Rapid urbanization and agricultural expansion pressurize ecosystems and their services. Long-term assessments integrating land-use and land-cover change (LULCC), transition drivers, future land-system dynamics and ecosystem service (ES) trade-offs remain limited in rapidly urbanising Mediterranean regions. This study proposes an integrated spatial modeling framework combining LULCC mapping, machine learning-based classification, MLP–Markov simulation, transition driver analysis, and ES valuation (ESV) in İzmir, Türkiye (1985–2024) and project future landscape dynamics through 2075. Results indicate that agricultural land expanded by 517 km2, primarily through the conversion of rangelands and mixed vegetation, while artificial surfaces nearly tripled from 380 to 1073 km2. Rangelands declined by > 850 km2, whereas forest area increased substantially; however, the decline in ESV suggests that forest extent gains do not necessarily translate into improved ecological functionality. Future projections indicate artificial surfaces reaching 1566 km2 by 2075 and total ESV decreasing from USD8043 million to USD7229 million annually, a net loss of ~10%. Proximity to urban centres and industrial areas is found to be a dominant factor influencing urban expansion, whereas agricultural growth is associated with favourable topographic conditions. These findings demonstrate that integrating LULCC modeling with ES assessment provides a decision-support framework for sustainable land-use planning and ecosystem conservation.
Fuel moisture (FM) monitoring is constrained by limited sensor availability, low-resolution data, and inadequate methods, necessitating advanced machine learning (ML) approaches. This study applied tree-based ML models, random forest, extreme gradient boosting (XGB), and light gradient boosting, to estimate dead FM in Gangwon State, South Korea. Input variables included meteorological data, GK-2A AHI radiance, topography, and temporal features, with hyper-parameters optimized via five-fold cross-validation. Results showed high accuracy: RMSE values of 1.20–1.41% and R2 values of 0.94–0.96 for the 10-h FM estimates at 10-min intervals. For daily and hourly mean FM predictions, RMSE values ranged from 0.03 to 0.45% with R2 values of 0.99–1.0. Among methods, XGB yielded the highest performance, while ensemble approaches provided consistent accuracy. For wildfire-risk thresholds (FM < 10%), the models demonstrated high estimation performance, with metrics including an ETS of 0.78 and an accuracy of 0.95. These findings demonstrate the efficacy of ML methods and diverse datasets for FM estimation, offering guidance on model selection based on data availability and computational resources. Accurate FM data can significantly enhance wildfire monitoring, prevention, and related research.
To overcome the limitations of limited feature diversity and insufficient exploitation of inter-variable correlations in hyperspectral prediction of soil organic carbon (SOC) and nitrogen (N), we propose a novel framework that integrates multi-feature fusion (MFF) with multi-target regression (MTR). The framework combines spectral features (SF) from the LUCAS 2015 dataset with land classification features (LCF) and remote sensing features (RSF), and further enriches the feature space using outputs from single-target regression (STR) models. A weighted MTR CatBoost model (WMTR-CatBoost) is then developed to systematically evaluate the contributions of MFF and MTR. Experimental results show that WMTR-CatBoost achieves coefficient of determination ([Formula: see text]) values of 0.945 for SOC and 0.923 for N, representing relative improvements of 2.5% and 4.6% over the SF-based baseline, and gains of 1%–3.3% for SOC and 1%–8.1% for N compared with STR-based models. These findings confirm the effectiveness of the proposed MFF–MTR framework and highlight its potential for advancing large-scale soil nutrient monitoring and precision agricultural management.
