establishes the Ventanilla spill as the largest ecological disaster in Peruvian history. This methodological framework is effective for the studied event and provides a reference that can be adapted to future oil spill events, contributing to the development of capabilities for rapid response to marine environmental disasters.
Faced with growing water scarcity, the delineation of groundwater potential zones (GWPZ) using remote sensing and machine learning techniques has become an indispensable tool. This article critically reviews the evolution and application of these techniques between 2000 and 2025. We synthesize prevailing approaches, document progress, address critical challenges, and identify urgent research opportunities to guide future sustainable groundwater exploration and management. Our review demonstrates that the advent and continuous advancements of remote sensing, providing critical long-term data, have significantly propelled the field of GWPZ. The availability of various multi-spectral sensors and sophisticated models has been instrumental in this progress. Additionally, emerging trends such as machine learning techniques (e.g., Random Forest and Support Vector Machine) and cloud computing platforms enabled effective multi-source data fusion, enhancing predictive accuracy and overcoming limitations of individual sensors. However, significant challenges persist. The scarcity of ground-based data impedes robust model validation, while a lack of high-resolution spatial and temporal data hinders precise and small-scale delineation. This is particularly acute in Sub-Saharan Africa, which remains critically understudied. To address these gaps, we recommend that future research work focus on using spatial cross-validation to achieve accurate results, assess data-scarcity solutions including data augmentation, use advanced approaches such as explainable artificial intelligence, hybrid and ensemble machine learning models to improve delineation precision and integrate these approaches into broader water resource resilience frameworks.
The urban thermal environment faces significant challenges during rapid urbanization, with intensifying extreme heat events combined with the urban heat island effect (UHI) threatening public health and environmental sustainability. As crucial green infrastructure, urban parks demonstrate notable cooling impacts contributing to UHI mitigation. This study advances prior work by quantifying nonlinear interactions between park landscapes and cooling effects under contrasting climatic conditions (extreme hot vs. normal weather), with a focus on dual-perspective (internal park and external surrounding) landscape patterns. The relationships between park cooling effects and landscape patterns were investigated using an eXtreme Gradient Boosting (XGBoost) model incorporating satellite-derived landscape metrics. The XGBoost model demonstrated strong predictive accuracy, with the coefficient of determination (R2) of the verification subset being 0.912 under normal days and 0.926 during extreme hot days. Analysis revealed a higher average cooling intensity of parks during extreme hot days (2.28°C) compared to normal conditions (1.86°C). It is noteworthy that the study revealed the drivers of park cooling vary between normal and extreme hot weather—external built-up coverage dominates under both conditions but with greater contribution during extreme heat, whereas water-related variables show significant effects only under normal conditions. These findings inform landscape optimization strategies for urban parks to mitigate climate change and urbanization challenges, especially for increasingly frequent urban extreme heat events.
Satellite-derived vegetation indices, derived from temporally dense time series, are widely used for monitoring post-landslide vegetation recovery modeling. However, integrating observations from multiple satellite sensors often introduces inconsistencies due to differences in spectral response, radiometric calibration, and spatial resolution. These discrepancies may affect the reliability of long-term vegetation monitoring and recovery modeling. This study develops a cross-sensor NDVI time series harmonization framework to generate a consistent, improved-resolution multi-sensor NDVI time series, and evaluate vegetation recovery following the Typhoon Morakot (2009) disturbance in Maolin District, southern Taiwan. The proposed framework integrates NDVI observations from Landsat 5, Landsat 8, and Sentinel-2 using three time series normalization approaches: global Range Preserving Affine Transformation Model (RPAM), pixel-based RPAM, and pixel-based standard deviation RPAM. The harmonized Sentinel-2 NDVI time series, merged with the Landst 5 and Landsat 8 NDVI time series, was subsequently used to model vegetation recovery trajectories by fitting a logarithmic recovery model to post-disturbance NDVI time series. Comparative evaluation using statistical metrics and spatial model performance indicates that the standard deviation–based RPAM approach provides the most consistent cross-sensor integration, producing the lowest RMSE (0.0548) and MAE (0.0043) compared with other approaches. This model performance further shows that the harmonized NDVI time series improves the reliability of recovery modeling, with coefficient of determination (R2) values reaching up to 0.96 across several landslide areas. An estimated recovery of about 22.5% is predicted by the model till 2025, while about 48% area sizes are expected to recover by 2050. The results demonstrate that harmonized multi-sensor NDVI time series significantly improve the reliability of long-term vegetation monitoring and enable spatially explicit estimation of vegetation recovery dynamics. The proposed framework provides a practical approach for integrating multi-sensor satellite observations and supports improved assessment of post-disturbance ecosystem recovery in mountainous landscapes.
