The rapid expansion of mussel raft aquaculture has increased the demand for accurate and scalable monitoring systems to support sustainable coastal management. Deep learning (DL) object detectors, particularly the YOLO family, enable automated extraction of aquaculture infrastructures from Very High-Resolution (VHR) satellite imagery. However, conventional detectors rely on horizontal bounding boxes that often include excessive background when objects are elongated or rotated, reducing localization accuracy. Oriented bounding box (OBB) detectors address this limitation by modeling object orientation and geometry. In this study, an open-source Python-based graphical user interface (GUI) was developed for automated mussel raft detection, and a comprehensive comparison of YOLOv8-OBB and YOLO11-OBB architectures was conducted using VHR imagery from Spain’s Ría de Arousa. A dedicated mussel raft dataset was compiled and annotated, and both model families were evaluated across five architectural scales (n, s, m, l, and x). Experimental results showed that the YOLO11 family generally outperformed YOLOv8 in terms of detection accuracy, localization quality, and performance under challenging maritime conditions. Among the evaluated models, YOLO11s achieved the highest F1-score (0.866) and mAP50–90 (85.21%), while YOLO11n achieved the highest Mean IoU (0.872) and mAP50 (96.87%). The results further demonstrated that compact architectures can provide a favorable balance between accuracy and computational efficiency, as larger models did not consistently yield superior performance. Large-scale experiments on a 60,309 × 61,569-pixel (50 cm resolution) VHR scene demonstrated the operational scalability of the proposed framework through a patch-based processing strategy. The comparative evaluation conducted in this study provides valuable insights into recent YOLO-OBB architectures and highlights their potential for scalable AI-based aquaculture monitoring.
Human‐wildlife conflict (HWC) is increasingly associated with rapid land‐use and land‐cover (LULC) changes and broader environmental shifts. While previous studies have focused on causal factors and coexistence strategies, the spatiotemporal dynamics of human-wildlife interactions under shifting LULC remain poorly quantified. This study examines the spatiotemporal patterns of LULC and their association with HWC around the Hluhluwe‐iMfolozi Park in South Africa. The detailed landscape dynamics were captured using Sentinel‐2 imagery at 5‐year intervals (2015, 2020, and 2025) with machine‐ and deep‐learning classification algorithms. Among the evaluated algorithms, the Two‐dimensional Convolutional Neural Network outperformed the others, achieving an average overall accuracy of 99.92% and a kappa coefficient of 0.9991. The findings show substantial LULC changes during the study, including a notable decrease in cultivated land, changes in vegetation structure, and an increase in built-up areas. Shrubland grew significantly by 2025, suggesting continuous landscape change, while woodland increased between 2015 and 2020. Anthropogenic activities and vegetation gains dominated earlier transitions, while natural vegetation transitions drove later changes. Outside protected regions, these trends were more pronounced, indicating greater anthropogenic pressure. Integration of LULC transitions with HWC data revealed that conflict events were increasingly linked to grassland, woodland, and transitional landscapes and were concentrated along protected area boundaries. Hotspots were closely associated with regions undergoing active land-cover change, and the frequency and spatial extent of HWC increased over time. These results demonstrate how landscape dynamics spatially overlap with human‐wildlife interactions and emphasize the need for sophisticated geospatial technologies to pinpoint conflict hotspots and guide targeted conservation and mitigation efforts.