Rising atmospheric CO₂ concentrations increase demand for spatially explicit, repeatable estimates of forest carbon stocks. Yet operational Earth observation (EO)-based above-ground biomass (AGB) products for temperate forests remain limited. This study presents the first national-scale EO-based forest AGB dataset for the Czech Republic, produced through a reproducible workflow combining cloud-based data processing and external modelling. Sentinel-1 radar backscatter and Sentinel-2 multispectral reflectance were processed in Google Earth Engine, with texture metrics derived separately. Forest AGB was modelled in Python using 107 in-situ reference plots, including DendroNetwork forest plots and non-forest low-biomass sites used to constrain the lower end of the biomass gradient. Spatial transferability was assessed using nested spatial cross-validation with 1 km blocking (5 × 5 GroupKFold). Random Forest, eXtreme Gradient Boosting, and LightGBM were evaluated. Random Forest was selected for map generation because of its robust operational behaviour and competitive spatially independent accuracy (R² = 0.69 ± 0.12; RMSE = 79.28 ± 16.14 Mg·ha⁻¹; MAE = 48.03 ± 10.20 Mg·ha⁻¹). The resulting maps represent national forest AGB conditions during the 2023 growing season. AGB estimates were converted to carbon stocks for comparison with national forest inventory data.
Due to the scarcity of high-quality training data and the limitations of traditional feature extraction methods, significant challenges exist in applying machine learning approaches to marine significant wave height (SWH) inversion using synthetic aperture radar (SAR). To this end, this paper proposes a novel method to construct a training dataset using Sentinel-1 SAR images over the Atlantic region in 2024, together with ERA5 reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF). A total of 12,853 paired datasets of Sentinel-1 and ERA5 were generated through spatiotemporal matching. These data were separated into training, validation, and testing datasets to develop a deep learning model named WaveFusionNet, incorporating residual structure and attention mechanism. The results show that the model inversion accuracy in the testing dataset reaches root-mean-square error (RMSE) 0.61 m, standard deviation (STD) 0.59 m, scatter index (SI) 0.28. In addition, validation at five independent buoy locations confirms high agreement between model predictions and buoy measurements, with a 0.54 m RMSE and a 0.51 m STD. It proves the feasibility of constructing the training dataset from ERA5 data and provides a new idea for the inversion of SAR ocean parameters.
Chalk grassland is a highly biodiverse but fragmented priority habitat in Europe, and its conservation is constrained by the lack of fine-scale, up-to-date habitat maps. In the UK, this challenge is compounded by limited systematically field data, especially where access to privately owned land restricts ecological surveys. This study developed a fine-scale mapping approach for chalk grassland in Surrey, UK, by integrating structured citizen science data, multi-temporal PlanetScope imagery, topographic variables, and soil information via random forest classification. Citizen science surveys, designed around UK Habitat Classification criteria and validated by Surrey Wildlife Trust, produced 155 confirmed chalk grassland presence quadrats and 669 high-confidence absence points. The models achieved high classification performance in both years, with overall accuracy of 0.9939 in 2023 and 0.9758 in 2024, alongside strong precision, recall, and F1-scores. Soil type and elevation were the most influential predictors, indicating that chalk grassland distribution is strongly shaped by stable environmental conditions, while spectral variables improved habitat discrimination within ecologically suitable settings. Independent validation using 116 long-term chalk grassland polygons supported the ecological realism of the predictions. The result produced 3 m habitat maps providing a promising and scalable framework for conservation planning, restoration targeting, and Local Nature Recovery Strategies.
The sun exposure on the northern and southern faces of the coffee planting rows may play a key role in recovering of coffee plants affected by frost damage. This study aimed to assess coffee plants with frost damage according to the solar exposure face along the planting rows. Three experimental areas, each with plantations of different ages (one, three and eight years), were used. Evaluations were performed on the northern and southern sides of the planting rows, where the frost damage were measured. The remotely piloted aircraft (RPA) was flown over the three areas to capture multispectral images, and the derived vegetation indices were analysed. Only plants with eight years of age showed differences in frost damage between the sun exposure faces. Greater of damage was observed in the north face, with 48% of plant compromise. For coffee trees of one and three years old, the sun exposure did not exert a significant influence on the levels of damage between the faces of the planting line. The use of remotely piloted aircraft equipped with multispectral cameras demonstrated high efficiency in identifying spatial variations of crop damage. Among the indices evaluated, SAVI and MCARI2 showed a strong correlation with frost damage levels.
