Environmental and Earth Sciences

Abstract: Flooded cropland in Sindh had not returned to its counterfactual vegetation condition four years after the 2022 floods, and the shortfall sits almost entirely in the summer (Kharif) season. We built 555,309 300-m cropland cells across 22 districts and tracked monthly Sentinel-2 normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI) from January 2021 to September 2026. Cells with three or more Sentinel-1 water detections during August to October 2022 were treated as flooded and matched one-to-one with non-flooded cells on terrain, drainage proximity, cropland fraction, and pre-flood vegetation level and trend, giving 87,447 pairs with all seven covariates balanced (absolute standardized mean difference below 0.03). An event-study difference-in-differences model shows no pre-flood divergence (k = −1 coefficient +0.003, p = 0.43). Flooded cells fell 0.20 NDVI below controls in Kharif 2022, recovered to −0.06 by Rabi (winter) 2023–24, settled at about −0.11 through Kharif 2026. The result survives a placebo shock, three calipers, leave-one-district-out estimation, and a monotonic dose-response across flood-duration terciles (−0.13, −0.19, −0.28). Integrated Kharif NDVI in flooded cells stayed 0.30 to 0.60 units below controls in every summer from 2023 to 2026, whereas integrated Rabi NDVI exceeded controls from the second winter onward (flood duration × Kharif interaction, γ = −0.32, p = 4.4 × 10−10). Flood duration and post-flood standing-water persistence were the strongest predictors of time to recovery across mixed-effects, spatial, and gradient-boosted models. Trajectory clustering places 17% of matched cells in a persistent-deficit archetype; Larkana and Dadu hold 36% of them.

Abstract: Sea ice classification models trained on manually labeled ice charts from public ice services are often limited by operational constraints and human annotation practices. These charts typically generalize spatial boundaries and provide a conservative representation of ice features. To address these challenges, we employ two label refinement strategies: spectral water masking and vector-to-raster registration. Specifically, spectral water masking resolves fine-scale open water features, while vector-to-raster registration aligns chart boundaries with satellite imagery to correct for sensor misalignment and temporal ice drift. To isolate the impact of these refinements, we apply them to Baltic Sea ice charts provided by the German Ice Service, utilizing the same dataset and neural network architecture as prior work. In controlled and balanced experiments, we find that these spectral and semantic refinements increase the Intersection over Union (IoU) from 0.68 to 0.77, with the incorporation of thermal emission channels providing further improvement to 0.81. The substantial increase in IoU and qualitative inference results demonstrate that the model more accurately captures physical boundaries and fine-scale ice structures. While this study focuses on Landsat-8 multispectral imagery, the vector-to-raster registration is sensor-agnostic and readily applicable to other optical or SAR platforms, such as Sentinel-2 and Sentinel-1.

Abstract: Retrospective mapping of aboveground biomass (AGB) using long satellite archives offers a valuable approach for assessing post-fire biomass change, but contemporary calibration does not necessarily ensure reliable transferability to historical conditions. The Global Ecosystem Dynamics Investigation (GEDI), a spaceborne lidar mission that provides observations of forest structure and biomass estimates, offers a potential source of contemporary reference data for model calibration. We evaluated this issue using a GEDI-calibrated Landsat workflow for the landscape affected by the 2009 Churchill wildfire in southeastern Australia. Random forest models combined optical, environmental and pre-fire predictors using 4,746 GEDI-linked reference observations from 2020–2023. Evaluation included reference-provenance auditing, spatial cross-validation across nine blocks with 300 m training exclusions, forward temporal and spatiotemporal tests within 2020–2023, paired predictor ablation, conditional-error diagnostics and an external field comparison. All reported provenance checks passed, supporting reference-processing integrity. The full model outperformed the training-mean baseline and reduced configurations in every block, although improvements were modest. Predictions showed a compressed range, with a prediction-to-reference standard-deviation ratio of 0.362 and regression slope of 0.104, overpredicting low reference biomass and underpredicting high reference biomass. Temporal and spatiotemporal transfer from 2020 to 2023 failed to outperform their corresponding training-mean baselines. Annual predictions declined after the fire and subsequently increased, but biomass loss and recovery were not independently validated. These findings support using the maps as descriptive reconstructions and highlight the need for independent structural observations across disturbance stages before quantitative recovery assessment or carbon accounting.