Land Use/Land Cover (LULC) change detection is fundamental for understanding landscape dynamics and supporting sustainable environmental management, particularly in ecologically sensitive and transboundary conservation areas such as Kruger National Park, South Africa. Rapid urban expansion, agricultural intensification, and environmental change have accelerated landscape transformation, highlighting the need for accurate and reliable LULC monitoring. This study aimed to evaluate long-term LULC dynamics in and around Kruger National Park during 2005, 2015, and 2025 using multi-temporal Landsat imagery and four machine learning classifiers: Random Forest (RF), Support Vector Machine (SVM), Gradient Tree Boosting (GTB), and K-Nearest Neighbors (KNN). In addition, RF-based transition matrices were employed to quantify long-term land cover transformations. The results showed that the Random Forest classifier consistently outperformed the other algorithms, achieving the highest Overall Accuracy (up to 92%) and Kappa coefficient (0.87) across all study years. Gradient Tree Boosting and K-Nearest Neighbors produced satisfactory results, whereas Support Vector Machine exhibited comparatively lower classification performance. Spatial analysis indicated that woodland/shrubland remained the dominant land cover class but experienced progressive fragmentation and conversion to grassland, agricultural land, and bare land. Built- up areas expanded substantially, particularly between 2005 and 2015, while agricultural land also increased, reflecting growing anthropogenic pressure. Water bodies exhibited a declining trend throughout the study period, suggesting increasing hydrological and environmental stress. These findings demonstrate that ensemble-based machine learning methods, particularly Random Forest, provide a robust and reliable approach for monitoring complex and heterogeneous savanna landscapes. The observed LULC changes highlight the increasing influence of human activities on ecosystem structure and emphasize the importance of continuous geospatial monitoring to support biodiversity conservation, sustainable land management, and evidence-based environmental planning. Overall, this study provides a reliable framework for long-term LULC mapping and change detection that can be applied to protected areas and other ecologically sensitive landscapes.
Atmospheric pollution associated with carbon monoxide (CO), nitrogen dioxide (NO2), and sulfur dioxide (SO2) represents an important environmental and public health concern in Andean regions influenced by urban growth, transportation, agricultural burning, mining-related activities, and complex topography. This study analyzed the spatiotemporal variability of CO, NO2, and SO2 in the Apurimac region, Peru, during 2020–2023 using Sentinel-5P/TROPOMI satellite products processed in Google Earth Engine. The original column-density data, expressed in mol/m2, were converted into column-derived estimated concentrations using a simplified effective lower-atmospheric layer height of 2,000 m. These values were used only as relative indicators and were not interpreted as direct ground-level air quality measurements. The results showed heterogeneous spatial patterns across Apurimac. CO presented relatively higher values mainly in Abancay, Chincheros, and Andahuaylas, while NO2 showed higher relative values in Cotabambas, Grau, Abancay, and Chincheros. SO2 exhibited a more irregular and uncertain behavior, with localized positive values, negative retrievals, and high variability, indicating greater sensitivity to satellite retrieval noise. Therefore, SO2 was interpreted as an exploratory indicator rather than robust evidence of province-level pollution hotspots. Temporal analysis showed seasonal fluctuations and isolated peaks, especially during dry-season months. Trend detection was performed using the Mann-Kendall test and Sen’s slope estimator. CO showed a statistically significant decreasing trend, NO2 showed a weak but statistically significant increasing trend, and SO2 did not show a statistically significant trend. Overall, Sentinel-5P/TROPOMI and Google Earth Engine proved useful for identifying relative pollutant patterns in a data-sparse Andean region. Future studies should integrate satellite observations with in situ measurements, meteorological data, boundary-layer information, emission inventories, and independent datasets to improve validation, source attribution, and regional air quality assessment.
Rapid tourism expansion and infrastructure growth in the Western Ghats are reshaping ecological stability in steep, topographically complex landscapes that still retain substantial forest cover. These changes are intensifying habitat fragmentation, degrading ecosystems, and placing increasing pressure on biodiversity-rich environments. Focusing on Idukki district in southern India, this study integrates the Urban Nature Access (UNA) and Habitat Risk Assessment (HRA) modules within the InVEST framework to quantify how accessibility-driven development influences habitat vulnerability between 2011 and 2025. Results reveal a widening spatial imbalance between population demand and accessible natural areas, with strongly negative urban nature balance values expanding across valley settlements and plantation corridors by 2025. Built-up land exhibits the highest mean habitat risk (R̄ = 0.42), followed by plantations (R̄ = 0.38) and croplands (R̄ = 0.34), while deciduous forests (R̄ = 0.22) and water bodies (R̄ = 0.05) remain comparatively less vulnerable. More than one-third of built-up and plantation landscapes fall within medium to high-risk categories. Spatial overlap between high-risk zones and documented landslide-affected areas during the 2018–2020 extreme monsoon events highlight the cumulative impact of tourism-driven development and associated forest conversion on steep slopes. The findings demonstrate that ecological vulnerability in Idukki is closely linked to localized forest loss and land-use transformation driven by concentrated human activities in hazard-prone terrain. These patterns highlight how steep and environmentally sensitive landscapes are increasingly affected by development pressures. This underscores the need to integrate ecosystem accessibility metrics with slope-sensitive land-use planning in tropical mountain systems.