Accurate mapping of soil organic matter (SOM) is critical for sustainable land management and climate change assessment. Taking Youyi Farm in Northeast China—located within one of the world's four major black soil regions—as a case study, 188 soil samples were collected and analyzed to construct a high-quality SOM dataset. Based on Sentinel-2 imagery, a multi-temporal remote sensing dataset representing three climatically contrasting years (2019 flood, 2020 normal, and 2021 drought) was developed, integrating spectral bands, vegetation indices, and environmental covariates. Four machine learning models—Random Forest (RF), Gradient Boosting Decision Trees (GBDT), eXtreme Gradient Boosting (XGBoost), and Particle Swarm Optimization–Support Vector Regression (PSO-SVR)—were systematically compared. PSO-SVR achieved the best performance using bare-soil spectral information alone (R2=0.544, RMSE=1.159%). Incorporating growing-season vegetation indices (NDVI, EVI, LSWI) improved R2 by 2.0%–5.5% relative to the bare-soil baseline depending on the climatic year, and further integrating 16 environmental covariates increased R2 to 0.617 in the drought year—a 13.4% relative improvement. SHAP analysis identified channel network base level, the green band (B3), and mean annual temperature as the dominant predictors, revealing how topographic drainage, spectral reflectance, and temperature-dependent microbial processes jointly control SOM spatial variability. Model accuracy followed the order drought > normal > flood year, providing practical guidance for selecting optimal remote sensing acquisition periods under different climatic conditions. Overall, explicitly accounting for interannual climatic variability substantially enhances the accuracy and robustness of SOM prediction in black soil regions.
Nigeria has a diverse range of forest types that conserve biodiversity and provide significant ecosystem services, including non-timber forest products, carbon sequestration, and climate regulation. However, Nigeria has historically lacked comprehensive, spatially explicit data on tree canopy cover (TCC) at the national scale, hindering the development of evidence-based forest management and policy. This is due to logistical limitations, which are further exacerbated by the insecurity crisis. However, the availability of high-resolution images, combined with advanced machine learning, can help mitigate these challenges. Therefore, this study presents the first national-level datasets of tree canopy cover and forest/non-forest areas for Nigeria. Using stratified random sampling, reference tree canopy cover was collected nationwide (n = 3,047) using Google Earth imagery. The tree canopy cover was modeled using a random forest regression, incorporating Landsat bands, vegetation indices, and AlphaEarth Embeddings in Google Earth Engine. Results showed that the inclusion of embeddings with multispectral bands and indices were helpful in accurately modeling tree canopy cover in a sub-Saharan region, such as Nigeria. The 2024 tree canopy cover model was used to predict tree canopy cover for other years (2017 and 2020), and the resulting tree canopy cover maps for the years were validated. Using the country’s adopted forest definition, the TCC maps were reclassified to generate binary classified forest and non-forest maps. Based on the analysis, 18.7% of Nigeria was forested in 2024, and the forest cover percentage followed a dynamic pattern from 2017 to 2024. The developed datasets are publicly available on Zenodo and are fundamental for continuous national-level forest monitoring and sustainable management.
The normalized difference vegetation index (NDVI) is vital for monitoring vegetation health. However, NDVI imagery imposes trade-offs in spatial resolution, temporal coverage, and historical depth. Sentinel-2 MSI has provided 10-m imagery since 2015, while the Landsat archive extends to 1982, though its 30-m resolution limits fine-scale vegetation analysis. We propose a knowledge distillation strategy to enhance Landsat 8 imagery to match Sentinel-2 resolution. Our novel deep channel attention super-resolution network (KD-DCASRN) serves as a lightweight student model guided by a deeper teacher, the widely used enhanced deep super-resolution network (EDSR). We experimentally validated the student model’s accuracy across diverse land-cover types, including dense forests, wetlands and water bodies, agricultural fields, urban settlements, sparse vegetation, and open land, using study regions in Assam (Palashbari and Majuli) and Rajasthan (Ajmer), India. KD-DCASRN outperformed all baseline methods across the three study regions, achieving an average PSNR gain of 56.7% over Bicubic interpolation, 21.6% over ESRT, 8.4% over SRCNN, 2.2% over EDSR, and 1.6% over the non-distilled DCASRN model. Correspondingly, KD-DCASRN reduced RMSE by 58.9%, 30.5%, 15.9%, 10.2%, and 9.2%, respectively, demonstrating its effectiveness for high-resolution NDVI reconstruction. Qualitative assessments show more structurally faithful reconstructions. The model’s strong cross-biome performance demonstrates its practical utility in vegetation monitoring.