Abstract: Lithological identification and mapping are crucial in understanding geological processes, mineral exploration, natural resource management, land-use planning, and other applications. Conventional field-focused mapping of geology is often requiring a lot of time and expensive, particularly in remote and rugged terrains, like in our area. Recent advances in satellite remote sensing provide efficient tools for lithological discrimination. This study explores the application of Sentinel-2 multispectral imagery to develop a lithological map in a case of Meda kebele. The Sentinel-2 satellite image was classified by using support vector machine (SVM) classification system to prepare a excellent geological map, which was validated by previous geological maps and field surveys For this purpose, several band combinations, spectral indices, and image processing methods, including OIF, ICA and PCA, were used in order to enhance the contrast among different rocks. Following that, a supervised classification supported by field survey and previous lithological maps of the area was carried out to create a lithological map. The results show that Sentinel-2 imagery can distinguish between the common geological units found in the area, which include metavolcanics, basalt, dolomite, limestone, marble, slate and sandstone, due to it has good spatial resolution of 10- 60 m and vast spectral extent from VNIR-SWIR. This paper highlights the significance of Sentinel-2 data for regional geological mapping, giving baseline information for mineral prospecting and resource management in Meda kebele.

Abstract: Land subsidence threatens infrastructure, ecosystems and coastal communities, yet predicting its spatially heterogeneous evolution from long-term interferometric synthetic aperture radar (InSAR) imagery remains challenging. Here, we propose the Large-scale Geospatial Time-series Model (LGTM), a spatiotemporal fusion network that combines a dual-path Informer-p temporal encoder with a simplified U-Net spatial decoder to model long-range temporal dependencies and multiscale spatial patterns. Using 136 cumulative SBAS-InSAR deformation images acquired from 10 January 2018 to 1 June 2023 in the Yellow River Delta, China, we evaluated three approximately 943.7 km² subregions representing circular subsidence funnels, irregular patches and radially contiguous subsidence clusters. Under identical data partitions, normalization procedures, training protocols and evaluation masks, LGTM was compared with ConvLSTM, TimesNet and BiLSTM, while ablation experiments examined Informer, Informer-p and Informer+U-Net. LGTM achieved the highest structural similarity index (SSIM) in all three regions, whereas histogram intersection over union (IoU) and normalized cross-correlation (NCC) were highest or tied for highest. The largest relative reduction in root mean square error (RMSE) and increases in SSIM, IoU and NCC were 36.90%, 7.57%, 3.10% and 6.39%, respectively. Although LGTM did not achieve the lowest RMSE in every region, it consistently outperformed Informer+U-Net across all four metrics in the ablation experiments. Informer-p alone did not consistently outperform the original Informer, indicating that the observed gains were associated primarily with the joint temporal-spatial fusion. These results demonstrate that LGTM provides a balanced approach to numerical prediction and preservation of complex subsidence structures in regional InSAR time-series forecasting.