Gross primary productivity (GPP) is a crucial indicator for understanding the global carbon cycle and climate change. Sun-Induced Chlorophyll Fluorescence (SIF) provides a direct link to plant photosynthesis, offering a novel approach for GPP estimation in terrestrial ecosystems. Given the complex factors influencing the canopy SIF–GPP relationship and the limited generality of empirical linear models, we developed a Light Use Efficiency (LUE) model incorporating the photochemical reflectance index (PRI) and structural vegetation indices, with a nonlinear function fitted using support vector machine regression. The model was evaluated across multiple vegetation types, SIF products, and vegetation indices to identify optimal SIF–VI combinations. Results indicate that training by vegetation type improves accuracy by 8.5%–14.4%, with shrublands and evergreen broadleaf forests showing the best performance. The combination of GOSIF with NDVI achieved the highest overall estimation accuracy across all vegetation types. Additionally, the combination of GOME-2 SIF with NDVI × NIRv performed best in inland arid low-GPP regions, while TCSIF with NDVI × NIRv was most accurate in cold high-latitude low-GPP regions. All three combinations effectively captured interannual GPP variations at individual sites. Using these combinations, monthly GPP from 2007 to 2014 was estimated and compared with GPP-MODIS and GPP-GLASS datasets, showing consistent temporal and spatial patterns that reflect seasonal dynamics in the Northern Hemisphere. These findings demonstrate that integrating SIF with targeted vegetation indices enhances GPP estimation and provides a robust framework for large-scale photosynthesis monitoring.
In visible-SAR cross-modal remote sensing target detection tasks, due to the significant differences between the two modalities in imaging mechanisms and feature representations, as well as the susceptibility of SAR images to speckle noise interference and the inadequate utilization of structural information, existing methods often struggle to balance detection accuracy with model lightweighting and practical deployment requirements. This is particularly true in resource-constrained scenarios, such as space-borne or airborne platforms, where models need to have low parameter counts, reduced computational overhead, and robust adaptability to complex environments. To address these issues, this paper proposes a lightweight dual-branch fusion network, DBF-YOLO, for visible-SAR cross-modal remote sensing target detection. Based on the YOLOv10 framework, the method constructs a visible-SAR dual-branch feature extraction structure and designs an intermediate cross-modal fusion path at three levels (P3, P4, P5) to achieve progressive interaction of dual-modal features at different scales. To overcome the strong speckle noise and edge degradation in SAR images, a SAR Gradient Enhancement Module (SGM) is introduced to enhance the structural representation capability of SAR inputs. Additionally, an Adaptive Gated Dual-Modal Fusion Module (AGD) is proposed to enable dynamic selection and effective complementarity of dual-modal information based on different scales and spatial positions. Experimental results on the OGSOD 1.0 dataset show that DBF-YOLO achieves 94.4% mAP50% and 70.1% mAP50-95, with only 5.0 M parameters and 20.5 GFLOPs, striking a good balance between detection accuracy and computational complexity. Furthermore, experimental results on the OSPRC dataset demonstrate the robustness of DBF-YOLO under different resolutions, polarization modes, and cloud cover conditions. Particularly in low-contrast dense target scenes and cloud cover conditions, the model demonstrates stronger environmental adaptability. This method can provide a reference for designing lightweight multi-modal remote sensing target detection models and deploying them on resource-limited platforms.