Accurate estimation of particulate matter (PM) concentrations is essential for air quality management, environmental monitoring, and public health research, particularly in regions with limited ground-based monitoring networks. This study evaluated satellite-derived Aerosol Optical Depth (AOD) products from MODIS and VIIRS for estimating daily PM₂.₅ and PM₁₀ concentrations across Thailand during 2012–2025. Daily observations from air quality monitoring stations were integrated with AOD, digital elevation model (DEM), land surface temperature (LST), normalized difference vegetation index (NDVI), meteorological variables, week of year (WOY), and year. Random Forest (RF) models were developed and evaluated using training, validation, and five-fold cross-validation datasets. Log-linear regression analysis showed significant positive relationships between AOD and PM concentrations (p < 0.05), with MODIS AOD explaining 10.8% and 9.1% of the variability in PM₂.₅ and PM₁₀, respectively, compared with 4.8% and 4.1% for VIIRS AOD. RF models substantially improved predictive performance. Validation R2 values reached 82% and 81% for PM₂.₅ and 77% and 76% for PM₁₀ using MODIS and VIIRS AOD, respectively. WOY, relative humidity, AOD, and topographic factors were the most influential predictors. Overall, MODIS and VIIRS provided comparable and reliable estimates of PM₂.₅ and PM₁₀, supporting satellite-based air quality monitoring in Thailand.
Arid oasis landscapes are characterized by strong spatiotemporal heterogeneity and complex planting structures, posing substantial challenges for consistent land cover mapping across years and at a fine spatial scale. Korla City was selected as a representative arid oasis case. A 10 m resolution land cover monitoring framework for 2017–2024 was constructed by integrating 64-dimensional deep semantic embeddings from the geospatial foundation model AlphaEarth Foundations (AEF) with the Extreme Gradient Boosting (XGBoost) algorithm. An automated sample migration strategy was implemented through high-dimensional vector similarity mapping, and the classification decision logic was quantitatively interpreted using the SHapley Additive exPlanations (SHAP) method. The proposed framework exhibited strong robustness, with overall accuracy (OA) ranging from 0.88 to 0.97 and F1-score ranging from 0.87 to 0.97, outperforming public products including CLCD and ESA WorldCover. Principal component analysis (PCA) revealed that AEF embeddings provided a more distributed and lower redundancy representation of land cover changes than traditional optical features, with the first two principal components capturing only 38.18% of the total variance. SHAP attribution further identified a complementary mechanism in which shared embedding dimensions constrained the environmental background, while class specific dimensions enhanced the discrimination of spectrally similar types. These findings demonstrate that geospatial foundation model embeddings provide an effective feature representation for spatiotemporally consistent land cover monitoring in heterogeneous arid environments.
Semantic segmentation of high-resolution remote sensing imagery (RSI) is essential for applications such as land-cover mapping, urban monitoring, and disaster assessment, yet it remains challenging due to complex backgrounds, high inter-class similarity, and large object-scale variation. We propose MBS-Net (Multi-Scale Attention Boosted Segmentation Network), a Multi-Scale Attention-Enhanced Network built on an encoder-decoder architecture to address blurred boundaries, loss of fine structures during downsampling, and insufficient multi-scale context modeling. The encoder integrates a ResNet-50 backbone augmented with Squeeze-and-Excitation (SE) blocks to recalibrate channel responses and highlight discriminative features. To reduce semantic misalignment in skip connections, we introduce a Fine-Scale Edge Attention (FSEA) module that combines a spatial attention stream for boundary cues with a channel attention stream for small-scale structure preservation, enabling precise fusion of encoder and decoder features. In the decoder, an Efficient Multi-Scale Attention (EMA) module aggregates parallel contextual paths with varied receptive fields to balance global semantics and local detail. By jointly leveraging SE, FSEA, and EMA, MBS-Net simultaneously strengthens feature representation, preserves fine structures, and adaptively models cross-scale context. Experimental results on the ISPRS Potsdam and Vaihingen datasets demonstrate that MBS-Net achieves consistent and substantial gains in segmentation accuracy and boundary delineation.