Abstract: On December 4, 2025, an M_W 5.8 earthquake struck in Aheqi County, Xinjiang. Its epicenter is located approximately 24 km southwest of the 2024 Wushi M_W 7.0 earthquake, and the seismic region lies in the hinterland of the Kalpintag fold-and-thrust belt on the western segment of the South Tianshan seismic zone. To clarify the seismogenic structure properties of this earthquake and its dynamic correlation with the strong Wushi M_W 7.0 earthquake, this paper constructs a time-series observation dataset covering 22 months before the earthquake to 5 months after the earthquake based on ascending and descending Sentinel-1A SAR images, and extracts the time-series deformation trend of Aheqi 22 months before the earthquake (after the Wushi earthquake) using the SBAS-InSAR technology. A step term is introduced into the time-series model to eliminate local unwrapping errors and residual atmospheric delays caused by severe convective weather in the summer of June 2025, so as to obtain a reliable interseismic creep rate. Combined with the co-seismic line-of-sight deformation field acquired by the two-pass D-InSAR method, a comprehensive study is carried out by integrating distributed fault slip inversion and Coulomb stress change simulation. The results show that: (1) The co-seismic deformation is dominated by uplift, with the maximum LOS deformation of approximately 5.0 cm for the ascending orbit and approximately 6.0 cm for the descending orbit. The deformation field presents a significantly asymmetric elliptical shape with no surface rupture, indicating that this event is a blind thrust rupture event; (2) The SBAS time-series analysis for the 22 months before the earthquake shows that after removing the observation disturbances caused by severe convection in June 2025, the tectonic rate of the hanging wall is approximately +26.9 mm/yr, that of the footwall is approximately +8.5 mm/yr, and the across-fault rate difference is approximately 18.4 mm/yr. After deducting the influence of post-seismic adjustment of the Wushi earthquake, it may indicate that the Aheqi fault was in a locked/strain-accumulating state before the earthquake, with continuous accumulation of elastic strain; (3) The inverted seismogenic fault has a strike of approximately 242°, a dip angle of approximately 40°, a slip angle of approximately 74°, a centroid depth of approximately 6.2 km, a maximum slip of approximately 0.2 m, and a moment magnitude of M_W 5.9. The rupture is dominated by thrust motion with a weak left-lateral strike-slip component, confirming that the Aheqi fault is the seismogenic structure of this earthquake; (4) The Coulomb stress calculation results show that the 2024 Wushi M_W 7.0 earthquake produced a static stress loading of more than 2 bar in the source region of the Aheqi fault, far exceeding the earthquake triggering threshold of 0.1 bar. Combining the pre-seismic locking characteristics, co-seismic slip distribution and Coulomb stress calculation results, it is considered that the Aheqi earthquake is a moderate blind thrust event triggered by post-seismic stress adjustment of the Wushi M_W 7.0 earthquake under the background of long-term tectonic loading. The results of this study can provide measured deformation constraints and reference for analyzing the seismogenic mechanism of the structural transition zone in the piedmont of South Tianshan and improving the regional seismotectonic model.

Abstract: Water is crucial for rice growth, particularly in semi-arid Australia, where climate change threatens water allocation for rice growers and challenges the sustainability of the rice industry. Estimating water depth using remote sensing can be beneficial, as it can help growers transition from traditional continuous flooding (CF) to irrigation practices such as Delayed Permanent Water (DPW) and Aerobic Rice (AR), which are promising pathways to ensure industry sustainability. Synthetic aperture radar (SAR) has been widely evaluated for water depth prediction in rice crops. To date, a significant gap remains in the performance of the proposed models, with authors suggesting the use of optical and SAR data to improve their estimates. To this end, this research evaluated three machine learning (ML) models to estimate water depth in rice under different irrigation practices in the Australian semi-arid region. Three different ML approaches (Tree-based, Neural Networks, and pretrained foundational model) were evaluated with three data configurations: Set 1 with SAR only, set 2 with Optical Only, and Set 3 with multimodal (SAR and Optical) data. The pre-trained foundational model achieved the best results, with RMSE of 14.19 cm, MAE of 10.01 cm, R² of 0.63, KGE of 0.73, and CCC of 0.78. Interestingly, the SHAP analysis reveals that optical remote sensing is more significant for the model than SAR. Specifically, the mean GVMI and VAWI_log standard deviation.

Abstract: While nighttime-light (NTL) data are increasingly used for disaster assessment, different indicator formulations may represent distinct dimensions of disruption and are rarely validated before cross-area comparison and spatial modeling. This study aimed to establish whether alternative NTL indicators derived from the same observations measure the same dimension of disruption, and then to identify the spatial correlates of flood-induced activity disruption and early recovery using the indicators that proved valid. This study examined county-level flood disruption and early recovery during the 2023 Haihe River Basin flood using daily Visible Infrared Imaging Radiometer Suite observations across the Beijing–Tianjin–Hebei region. Three disruption indicators were evaluated against official disaster records, and the retained indicators alongside an early recovery ratio (ERR) were subsequently analyzed using multiscale geographically weighted regression. Spatial disruption extent (SDE) and aggregated disruption magnitude (ADM) strongly discriminated documented flood impacts (area under the curve [AUC] = 0.931 and 0.865, respectively), whereas affected-area disruption intensity performed poorly (AUC = 0.623) and was excluded from spatial modeling. After controlling for baseline luminance and screening spatial associations for robustness, young-child population ratio was positively associated with SDE, mean slope (MS) was negatively associated with SDE and ADM, and cumulative precipitation was negatively associated with ADM. MS was the only retained correlate of early recovery, with steeper terrain associated with lower ERR regionwide. These findings show that NTL indicators derived from the same observations measure distinct dimensions of disaster response and are not interchangeable; each should be interpreted based on the dimension it measures in NTL-based disaster assessment.

