Environmental and Earth Sciences

Abstract: Low-light remote sensing images acquired by a TDI-ICMOS system under photon-limited conditions usually exhibit severe underexposure, amplified random noise, unstable local contrast, and unnatural highlight transitions. These challenges are more difficult in our target setting because the data are real 16-bit grayscale remote sensing images without paired clean references. We propose WISE-Net, a Wavelet-guided Illumination Self-supervised Enhancement framework for low-light remote sensing images. WISE-Net follows a two-stage fully self-supervised design. In Stage 1, the input image is decomposed by Haar wavelets, and only the low-frequency component is modulated by a learned brightness-adaptive gain while the high-frequency sub-bands are preserved for structure-consistent reconstruction. A composite self-supervised objective stabilizes enhancement, protects bright regions, and reduces gain artifacts. In Stage 2, the Stage-1 enhancement network is frozen and a self-supervised blind-spot refinement network suppresses enhancement-amplified noise and local artifacts. Experiments on three real low-light remote sensing test subsets show that WISE-Net achieves the lowest NIQE on all three subsets, the lowest PIQE on test1_2048×1440 and test2_4096×1440, and the highest entropy on all three subsets.

Abstract: Avalanche companion rescue provides the only realistic opportunity to locate and extricate a buried victim during the critical 10–15 min survival interval before organized rescue teams can normally arrive. Motivated by this stringent time constraint, this paper presents a physics-based waveformdomain framework for three-dimensional avalanche victim localization using a helmet-mounted Sparse Uniform Circular Array (SUCA) receiver and a compact cooperative RF beacon. Unlike conventional avalanche transceivers, which require sequential signal-search, coarse-search, and finesearch procedures, the proposed approach determines the victim’s three-dimensional coordinates directly by correlating the received waveform with a precomputed electromagnetic dictionary, enabling immediate guidance toward the estimated burial location. A comprehensive electromagnetic model incorporates realistic snow dielectric properties, propagation through snow and air, snow–air refraction, LOG-based deterministic multifrequency waveform generation, and dictionary-based localization. Numerical simulations demonstrate unique waveform representation throughout the investigated search region, complete recovery of all candidate victim locations under noiseless conditions, practical localization accuracy using a 0.5 m spatial grid, and a calculated 38 dB link margin under conservative very-wet-snow conditions. The proposed framework also supports several promising extensions beyond victim localization, including respiration monitoring, approximate body-orientation estimation, and enhanced victim detectability using passive conductive textile threads integrated into avalanche garments. Although demonstrated for avalanche companion rescue, the proposed waveform-domain localization framework represents a new operational concept for time-critical search-and-rescue and may be extended to a broad class of cooperative and non-cooperative RF sensing applications.

Abstract: Multimodal remote sensing classification benefits from complementary spectral, structural, and elevation information. Convolutional neural networks (CNNs) have limited long-range modeling, whereas Transformers incur quadratic computational and memory costs. Although Mamba enables linear-complexity sequence modeling, multimodal Mamba methods usually fuse modalities before selective scanning, limiting the direct use of auxiliary information in parameter generation. We propose Cross-Modal Selective Scan (CM-S6), which uses lightweight projections to condition selective state-space parameters Δ, B, and C on auxiliary information. These parameters regulate state retention, information writing, and state readout, respectively. With input injection, output calibration, and five bounded learnable coefficients, CM-S6 forms an input-parameter-output fusion mechanism without increasing backbone width or depth. Across ten independent runs, CM-S6 achieves OA values of 96.16±0.18%, 91.91±0.49%, and 82.18±0.46% on Houston 2013, Augsburg, and MUUFL Gulfport, respectively, using fewer than 1 M parameters-1.8-2.9% of the strongest compared baselines. Paired tests detected no significant OA difference from the strongest baseline on any dataset, while CM-S6 showed significant OA gains over PICNet and MCAMamba/MTMixer on the respective datasets. CM-S6(L) retains parameter-level modulation, reduces model size to approximately 0.19 M parameters, and increases throughput by 3.86-5.82×, with a dataset-dependent accuracy trade-off. CM-S6 offers an alternative for multimodal fusion under resource constraints.

Abstract: Perennial crops are a key component of sustainable agriculture due to their higher carbon storage compared to annual systems. Improving their mapping is essential to support effective public policies. This study proposes an integrated approach combining field observations, waveform LiDAR data from Global Ecosystem Dynamics Investigation (GEDI), and hyperspectral imagery from EnMAP to enhance perennial crop classification.Field data were used to develop an interpretation key for the analyzed classes. GEDI Level 1B waveforms were normalized to extract metrics such as the number of peaks, while structural attributes, including canopy height (RH95), were derived from GEDI Level 2A. From EnMAP imagery, spectral reflectance was obtained and processed using mean values, first- and second-order derivatives, and continuum removal. Differences among classes were evaluated using Tukey’s test and principal component analysis (PCA). Results show that waveform structure and canopy height significantly distinguish secondary vegetation from perennial crops. Additionally, spectral regions in the visible and shortwave infrared present significant differences between these classes. These findings demonstrate that integrating structural and spectral variables improves the discrimination of perennial crops and helps overcome challenges in separating them from secondary vegetation.