Accurate and efficient detection of individual trees is crucial for sustainable forest management, particularly in large-scale production forests such as Indonesia’s pine (Pinus merkusii) plantations. Conventional manual counting of trees from aerial photographs is labour-intensive and prone to human error. This study aims to enhance the precision of pine tree detection using the local maxima (LM) method applied to UAV-based aerial imagery collected in Sagaranteun, Sukabumi, West Java. The LM algorithm identifies canopy peaks based on the canopy height model (CHM), with detection performance evaluated through two parameters: canopy height and window size. Four sample plots were analyzed to quantify detection accuracy using commission and omission error metrics. Results indicate an average commission error of 0.07, an omission error of 0.14, and an overall accuracy of 0.81, suggesting reliable detection performance. However, higher omission errors indicate potential under-detection in densely forested areas, emphasizing the need for further parameter optimization. The study demonstrates that the LM approach is a robust and adaptable method for automatic pine tree detection and can significantly accelerate forest inventory processes for Perum Perhutani and similar forestry institutions.
Accurate observations of water surface elevation (WSE) are essential for small waterbodies (WBs) due to their role in agriculture, biodiversity, and hydrology. This study (i) evaluates SWOT PIXC WSE accuracy over synthetic small WBs (0.5-50 ha) built along the shoreline of Lake L & eacute;man, Switzerland and (ii) assesses filtering criteria to improve WSE accuracy while preserving data. Pixel-scale accuracy is computed using SWOT pixels within a 50 m shoreline buffer, while WB-scale accuracy uses median pixel elevation per date. After outlier removal, pixels classified as water-edge (class 3) and open-water (class 4) provide the best accuracy, with a 1-sigma absolute error of 24.2 cm at the pixel scale. Aggregation at the WB scale substantially improves accuracy, reducing errors from similar to 17.4 cm (0.5 ha) to similar to 8 cm (50 ha). Eight filtering strategies were tested: a multi-variable filter achieved the highest pixel accuracy (19 cm) but excluded similar to 58% of pixels, whereas simpler filters based on one or two variables produced similar accuracy (similar to 20 cm) while removing only 17%-37% of observations, offering a more practical balance. At the WB scale, filtering is most beneficial for small WBs (0.5-3 ha), reducing error to 15.7 and 12.4 cm, respectively, while removing 12%-15% of days. For larger WBs (similar to 50 ha), all strategies achieve similar to 8 cm accuracy, so strict filtering provides minimal benefit.
Semantic annotation of remote sensing imagery aims to automatically assign a semantic category to each pixel based on a limited set of manually labeled samples. While supervised deep segmentation networks have achieved remarkable success in this task, their performance heavily relies on abundant pixel-level annotations, which are often prohibitively expensive to obtain. In practical scenarios, the scarcity of annotated samples leads to severe overfitting, significantly degrading generalization ability. To address this challenge, we propose a spectral super-resolution enhanced deep annotation framework that advocates enhancing the spectral representation of the input imagery as a complementary direction to network architecture design, with the goal of improving pixel discriminability in few-shot annotation scenarios. Specifically, motivated by recent advances in transforming RGB images into hyperspectral representations using deep neural networks, we design a heterogeneous multi-task learning framework that jointly trains an image annotation network and a spectral super-resolution network in an end-to-end fashion, with a shared spectral reconstruction module. Furthermore, a joint optimization strategy is integrated with a semi-supervised clustering algorithm to fully exploit the limited supervision signals from scarce labeled data. Extensive experiments on four benchmark datasets demonstrate that our method consistently outperforms state-of-the-art few-shot annotation approaches.