Abstract: This study advances global gas-flaring monitoring from a sequence of annual inventories to a monthly, source-resolved, multi-satellite record of upstream gas flaring from March 2012 through December 2025, using VIIRS Nightfire observations from SNPP, NOAA-20, and NOAA-21. The analysis integrates a persistent multiyear flare catalog, a calibration derived from controlled flare tests relating radiant heat to gas flow rate, and an aggregation framework that separates clear non-detections from cloud-affected, sunlit, or otherwise invalid observations. The framework yields consistent monthly and annual estimates from individual flare sites to country and global scales. Global upstream flaring reached 164.5 BCM in 2025, the third consecutive annual increase, and was highly concentrated: 4.3% of active sites accounted for 80% of the global volume. Summed monthly estimates are internally consistent with direct annual estimates (median ratio 98.1%) while revealing seasonal patterns and short-term changes hidden in annual inventories. Global estimates from the three VIIRS instruments agree within 1–4%, and comparisons with regulatory records for Texas (Permian and Eagle Ford) and New Mexico, regions with reliable reporting but atypically small and intermittent flares, show strong temporal agreement (monthly r = 0.85–0.91) and recovery of 69–96% of reported volumes. Comparisons with legacy estimates further show that differences among flaring products reflect combined changes in source coverage, calibration, satellite sampling, and aggregation rather than a uniform rescaling of annual totals. Cloud-covered overpasses must be excluded rather than counted as non-detections, which would lower global volumes by 27–30%, and the catalog must be refreshed every 6–12 months to remain representative.

Abstract: Change detection in high-resolution remote sensing imagery is susceptible to adversary effects of within-class spectral heterogeneity, shadows, and complex spatial structures. Meanwhile, change detection methods based on land-cover posterior probabilities generally rely on reliable semantic training information, making complete manual interpretation costly. To address these issues, this study develops a change detection framework that integrates SEEDS-guided SAM object construction and active learning-based interpretation with FCM-CSBN-CVAPS posterior probability inference. First, representative points of SEEDS superpixels are used as prompts for SAM to generate candidate objects and construct a mutually exclusive object partition. Second, based on the spectral statistical features of the superpixels and an XGBoost classifier, active learning is employed to select highly uncertain samples for limited manual annotation, thereby obtaining object-level semantic information. Finally, FCM is used to characterize pixel-level spectral uncertainty, while CSBN incorporates spatial context to estimate posterior probabilities of land-cover classes. Bi-temporal change measurement and changed-area extraction are then performed in the posterior probability space. Experimental results show that the proposed framework achieves stable change detection performance with only approximately 10% manual annotation. On DS1, DS2, and DS3, the Kappa coefficients reach 83.87%, 86.05%, and 90.21%, respectively, while the corresponding F1 scores are 87.62%, 91.78%, and 91.31%, outperforming the compared methods. These results indicate an effective balance between false-alarm suppression and missed-detection control. In adaptive evaluation experiments, the mean Kappa and F1 scores reach 84.73% and 87.88%, respectively, further confirming the effectiveness and stability of the proposed framework across diverse remote sensing scenarios.