Abstract: Frequent droughts have severely undermined agricultural productivity. Using the standardized precipitation evapotranspiration index (SPEI), we identified that drought events predominantly occurred from July to September during 2000–2021 in Jianghan Plain, southern China. Over 50% of the total paddy field area experienced drought conditions in the actual rice growth periods in eight years (2001, 2003, 2006, 2009, 2011, 2018, 2019, and 2021), with the most severe and prolonged event in 2019 lasting nearly 4 months. To assess the ecological impacts, we employed solar-induced chlorophyll fluorescence (SIF) to examine the differential drought responses of three rice cropping systems: common single cropping rice (CSCR), double cropping rice (DCR), and integrated farming of rice and aquaculture animals (IFRA). The main findings are: (1) Under drought years, average SIF reductions were -3.33% for CSCR, -5.09% for DCR, and -2.17% for IFRA,  indicating DCR was the most sensitive and the IFRA the least. (2) The sensitivity of rice SIF to water stress declined from the vegetative stage to the ripening stage. (3) Pre-transplanting precipitation anomalies affected the subsequent drought resilience. These findings provide insights into the mechanistic responses of rice photosynthetic capacity to water deficits and offer a framework for evaluating drought impacts on paddy ecosystems.

Abstract: Accurate land-cover mapping of urban botanical gardens is essential for biodiversity assessment, ecosystem monitoring, and sustainable landscape management. However, their high structural complexity creates significant challenges for remote sensing-based classification. This study evaluates the influence of spatial resolution on classification performance by comparing 3 m PlanetScope satellite imagery and centimeter-scale unmanned aerial vehicle (UAV) data within a 28 ha heterogeneous area of the Yerevan Botanical Garden (YBG) in Yerevan, Armenia. Three pixel-based machine learning (ML) algorithms, Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosting (GB), were compared with the spatial-context-based U-Net deep learning (DL) semantic segmentation model. The results revealed a strong relationship between sensor resolution and classifier performance. For satellite imagery, U-Net achieved the highest statistical accuracy, with an Overall Accuracy (OA) of 78.56% and a Kappa coefficient (κ) of 0.726, demonstrating the benefit of contextual feature extraction for reducing pixel-level noise. However, spatial evaluation indicated that RF generated land-cover patterns with better preservation of local landscape structures and class boundaries. In contrast, UAV-based classification showed superior performance for pixel-based ML methods, with GB achieving the highest accuracy (OA = 85.9%, κ = 0.839), followed by SVM (OA = 84.7%, κ = 0.825) and RF (OA = 84.3%, κ = 0.821). Area comparison revealed that medium-resolution satellite imagery overestimated continuous tree canopy coverage (30.17% vs. 16.53% from UAV mapping) and underestimated fragmented vegetation classes due to mixed-pixel effects. The findings highlight that classifier selection should be adapted to sensor characteristics and landscape complexity. For satellite imagery, U-Net achieved the highest overall accuracy while RF preserved landscape structure and class boundaries more effectively; for UAV imagery, GB achieved the best overall performance, confirming that spectral-based pixel classifiers perform strongly at fine spatial resolutions. This is probably due to reduced mixed-pixel effects rather than enriched spatial information, as the latter was not leveraged in the classification process used.

Abstract: The suspended matter concentration (TSM) and Secchi disk depth (SDD) are key parameters describing seawater quality, directly reflecting the turbidity of seawater and the degree of light absorption and scattering by seawater. In this study, based on GOCI data, three different atmospheric correction (AC) algorithms were employed to retrieve the TSM and SDD of the Yellow Sea and Bohai Sea, and conducted with in-situ data. Leveraging the high observation frequency of GOCI, differences in the hourly variation characteristics of TSM and SDD under different AC algorithms were further analyzed. The results show that at the same pixel, the values of TSM and SDD retrieved using different atmospheric correction algorithms exhibit non-negligible differences. In some sea areas, there are obvious anomalies in the coupled response of TSM and SDD obtained using the same atmospheric correction algorithm. Most critically, the hourly trends of TSM and SDD obtained by different AC algorithms may diverge or even be completely opposite, especially in highly turbidities waters. This suggests that AC algorithmic selection is not just a preprocessing detail, but a factor that can radically alter the interpretation of short-term biogeochemical dynamics. The findings suggest an important methodological warning for high-frequency time-series research in ocean color remote sensing.