Accurate monitoring of dynamic changes in vegetation cover is of great significance for balancing ecological and agricultural development in Northeast China. Therefore, in this paper, a comprehensive detection system was constructed for vegetation cover dynamic changes that includes trend and mutation detection. The BEAST-Optimization (BEAST-OP) was developed for detecting mutation of the time-series Enhanced Vegetation Index (EVI) data in this paper. Furthermore, the co-occurrence relationship between mutations and climate has been discussed. Meanwhile, the Theil-Sen median and Mann-Kendall methods were combined to construct the Sen-Mann-Kendall (SMK) trend analysis algorithm, which was used to detect the overall trend of vegetation cover in the stable vegetation type area of Northeast China. The results showed that: (1) From 2001 to 2022, 82.76% of the vegetation in the stable vegetation type area of Northeast China showed an increasing EVI trend. (2) The EVI of the vegetation cover areas in Northeast China was dominated by abrupt declines in 2003, 2009, and 2012, and the months of abrupt declines were mainly from April to September. (3) Compared to the BEAST, BEAST-OP significantly improved the reliability of mutation detection. Statistical analysis revealed a significant co-occurrence relationship between abrupt declines and climate disasters in some vegetation zones.
Evapotranspiration is an important process in croplands, linking water, energy, and carbon cycles. While direct measurement is resource-intensive, remote & times; sensing-based models offer spatial estimates but often require local calibration due to static parameterization. This study compares two energy balance models, Two-Source Energy Balance (TSEB) and Priestley-Taylor Jet Propulsion Laboratory (PT-JPL), and explores combining the better-performing model with machine learning to enhance accuracy without ground-based calibration. The models were driven by Landsat 7-9 and ERA5-Land reanalysis data and evaluated against eddy covariance data from 137 quality-filtered cropland stations in a wide range of climates. The PT-JPL model outperformed TSEB (R-2 = 0.39, RMSE = 86.8 W m(-2), similar to 25% lower RMSE) and was selected for optimization via a gradient boosting algorithm. A rigorous 70%-30% train-test split with spatial cross-validation prevented data leakage and overfitting. Five key residual drivers were identified: SAVI, NDMI, NDVI, relative humidity, and surface net solar radiation. Relative humidity was the most important predictor, with low values causing the largest PT-JPL errors. The final hybrid PT-JPL model showed improved performance (test: R-2 = 0.63, RMSE = 66.1 W m(-2); train: R-2 = 0.69, RMSE = 62.2 W m(-2)). This hybrid approach provides a refined method for spatial evapotranspiration estimation and offers interpretable insights into the environmental factors underlying PT-JPL errors.
Chinese high-resolution Earth observation satellites have developed rapidly over the last two decades and have made continuous progress in terms of their mapping capabilities. In this study, we analyzed the geometric accuracy potential of Jilin-1 satellite imagery covering the area around the Marmara Sea of T & uuml;rkiye. The Jilin-1 image was received as four quadrants, comprising 19 multispectral bands. We used different mathematical models in the accuracy assessment phase, including polynomial transformation, 2D- and 3D-affine projections, projective, DLT, and RFM methods. The reference datasets consisted of the points collected over Jilin-1 images, and their object coordinates were measured on the HGM-K & uuml;re. Then, these image and ground coordinates points were input into the GeoEtrim program, in which various transformation models can be implemented. The findings demonstrate that the highest accuracy values were obtained with the RFM using the panchromatic B0 band with 5 m GSD, reaching the sub-pixel accuracy level in both the row and column directions after applying second-order terms. A similar tendency was observed with the B7 and B15 bands with 10 and 20 m GSD, respectively. Above all, it was shown that it is possible to achieve higher accuracy with Jilin-1 imagery, and this level of accuracy meets the requirements needed for topographic mapping applications.
Ground-mounted photovoltaic systems are expanding rapidly to meet decarbonisation targets, but their growth raises concerns about land take, farmland conversion, and biodiversity impacts. Addressing the lack of tools to monitor local land-use change across Italian municipalities, this study presents an open-access application developed on Google Earth Engine. Through an interactive interface, users can select an Italian municipality, define the year and compositing method for Sentinel-2 imagery, draw training and validation polygons, and choose among three classifiers to generate land-cover maps. The tool automatically evaluates classification accuracy, filters pixels, and converts the photovoltaic class into vector polygons. Users can then select the dataset (CORINE Land Cover + Backbone or EUCropMap) and reference year to reconstruct previous land cover and agricultural use. All results, including classified maps, photovoltaic polygons, summary tables, and charts, can be exported. Developed for Montalto di Castro (Lazio), the workflow achieved 91.05% overall accuracy and mapped 762.14 ha of installations, covering 4.02% of the entire municipal territory. Results show that most installations replaced herbaceous farmland. The workflow was successfully tested in the municipality of Guillena (Andaluc & iacute;a, Spain), confirming its adaptability. The application offers a practical, scalable solution for quantifying photovoltaic expansion and supporting spatial planning in Italy and across Europe.