Abstract: Forests worldwide face intensifying natural and anthropogenic pressures that progressively erode their resilience, bringing them closer to abrupt, potentially irreversible shifts. Such loss of resilience can manifest as critical slowing down (CSD), a phenomenon characterized by slower recovery rates. It is often measured through statistical indicators, known as early warning signals, which are expected to rise in advance of a critical transition such as a tipping point. Earth Observation (EO) is well suited to monitoring CSD indicators at global scales. However, the reliability of these indicators depends strongly on the quality and consistency of the underlying satellite record given their derivation from optical biophysical proxies. This is particularly true for what is arguably one of the most useful EO dataset – the Landsat Collection 2 archive, given its extensive temporal coverage (1984-present) and 30-m spatial resolution. Here, we present a scalable and reproducible workflow for generating high-quality, long-term, cross-calibrated Landsat datasets specifically designed for the computation of statistical indicators of resilience over large extents of forest ecosystems. The workflow integrates robust imagery processing, consistent radiometric cross-calibration, and a multiscale grid-based implementation to improve the reliability of the Landsat imagery used for resilience analysis. Using an undisturbed boreal forest region in Quebec, Canada, as a case study, we demonstrate that the proposed workflow reduces inter-sensor discontinuities, enhances spatial consistency and preserves ecologically meaningful temporal dynamics across the 40-year Landsat record. Applied across ecoregions and biomes, this framework seeks to enable robust and comparable assessments of forest resilience, supporting large-scale ecosystem monitoring, early-warning assessments, and evidence-based conservation prioritisation.

Abstract: Emergency forecasts benefit from scene detail when it constrains the relevant physical process and reaches decision-makers in time. This critical review examines flood inundation, wildfire spread and marine drift through 148 sources and 16 purpose-selected study families. Within this selected corpus, four families develop dedicated 3D conversion, chiefly for hydraulic-domain or fuel-material preparation; the coding rule excludes routine terrain input and existing ocean analyses. Forecast cases mainly update states, controls or initial trajectories. Hydraulic connectivity can alter local inundation. Corrected fire fronts remain exposed to propagation error, while wider marine search regions improve coverage at increased search cost. These mechanisms explain why geometric accuracy and analysis fit alone cannot establish predictive value. We propose a framework that selects scene attributes, update targets and observation timing jointly. Matched alternatives share a future response quantity, information cutoff and resource budget; deployment-specific tolerances connect their comparison to evacuation requirements and search capacity. Physically coupled neural representations require calibrated material properties and an online delivery-time account. The proposed tests assess the forecast effect of additional scene detail, the persistence of a correction and the time left for response.

Abstract: In the task of building extraction from high-resolution remote sensing images, challenges remain due to large variations in building scales, confusion between buildings and complex backgrounds, and oversmoothing of boundary details. Existing deep learning methods based solely on convolutional neural networks (CNNs) or Transformers often fail to simultaneously capture local geometric details and global semantic consistency. This study proposes a dual-branch building extraction network that combines CNN-based multi-scale spatial features with Transformer-based global semantic features. An attention-guided multi-scale feature fusion module is designed to alleviate feature conflicts and to suppress complex background noise. A boundary-aware loss and a multi-scale deep supervision strategy are introduced to enforce geometric constraints on building contours and intermediate features. Experiments on the Gaofen-7 (GF-7) dataset show that the proposed method achieves an IoU of 0.8065 and an F1 score of 0.8929, outperforming the compared methods. When using only 50% of the training samples, the model still attains an IoU of 0.7701, indicating good stability and application potential under limited labeled data. The method enables fine extraction of buildings from high-resolution remote sensing images and provides technical support for geographic information applications such as land-use survey, urban planning, and disaster assessment.

Abstract: An integrated signal-processing, performance-characterization, and atmospheric-measurement framework is presented for an all-fiber coherent Doppler lidar operating near 1.5 µm for wind sensing and atmospheric boundary-layer observations. The lidar uses pulsed heterodyne detection at a pulse repetition rate of 20 kHz, with atmospheric return signals sampled at 400 MS/s and preprocessed in real time using a field-programmable gate array (FPGA). Two embedded processing architectures are consolidated: direct range-gated fast Fourier transform (FFT) processing with accumulated power spectra, and digital in-phase/quadrature demodulation followed by down-sampling and autocorrelation accumulation. The autocorrelation approach preserves flexibility for subsequent range-dependent processing. Receiver performance is characterized by normalizing measured spectra to a reference local-oscillator shot-noise spectrum to compensate for the non-flat receiver response. For 10,000-pulse accumulation, the measured normalized noise fluctuation is approximately 0.015–0.0155, compared with an idealized shot-noise value near 0.014, and a normalized detection threshold of approximately 0.024–0.025 is used. Range-dependent coherent return is interpreted through heterodyne efficiency and loss of spatial coherence caused by distributed aerosol scattering and atmospheric refractive-index fluctuations. Range-corrected coherent-lidar measurements are compared with direct-detection lidar observations, while coordinated microwave-radiometer measurements provide thermodynamic context. The combined results show that FPGA processing, statistical receiver characterization, range-dependent coherence analysis, and multi-sensor observations form a unified framework for quantitative coherent Doppler lidar measurements of urban boundary-layer wind and turbulence.