IntroductionIndustrial hemp (Cannabis sativa L.) is a multipurpose bio‐economy crop capable of producing fiber, grain, and biomass while contributing to soil health and carbon sequestration. Realizing this potential requires optimized nitrogen (N) management, as N strongly regulates plant growth, yield, and fiber quality, while inefficient or excessive N use can cause environmental harm. Conventional N diagnostics based on destructive sampling are labor‐intensive and lack the spatial resolution needed for precision management. Uncrewed aerial vehicle (UAV) multispectral imaging offers a high-throughput, non‐destructive alternative; however, hemp remains underrepresented in UAV‐based N studies, particularly in linking spectral data to physiological traits associated with N metabolism.MethodsTo address this gap, two field experiments were conducted at the North Carolina A&T State University research farm using dual‐purpose and fiber‐type hemp cultivars under contrasting N regimes during the 2024 and 2025 growing seasons. Multispectral imagery was collected using a WingtraOne GEN II UAV equipped with a MicaSense RedEdge‐P camera. From reflectance mosaics, 33 vegetation indices (VIs) were computed, and the top seven were selected using Spearman's correlation analysis. Ground measurements included SPAD chlorophyll readings and gas‐exchange traits, i.e., net photosynthetic rate (Pn), stomatal conductance (Gs), and transpiration rate (E), using a LI‐COR 6800 system. Using SAS Viya, multiple supervised learning models were developed to predict SPAD, Pn, Gs, and E from UAV‐derived VIs.ResultsRed‐edge and green‐based indices showed very strong correlations with SPAD (ρ = 0.9078−0.9375), while NDWI showed a strong negative relationship (ρ = −0.9623). For Pn, GNDVI and CIG were strongly correlated (ρ ≈ 0.89), with NDWI negatively associated (ρ = −0.8934). Gs and E exhibited moderate correlations (ρ ≈ 0.73−0.84 and 0.75−0.80). Linear regression achieved R2 = 0.88 for SPAD, while a generalized additive model predicted Pn with R2 = 0.87. Quantile regression performed best for Gs and E (R2 = 0.82 and 0.75; Gs: ASE ≈0.013, MAE ≈0.086; E: ASE ≈0.0003, MAE ≈0.014).DiscussionThese results demonstrate a scalable framework for UAV‐based phenotyping to support precision N management in hemp.
Introduction Remote sensing image classification is an important task in Earth observation. However, achieving high accuracy is still challenging. This is mainly due to high-dimensional feature redundancy, large intra-class variability, and the difficulty of capturing both fine spatial details and long-range contextual information. To address these challenges, this paper proposes a unified classification framework based on a novel Multi-Scale Dual-Path Shifted Pyramid Vision Transformer (M-DSPViT). Methods The proposed model improves standard Vision Transformer architectures by introducing dual-path shifted patch embedding and content-adaptive attention gating. It also incorporates multi-scale feature pyramid fusion, dynamic expert routing, and gradient-based attention masking to better capture spatial and contextual features. In addition, a hybrid feature selection method (HSIC-HFS) is introduced to remove redundant information and retain discriminative features. This method combines the Hilbert-Schmidt independence criterion, Shannon entropy, and L 1 -regularization. The refined features are then integrated with CNN-based spatial descriptors extracted using EfficientNet through a Sequential Feature Aggregation (SFA) framework. Results The proposed method is evaluated on the WHU-RS19, UC Merced, and AID benchmark datasets. It achieves state-of-the-art performance in land-use classification. The robustness and generalization ability of the model are further validated through statistical analysis, including one-way ANOVA and F-statistic testing. Discussion The results confirm the stability of the proposed approach across different remote sensing scenarios.
Given the similarities in their instrumental characteristics, a comparative analysis was conducted between hyperspectral missions of PRISMA, operated by the Italian Space Agency (ASI), and EnMAP, developed by the German Aerospace Center (DLR). This work investigates both the similarities and discrepancies between these two satellite optical data, providing an outline of the spectral regions where data from both sensors can be reliably interchanged. The analysis was performed by analyzing the full visible and near-infrared (VNIR) spectrum and by examining selected spectral ranges sensitive to key environmental parameters, such as nitrogen dioxide, chlorophyll and NDVI. A total of 11 PRISMA and EnMAP co-located hyperspectral image pairs acquired across different seasons over the Euro-Mediterranean region were analyzed, focusing on representative land cover classes (bare soil, vegetation, urban areas and waters). Preprocessing steps were applied to standardize the acquisition viewing geometries, and the top of atmosphere (TOA) level of products from the two sensors were used to avoid inconsistencies introduced by atmospheric correction schemes adopted by each operating agency. Overall, the quantitative and qualitative assessments confirm a strong correspondence between the PRISMA and EnMAP hyperspectral measurements across the VNIR spectrum, with SAM always below 0.1 rad. However, the statistical analysis reveals relatively low consistency between the two sensors for the vegetation sensitive spectral window (750–940 nm), where larger deviations are observed with STD of 0.067 and SAM of 0.097, while the NO2 spectral range (420–500 nm) shows lower SAM of 0.048 and STD of 0.045.