Forests are imperative to sequester carbon, conserve biodiversity and regulate climate, making forest site suitability a prerequisite for sustainable ecosystem restoration. However, efficient planning requires a generalizable methodology that moves beyond the constraints of single-model or species-specific approaches to identify land with high forest growth potential. This study proposed a framework that ensembled the predictive performance of Random Forest (RF), the probabilistic strength of Maximum Entropy (MaxEnt) and the expert-driven weighting of Analytical Hierarchy Process (AHP) to provide forest site suitability analytics using key edaphic, climatic, and topographic parameters. RF achieved high predictive accuracy (R & sup2; = 0.87), with MaxEnt (AUC = 0.86) and AHP overall accuracy of 77.8% for validation of spatial coherence. Evaluation against existing forest cover validated the model's reliability with (24,210 km2 52.9%) of forests located in highly suitable zones and only (1742.7 km & sup2; 0.6%) in the least suitable zone. The study identified major forest expansion zones in Balochistan (42,767.7 km & sup2;), Punjab (29,692.5 km & sup2;) and Sindh (8576.8 km & sup2;), indicating that approximately (16%) of the analyzed area is suitable for forestry. This resource-optimized, actionable spatial decision-tool enables stakeholders to prioritize afforestation investments in high-potential, low forest-cover zones maximize carbon sequestration and achieve Sustainable Development Goals 13 and 15.
Accurate estimation of forest growing stock volume (GSV) at fine spatial scales is essential for sustainable management, carbon accounting, and local decision-making. However, traditional inventories often lack sufficient sampling density for small areas. This study evaluates two small-area estimation (SAE) approaches: the Empirical Best Predictor (EBP), based on a nested-error linear regression model, and the Mixed-Effects Random Forest (MERF), using multi-source remote sensing data. The analysis was conducted in the Vallombrosa Nature Reserve (Italy), integrating field measurements from 101 plots with auxiliary variables from Sentinel-2 and airborne LiDAR. Both methods estimated mean and total GSV across 658 forest stands, many lacking direct observations. Performance was assessed via spatial cross-validation, and uncertainty was quantified using root-mean-square error (RMSE). Results show that MERF outperformed EBP in predictive accuracy, achieving higher R & sup2; (0.67 vs. 0.37) and lower RMSE (151 vs. 202 m & sup3; ha(-)& sup1;). MERF produced more stable uncertainty estimates with improved coverage. While both methods yielded comparable total GSV, EBP exhibited greater sensitivity to model assumptions. Conversely, MERF effectively captured non-linear relationships and handled multicollinearity, despite reduced interpretability and higher computational demand. Overall, findings highlight the advantages of integrating machine learning with mixed-effects modeling for SAE under sparse sampling.
Nighttime light (NTL) remote sensing imagery supports diverse applications, from demographic analysis to urban studies, but coarse spatial resolution (500 m) of NPP-VIIRS NTL data limits fine-scale analyses. This study presents a CNN-based downscaling approach to enhance NTL data resolution.The proposed method improved NPP-VIIRS resolution from 500 m to 120 m, yielding more detailed spatial information and clearer boundaries. Evaluated via urban built-up area extraction, the downscaled data outperformed original NTL data in overall accuracy and spatial detail, effectively mitigating resolution-related limitations.The enhanced-resolution NTL data demonstrates strong potential for detailed urban mapping and socioeconomic estimation. Future work will address multi-source temporal consistency challenges through real-time social sensing integration and advanced fusion techniques.