Plant phenotyping is central to developing climate-resilient, stress-tolerant crop varieties, particularly as rising food demand and climate change intensify pressure on agricultural systems. Nitrogen (N) is an essential nutrient that significantly influences the growth, health, and productivity of grain-producing crops. Accurate identification of N status in plants is critical in phenotyping studies, enabling timely interventions and optimal use of fertilizers to improve plant health and production. Hyperspectral imaging (HSI) acquires rich spectral information across a broad wavelength range, facilitating precise characterization and analysis of complex phenotypic traits, such as early identification of water and nitrogen stress. HSI, when integrated with unmanned aerial vehicles (UAVs), is capable of accelerating crop phenotyping by offering extensive coverage and high spectral, spatial, and temporal resolutions. This study leverages UAV-based HSI (400–1000 nm) combined with machine learning (ML) techniques to classify nitrogen treatment in sorghum canopy. A comprehensive end-to-end data analysis framework is presented, encompassing calibration, quality assessment, denoising, band selection, and classification for UAV-acquired hyperspectral images. A hybrid ensemble band selection (hyb-EBS) algorithm is proposed, leveraging homogeneous and heterogeneous ensemble feature selection and clustering strategies to identify key wavebands sensitive to canopy nitrogen status in sorghum. Wavebands centered at 699.13, 707.66, 720.47, 754.79, and 938.59 nm were identified as most effective in classifying nitrogen treatment in sorghum, achieving classification accuracies of 96.29%, 93.11%, 95.89%, and 95.76% at late-vegetative, flowering, milking, and grain filling stages, respectively. The HSI datasets and associated code will be made publicly available to facilitate further research in UAV-based crop phenotyping 11https://github.com/sankaraug/HSI-Sorgh-Stress..
The European Space Agency's (ESA) Copernicus program has provided free access to satellite remote sensing data, making it an attractive option for public institutions. Government institutions are currently evaluating and identifying use cases for satellite imagery. Agricultural administration stands to benefit particularly, with the potential uses being substantial, especially in this area. This is primarily driven by two factors: (1) legal requirements within the framework of the Common Agricultural Policy (CAP), and (2) the substantial proportion of agricultural land within the total land cover in a mid-European study area. In the domain of agriculture and environmental discourse, there are a variety of applications for agricultural remote sensing in the context of providing a database for decision makers within the agricultural administration.The administration's activities are connected with the careful handling of the created geodata products, depending on the data requirements and the uncertainties inherent in the products as a result of the creation process. Government institutions have high accuracy requirements for the products they use to make decisions. In this regard, this thesis examines the factors that influence the informative value of agricultural remote sensing products and proposes possible solutions in understanding and communicating data uncertainties. This applied approach aims to increase the social benefits of remote sensing. The objective of this study was to assess the potential and limitations of applications by remote sensing data from the point of view of the agricultural administration. To this end, a total of 53 uncertainty factors in agricultural remote sensing were compiled. A selection of practices for communicating and visualizing uncertainty was created to address these factors.
Accurate, up-to-date, and high-resolution land use/land cover (LULC) maps are essential for environmental monitoring, spatial planning, climate resilience, and sustainable land management. Despite the increasing availability of Sentinel-2 imagery, existing publicly available datasets often lack explicit seasonal representation, consistent multi-year annotations, or thematic detail suitable for heterogeneous Mediterranean landscapes. To address these limitations, this study introduces S2GAIA, a multi-year, seasonally aware Sentinel-2 dataset for pixel-wise LULC mapping in Greece. The dataset is generated by harmonizing multiple geospatial sources, including Copernicus products and national datasets, into a unified 22-class taxonomy. S2GAIA comprises 34,030 annotated Sentinel-2 image patches (256 × 256 pixels) at 10 m spatial resolution, spanning the period 2017–2024. Each year is represented by four cloud-free seasonal composites, enabling the exploitation of phenological variability and improving the discrimination of temporally dynamic land-cover classes. The dataset was evaluated using four semantic segmentation architectures, with ResUNet achieving the highest performance, obtaining a Cohen’s Kappa coefficient of 91.23% and a weighted F1-score of 92.23%. Additional class-wise, confusion, and seasonal analyses demonstrate the effectiveness of the proposed multi-season design while identifying the remaining challenges associated with fragmented and spectrally similar land-cover classes. The resulting national-scale LULC maps demonstrate the practical applicability of S2GAIA for environmental monitoring and provide a valuable resource for developing and evaluating deep learning models for Mediterranean land-cover mapping. The S2GAIA dataset and the associated source code are publicly available through Zenodo (https://zenodo.org/records/21883467) and GitHub (https://github.com/eugeniapapathe/S2GAIA), respectively.