The accurate assessment of forest aboveground carbon (AGC) is crucial for improving the efficiency of forest resource management, mitigating climate change, and fostering sustainable development. However, the extensive distribution of forests, their complex ecosystem structures, insufficiently representative assessment data, and methodological inconsistencies generally lead to estimates with low accuracy and high uncertainty. To address these issues systematically, this study introduces a novel framework that integrates remote sensing features with field-measured plot data. This framework leverages an Optuna-optimized eXtreme Gradient Boosting (XGBoost) model to achieve accurate estimation of forest AGC. A key contribution of this study is the application of Optuna to optimize the hyperparameters of the XGBoost model, which improves both its predictive performance and generalization ability. For feature selection, we employed a combination of the Pearson correlation coefficient and the Boruta algorithm, which identified ten core feature variables from the initial set. This process effectively improved the relevance and interpretability of the model inputs. The experimental results demonstrate that feature selection markedly improved model performance: R2 increased by 0.2, RMSE decreased by 2.33 Mg C/ha, and MAE was reduced by 6.21% compared to the model using the unselected feature set. The Optuna-optimized XGBoost model demonstrated excellent performance, achieving an R2 of 0.72, an improvement of 0.12 over the baseline XGBoost model, with an RMSE of 24.48 Mg C/ha and an MAE of 18.56%. These results indicate superior predictive accuracy and stability. In conclusion, the integrated framework developed in this study, which combines multi-source remote sensing data with machine learning, effectively enhances the estimation accuracy of forest AGC at a regional scale. This approach provides a reliable theoretical basis and a practical methodology for the dynamic monitoring and management of forest carbon sinks.
Timely and accurate crop-type mapping is fundamental for sustainable agricultural management and food security in semi-arid regions, where climate variability and fragmented landscapes present persistent challenges. This study develops and validates an operational, multi-temporal framework for classifying five key agricultural classes (i.e., soft wheat, durum wheat, barley, trees, and other crops) across diverse Moroccan agroecosystems. By integrating monthly Sentinel-1 Synthetic Aperture Radar and Sentinel-2 optical time series spanning six growing seasons (2018–2025), we extracted 156 features comprising 13 spectral indices across 12 monthly composites. Ground truth data from the national Al Moutmir database, strategically balanced to address natural class imbalances, supported comprehensive training and validation of six machine learning models (i.e., Random Forest, Extra Trees, XGBoost, LightGBM, Voting Ensemble, and Stacking Ensemble). Our findings showed that the LightGBM and Stacking Ensemble achieved the highest performance with 88.04% overall accuracy, followed closely by XGBoost (87.93%). Feature importance analysis revealed that monthly temporal resolution significantly outperformed traditional phenological-stage approaches, with March and April indices (particularly Normalized Difference Vegetation Index and Normalized Difference Red Edge) contributing most to class discrimination. Notably, early-season radar features (Vertical-Vertical polarization in September) provided valuable complementary information when optical data were limited. The framework demonstrated robust generalization through 10-fold cross-validation while explicitly quantifying a 12.56% overfitting gap (train-CV difference), acknowledging a non-negligible overfitting risk. Offering transparent performance assessment. Error analysis identified persistent confusion between spectrally similar cereals, particularly durum and soft wheat, highlighting priority areas for future sensor integration. This scalable, cloud-based pipeline directly supports Morocco’s Green Generation strategy by providing a reproducible, high-accuracy solution for annual crop inventories, with transferable applications across similar Mediterranean and semi-arid agricultural systems.