Abstract: Burned area mapping (BAM) provides essential spatial information for fire monitoring, emissions accounting, and ecological impact assessment, but reported accuracy cannot be interpreted independently of mapping targets, reference data, validation design, and application context. We systematically reviewed 1339 remote sensing BAM studies published from 1999 to 2026 to examine technological advances and the evidence supporting validation, transferability, and downstream use. Existing studies address multiple targets, including burned extent, total burned area, fire perimeter/event, burn date, and fire progression, each of which requires distinct evaluation criteria. High-spatial-resolution data, dense time series, multi-source fusion, and advanced learning methods have expanded mapping capabilities, yet small-fire omission, mixed pixels, cloud and smoke interference, confusion with non-fire agricultural disturbances, reference-data limitations, and long-term nonstationarity remain important sources of error. Most studies emphasize local validation, whereas explicit external-transfer tests are comparatively scarce, and random pixel or patch splits do not demonstrate transfer across fire events, regions, years, or sensors. Evidence that BAM errors alter downstream outcomes is even more limited. BAM reliability should therefore be interpreted in relation to the mapping target, tested conditions, transfer evidence, and task-specific fitness for use rather than a single accuracy metric.

Abstract: Retrieval of soil organic carbon (SOC) from uncrewed aerial vehicle (UAV) hyperspectral imagery depends on how laboratory measurements are linked to pixels at centimetre resolution. This study presents a quality-controlled workflow in which radiometric calibration, harmonisation against proximal spectroradiometer measurements, pixel-level screening, spectral support definition and model validation are treated as interdependent components of a single procedure. Imagery was acquired with a tunable Fabry–Pérot interferometer camera in six bands between 670 and 1100~nm over eight bare soil fields in Flanders, Belgium (85 topsoil samples) and five fields in Central Macedonia, Greece (42 samples). Single pixel spectra, neighbourhood aggregation and local pixel augmentation were compared at radii of 0.25, 0.50 and 1.00~m using Random Forest, XGBoost, a 1D convolutional neural network and TabPFN, with all pixels from a given sampling location assigned to the same cross-validation fold. In Flanders, 23 of 24 neighbourhood configurations reduced RMSE relative to single pixel extraction, the best reaching \(R^2 = 0.784\) and RMSE =0.155% SOC. In Central Macedonia, adverse acquisition conditions and a narrow target range prevented a usable SOC model, although electrical conductivity, CaCO₃ and pH reached \(R^2\) between 0.61 and 0.69. These results indicate that the spectral support should be defined to match the field sampling support, and that this correspondence belongs alongside radiometric calibration, pixel screening and validation in a reproducible workflow for field-scale soil mapping.

Abstract: Industrial parks and company growth are vital for urban planning and development. With temporal data collection, mathematical model based growth predictions of industrial parks becomes feasible. In this study, an industrial park Land development Simulation and Company Growth Projections (LSGP) framework is proposed by applying satellite data for land development simulation and social big data for company growth analysis. The cellular automata model is developed to simulate land use development, and the logistic function is applied to set up company growth projection model for selected industrial parks. Results show that the LSGP framework can effectively simulate industrial park land expansion and growth patterns with the following main findings: (1) Underlying land growth follows similar patterns, and key influencing factors are similar, such as traffic convenience, adjacent to urban centers; (2) Company growth pattern follows a logistic growth curve, with each subgroup of companies following the same pattern; (3) Models developed based on a selected industrial park can be transferred to another, i.e., case-based studies can be applied to newly planned industrial parks; (4) The LSGP framework may be further developed to simulate various land use projects and emulate growth patterns for additional areas of interest.

