Water quality in sub-Saharan Africa is increasingly threatened by rapid land cover change and limited monitoring. This study used Landsat imagery (1990–2020) and field data from three dams in Bukuru and environs in north-central Nigeria to test satellite indices as water quality proxies. Land use/land cover was classified using Support Vector Machine, while NDVI (Normalised-Difference Vegetation Index), NDWI (Normalised-Difference Water Index), and NDTI (Normalised-Difference Turbidity Index) were correlated with pH, BOD, and TDS using regression and Moran's I. Results show that built-up areas increased by 191% and water surface declined by 45%. NDVI strongly correlated with pH (R2 = 0.92), reflecting vegetation buffering and organic inputs. NDTI captured BOD variability linked to turbidity (R2 = 0.57). No spectral index predicted TDS (R2 < 0.1). NDVI and NDTI are effective for organic pollution monitoring, whereas TDS remains optically undetectable. Findings provide framework for satellite-based water quality monitoring and stress riparian and sediment management in urbanising basins.
This study compared U-Net models with ResNet-18, ResNet-34, ResNet-50, and ResNet-101 encoders for oil-palm segmentation using 10 m Sentinel-2 RGB true-colour imagery in Riau, Indonesia. A pooled multi-year dataset comprised 1,283 image–mask patches: 1,026 for training and 257 for patch-level testing. Sixty matched configurations produced 240 model fits. Mean intersection over union (IoU) ranged from 0.789 to 0.812, with no significant overall backbone effect (χ2(3) = 1.858, p = 0.602). ResNet-18 exhibited the lowest IoU coefficient of variation (18.3%) while requiring approximately 72% fewer parameters and 77% fewer floating-point operations than ResNet-101. Increasing encoder depth did not consistently improve performance, whereas ResNet-18 offered a favourable observed accuracy–efficiency trade-off. These findings do not establish statistical equivalence or generalisation beyond the evaluated RGB input, training configurations, reference masks, and patch-level test set.
Deep learning has rapidly expanded in agricultural remote sensing, but its thematic evolution and technical configurations remain insufficiently examined together. This study analysed 1,706 Web of Science Core Collection articles published from 2017 to 2025 using CiteSpace and VOSviewer, and applied Task–Data–Model (T–D–M) coding to a publication-year-stratified random sample of 500 articles. Publications increased markedly after 2020, with broader collaboration networks and interdisciplinary participation. Four major themes were identified: crop growth and yield prediction, land-cover classification and segmentation, object-level recognition and fine-scale monitoring, and broader deep-learning applications. Temporal analyses showed growing attention to time-series modelling and multi-source integration. In the coded sample, Task–Data association was strongest (Cramér’s V = 0.362), exceeding Task–Model (0.215) and Data–Model (0.186). These results show that data-source selection is more task-structured than broad model-family use.
The Shebelle River Basin in Somalia is perennially threatened by devastating floods, yet it remains profoundly understudied, with existing conventional models, incomplete variable selection, and a lack of model interpretability. This study addresses these gaps by proposing an Explainable Stacking Ensemble Model designed for high-precision and transparent flood risk assessment. Our sophisticated stacking architecture integrates five diverse base models, XGBoost, LightGBM, and CatBoost, whose predictions are synthesised by a CatBoost meta-classifier to achieve superior predictive performance. The model was trained on a comprehensive set of seventeen conditioning factors using a balanced dataset of 4,154 locations. Our proposed model demonstrated exceptional performance, achieving (Accuracy: 91.26%, F1-Score: 91.36%, ROC-AUC: 0.9723) and outperforming Gradient Boosting Machines, Extra Trees, XGBoost, LightGBM, and CatBoost. Crucially, to dismantle the ‘black-box’ nature of complex ensembles, we embedded SHAP (SHapley Additive exPlanations) analysis directly into the framework. This provides unparalleled interpretability, revealing Topographic Wetness Index (TWI), elevation, rainfall, and stream order as the most influential factors driving flood susceptibility. The resulting high-resolution risk map accurately identifies high-risk zones along the middle and lower reaches of the Shebelle River, aligning with historical events.
The present study aims to predict soil erosion-prone regions in Kangra District using ML algorithms; Support Vector Machine (SVM), Random Forest (RF), XGBoost and CatBoost and their stacking ensembles. Additionally, the study also utilizes the SHAP (SHapley Additive exPlanations) model to enhance the interpretability of the best-performing RF algorithm. For this, a comprehensive soil erosion point dataset was developed using empirical field survey integrated with visual interpretation of Google Earth historical images. The results show RF as the best-performing model, predicting 14.7% of the area into the very high susceptibility zone, followed by XGBoost (27.1%), SVM (24.4%) and CatBoost (16.4%), whereas the stacking ensemble predicted 25.2%. The SHAP analysis of the best-performing model (RF) indicates NDVI, slope and LULC as the primary governing factors influencing soil erosion in the region. This study also provides a scientific base for policy formulation and strategic planning to mitigate soil erosion in Kangra District.