IntroductionHydrological hazards, such as floods and landslides are frequently driven by extreme rainfall events (ERE). Thus, understanding the spatio-temporal patterns and intra-event behavior of these events is important for identifying vulnerable regions, improving early warning systems, and enhancing water management.MethodsThis study aimed to analyze the spatio-temporal intra-event characteristics of ERE, using high-resolution (5min) weather radar data focusing on their internal structure and spatial distribution. The study was conducted in the headwaters of the Paute basin (2,200–4,400 m a.s.l.) in southern Ecuador. Based on three ERE classes, four intra-event rainfall features were analyzed: area, maximum rainfall, cohesion, and the locations of rainfall hotspots.ResultsThese features revealed different rainfall patterns for the three distinct rainfall classes. Class 1 is characterized by the highest rainfall peaks, concentrated between 12:00 and 19:00 (afternoon). Class 3 shows the lowest rainfall peaks. Class 2 shows the least cohesive rainfall core and a mixed behavior in features. Regarding the locations of rainfall hotspots, classes 1 and 2 show hotspots located at the catchment outlets and at the urban (City of Cuenca) areas of the sub-catchments (around 2,500 m a.s.l), while those in class 3 are found at headwaters (above 3,500 m a.s.l).DiscussionIdentifying these rainfall characteristics and hotspot location provides a better understanding of extreme rainfall behavior in the tropical Andes, which enhances knowledge of hydrological processes, and improves flood forecasting.
The Qilian Mountains, a crucial ecological security barrier and water conservation region in northwestern China, are highly sensitive to climate change. Reliable climate projections are essential for regional environmental management, yet the performance of statistical downscaling methods in this complex mountainous region remains inadequately evaluated. This study evaluates four statistical downscaling methods—Delta Change Method (DCM), Quantile Mapping (QM), Multiple Linear Regression (MLR), and Random Forest (RF)—using station-based observations (1951–2025) and outputs from 34 CMIP6 climate models. Validation results indicate that RF achieved the highest R2 and the lowest RMSE for both temperature and precipitation, with superior stability across models, scenarios, and stations. Based on RF downscaling, future climate projections under SSP1–2.6, SSP2–4.5, SSP3–7.0, and SSP5–8.5 indicate a persistent warming and moderate wetting trend throughout the 21st century. Temperature is projected to increase at rates of 0.03 °C–0.31 °C/decade, while precipitation is projected to increase by approximately 3.2%–8.1% by the end of the century, although inter-model uncertainty remains substantial. Future climate change also exhibits pronounced seasonal and spatial heterogeneity, characterized by winter-dominated warming, reduced summer precipitation but increased autumn–winter precipitation, and strong elevation-dependent responses concentrated in high-elevation areas. The 0 °C isotherm is projected to rise by approximately 100–400 m by the late 21st century. These findings highlight the applicability of RF downscaling for station-scale climate projections in the Qilian Mountains and provide scientific support for regional climate adaptation, water resource management, and ecosystem conservation.
IntroductionSatellite-based deforestation monitoring is critical for environmental sustainability and climate change mitigation. However, conventional remote sensing approaches face significant limitations in cloud-covered regions and are often inadequate for capturing temporal change dynamics, constraining their effectiveness in continuous forest surveillance.MethodsWe propose the Siamese Attention U-Net with Multimodal Temporal Fusion (SAU-MTF), a novel deep learning architecture that integrates optical (Sentinel-2) and Synthetic Aperture Radar (Sentinel-1) imagery within a tri-temporal framework. The model employs EfficientNet-based encoders and attention gates for discriminative feature extraction, alongside temporal context blocks designed to capture change dynamics across time steps. A multimodal fusion strategy is adopted to exploit the complementary strengths of SAR’s cloud-penetrating capability and the spectral richness of optical data.ResultsEvaluated on large-scale deforestation datasets, SAU-MTF achieves a classification accuracy of 94.7% and an Intersection over Union (IoU) of 0.93, outperforming existing state-of-the-art models across benchmark comparisons.DiscussionThese results demonstrate that the joint exploitation of temporal, spectral, and spatial information substantially enhances deforestation detection performance. The architecture proves especially effective in challenging environments characterized by persistent cloud cover and seasonal variability in remote forest regions, highlighting its potential for operational deployment in global forest monitoring systems.