Abstract: Analysing the spatial and temporal variability of near-surface air temperature (T_a) is essential for assessing climate change, particularly in polar regions such as the maritime Antarctic, where observational studies remain comparatively scarce. In areas with sparse meteorological networks, satellite observations provide a valuable means of extending station-based analyses and generating spatially distributed Ta maps. This study analyses updated T_a records from six meteorological stations on the Hurd Peninsula, Livingston Island, for 2000–2024, with particular attention to temperature changes since 2016, and evaluates T_a estimation models based on MODIS Terra daytime LST Collection 6.1 data. The analysis of the T_a time series showed significant warming over the complete 2000–2024 period, with trends ranging from 0.53 to 1.27 °C decade⁻¹, and a clear shift towards warming after 2016. While the 2000–2015 period showed spatially contrasting trends, including cooling at JCI, positive trends were observed at all analysed stations during 2016–2024, ranging from 0.40 to 1.33 °C decade⁻¹ and highlighting substantial local variability across the Hurd Peninsula. The global Ta model, developed using data from all six stations and based on MODIS Terra daytime LST and spatiotemporal variables, achieved a performance comparable to that obtained with Collection 6, with R² = 0.77 and a cross-validated RMSE of 2.23 °C. The results also showed that rigorous filtering of the MODIS LST observations is essential: retaining only good-quality data and excluding observations with |view angle| > 45° substantially reduced dispersion in the Ta–LST relationship and removed unrealistically low LST values. The introduction of this view-angle threshold represents a new filtering step with respect to our previous studies.

Abstract: Hyperspectral Image Classification is an importanttopic in remote sensing with numerous applications in precision agriculture, environmental monitoring, defence, and more. It helps classify material composition by mapping spectral signatures from each pixel. Despite its potential, it faces persistent challenges due to the scarcity of labelled data and high dimensionality. This makes it difficult to achieve robust performance that can generalize well. To address this problem, we propose SRSSF (Stability-Reliability Spectral-Spatial Fusion), a hybrid mixture-of-experts framework for hyperspectral image (HSI) classification. It fuses a compact, stability-selected spectral representation with full-spectrum spatial and joint spectral-spatial descriptors (extended morphological, attribute-like, and Gabor texture profiles), combined through validation-derived reliability weights rather than fixed or heuristic fusion rules. SRSSF is evaluated under a strict, footprint-level spatially disjoint protocol in which entire patch receptive fields are guaranteed non-overlapping across training, validation, and test partitions, not mere center pixels. It has an automated, geometry-only (label- and accuracy-independent) seed-acceptance procedure and requires a minimum dataset-specific test-set size for training. The framework is benchmarked against fourteen classical and CNN spectral/spatial baselines and a dedicated ablation suite, across four benchmark datasets (Indian Pines, Pavia University, Salinas, and the Kennedy Space Center (KSC)). Our framework achieves the best OA, AA, and κ on every dataset, with OA scores of 56.43,91.94,90.50, and 75.96 at Indian-Pines, Pavia University, Salinas, and KSC, respectively.

Abstract: Total suspended matter (TSM) is a key coastal water-quality indicator, yet the operational GOCI-II retrieval algorithm is constrained by an empirical ceiling of approximately 28.13 g m⁻³ because of the inherent dynamic-range limitation of visible band-ratio formulations in the current operational product; this limitation precludes accurate monitoring of turbid nearshore environments, which is critical for ecosystems and coastal management. Here, we developed and evaluated a physically interpretable stacking ensemble (PISE) model that fuses three complementary base learners via a ridge meta-learner on five physically interpretable GOCI-II Rᵣₛ (555, 620, 709 nm) features trained on 212 matchup samples (2022–2025). On an independent held-out test set, PISE achieves R² = 0.793 and RMSE = 3.670 g m⁻³, improving the operational GOCI-II product by 8.9 percentage points in R² and 16.3% in RMSE, with 90% conformal prediction intervals achieving 90.6% empirical coverage. On 4 May 2025 GOCI-II imagery, PISE captures the Gyeonggi Bay flood-ebb TSM cycle, with a +0.006 g m⁻³ per cm regression against the KHOA tidal-stage gauge. PISE thereby provides a quantitative satellite-derived characterization of diurnal TSM dynamics across the flood-ebb tidal cycle over Korean coastal waters, extending operational TSM retrieval above the 28.13 g m⁻³ retrieval ceiling up to its trained limit of 50 g m⁻³.