In Part III of the series, we evaluate the accuracy and applicability of the tomographic algorithm introduced in Part I and applied to real measurements by the research scanning polarimeter in Part II. We focus on the core part of the algorithm, producing a nested family of cloud shapes corresponding to a range of brightness thresholds. This family is then used to derive a 2D field of cloud extinction coefficient. We relate the resolution of the multi-angle measurements to the spatial accuracy of the cloud shape retrievals and determine constraints on the cloud aspect ratios required for the applicability of the algorithm. The expressions for overpass length and time derived in this study allow for estimating how much the cloud can move or change during the measurement process. We estimate biases in cloud size and position retrievals caused by the cloud’s advection during the measurements. Our accuracy estimation techniques are applied to previously published examples of clouds, both simulated and real.
Canopy segmentation is a crucial step in obtaining canopy parameters in forestry remote sensing, and it holds significant importance for research areas such as forest carbon sequestration. However, canopy segmentation based on UAV imagery still faces challenges including confusion between edges and background, as well as insufficient edge details. To address these issues, this paper proposes a semantic segmentation network based on U-Net which implements boundary and key feature enhancement named BKFE-UNet, which is built upon the U-Net model and integrates a differential boundary attention module (DBM) and a key feature enhancement module (KFEM). The DBM enhances canopy edge information through differential computation, alleviating problems such as adhesion between adjacent canopy edges and inadequate edge details. The KFEM introduces Ghost convolution; by stacking Ghost features, it simultaneously filters out redundant information and enhances the key semantic features of canopy objects, thereby reducing background interference. On the UAV tree canopy segmentation dataset, BKFE-UNet achieves an mPA of 92.62%, an mIoU of 86.50%, an Accuracy of 96.30%, and an F1-score of 0.90, representing improvements of 2.23%, 2.27%, 2.09%, and 0.02, respectively, over the baseline U-Net, demonstrating a significant improvement over the baseline U-Net model.
IntroductionThe spatial dynamics of wetlands require advanced geospatial modelling approaches capable of capturing nonlinear ecological interactions across heterogeneous landscapes. However, many wetland studies in sub-Saharan Africa rely on single-classifier approaches and limited predictor variables, resulting in reduced classification reliability and weak ecological interpretability.MethodsThis study addresses this gap by comparatively evaluating Random Forest (RF), Support Vector Machine (SVM), and Classification and Regression Trees (CART) for integrated wetland modelling within a geospatial big data environment in the Driefontein Grasslands, Zimbabwe, a Ramsar-designated wetland ecosystem. Multi-temporal Landsat imagery acquired for 2015, 2020, and 2025 was processed using Google Earth Engine, Python 3.10, and QGIS 3.44.6 Solothurn within a scalable cloud-supported analytical framework. To improve wetland discrimination, a comprehensive suite of remotely sensed spectral indices was integrated into the modelling workflow, including NDVI, EVI, SAVI, OSAVI, MSAVI, SIPI, GCI, RECI, NDWI and MNDWI. Model performance was evaluated using Receiver Operating Characteristic (ROC) curves, Area Under the Curve (AUC) statistics, and inter-model Pearson correlation analysis.ResultsRF demonstrated superior predictive stability and discriminatory performance across all epochs (AUC = 0.880–0.891), followed by SVM (0.850–0.873), while CART exhibited comparatively lower performance (0.749–0.789) and structural divergence through negative inter-model correlations. Spectral-index analysis revealed progressive vegetation decline, hydrological fragmentation, increasing vegetation stress, and accelerated conversion of vegetated wetland surfaces to bare substrates by 2025, signalling intensifying anthropogenic disturbance and ecological degradation.DiscussionThe findings demonstrate that integrating multi-index remote sensing analytics with ensemble machine learning significantly enhances wetland detection accuracy and ecological interpretation, providing a transferable GeoAI framework for scalable wetland monitoring, ecosystem restoration planning, and evidence-based environmental policy in Africa.