Abstract: Extreme heat is the deadliest weather-related hazard globally, yet the role of agricultural land management, specifically post-harvest transitions from dense vegetative canopies to exposed bare soil, in amplifying urban heat exposure remains critically unquantified in disaster risk management (DRM) frameworks. This study investigates Land Surface Temperature (LST) dynamics before and after the annual harvest season across seven consecutive years (2020–2026) at the interface between an agricultural plain and the city of Lleida, Catalonia, Spain. Framing the analysis explicitly within the Sendai Framework for Disaster Risk Reduction 2015–2030 , we utilize open-access Landsat 8/9 Collection 2 Level 2 surface temperature products alongside Sentinel-2 Level-2A surface reflectance imagery to derive LST, Normalized Difference Vegetation Index (NDVI), Normalized Difference Moisture Index (NDMI), and Short-Wave Infrared (SWIR) reflectance at a co-registered 20-meter spatial resolution. Across all seven years, post-harvest mean city LST consistently exceeded pre-harvest baselines by 16.4 °C to 38.1 °C, surpassing local emergency thermal risk thresholds in every year evaluated. Strong, spatially co-registered pixel-wise correlations between LST and NDMI (mean r≈-0.74) and LST and NDVI (mean r ~-0.70), along with positive correlations with SWIR (mean r≈+0.74), confirm that rapid loss of canopy moisture and soil evapotranspiration drives severe microclimatic heating beyond natural background summer warming. Since agricultural harvest dates follow highly predictable calendar windows, this phenomenon represents a unique class of predictable spatial hazards. We demonstrate how integrating satellite-tracked agricultural phenology into urban early warning systems and spatial planning directly advances Sendai Framework Priorities 1 and 4.

Abstract: Marine cage aquaculture contributes to nutrient enrichment of the surrounding waters, with ammonium (NH_4⁺) being the predominant ionized form of inorganic nitrogen (N) released. Elevated NH_4⁺ concentrations in marine waters may promote excessive phytoplankton growth, which may subsequently affect the ecological status of the surrounding ecosystem under certain environmental conditions. This study aims to investigate the potential of Sentinel-2 MSI imagery for estimating NH_4⁺ in an oligotrophic semi-enclosed cove where marine cage aquaculture operates. Firstly, the relationship between in situ NH_4⁺ and chlorophyll a (chl a) measurements was evaluated, both on a seasonal basis and across all seasons, to determine whether chl a could serve as a proxy for NH_4⁺ estimation. Given the weak correlation observed, the feasibility of direct NH_4⁺ retrieval was further evaluated. A band-selection procedure was applied to retain Sentinel-2 MSI spectral bands with the highest correlation with NH_4⁺ while minimizing multicollinearity arises from spectral “overlap” among bands. To evaluate the effectiveness of the proposed band-selection procedure in improving the generalization ability of three machine learning (ML) models, the complete set of Sentinel-2 MSI spectral bands at 10 and 20 m spatial resolutions was also used. The obtained RMSE and MAE values (0.29 and 0.22 μM, respectively) for the Random Forest–Bayesian Optimization with Tree-structured Parzen Estimator (RF-BO-TPE) and Support Vector Regression–Bayesian Optimization with Tree-structured Parzen Estimator (SVR-BO-TPE) models in the testing dataset, using the selected spectral bands (B03, B04, B05 and B08), indicated reasonable agreement with the actual NH₄⁺ concentrations (Range: 0.00-1.46 μΜ, SD = ±0.20 μM). Furthermore, the observed robustness against overfitting (ΔR² = 0.07, OR_RMSE = 1.10–1.11, and OR_MAE = 1.01–1.02) suggests that the developed models could provide useful NH₄⁺ estimates. These findings highlight the potential of Sentinel-2 MSI for NH₄⁺ estimation in oligotrophic coastal waters affected by cage aquaculture.

Abstract: This study evaluated spatial patterns and trends in surface urban heat island intensities (SUHIIs) across three major cities in Alabama, USA: Huntsville, Birmingham, and Mobile. Spatial expansion of developed areas and magnitudes and trends of SUHIIs were mapped and quantified using land-use/land-cover (LULC) data from the National Land Cover Database (NLCD) and land-surface temperature (LST) products from the Moderate Resolution Imaging Spectroradiometer (MODIS), respectively. Our findings reveal urban expansion over all three study areas, but at varying rates and patterns, with the highest rates observed over the Huntsville city area. The ISODATA clustering approach using LST time series mapped surface urban heat islands (SUHIs) as distinguished clusters of significantly (p=0.01) warmer surface temperatures compared to their surrounding non-urban areas. The SUHI clusters also resembled spatial distributions of NLCD-developed LULCs, with the highest resemblance between NLCD-developed LULC clusters and SUHI clusters mapped using summer daytime time series. Our SUHIIs estimated using the derived SUHI clusters correspond well with the values reported in SUHI literature. Yearly seasonal average SUHIIs varied significantly (p = 0.05) among the three study areas and during seasonal and diurnal cycles. However, SUHIIs during summer daytime were consistently larger across all sites and throughout the study period. During both seasons (winter and summer) and over all sites, the highest SUHIIs were reported over the SUHI-3 clusters that resembled high-intensity developed areas. These findings indicate novel insights suggesting our ISODATA clustering approach using time series of satellite-derived LST products is a promising approach for local-scale mapping and characterization of SUHIs. Our findings also provide detailed insights for improved understanding of the spatial distributions and temporal variations of SUHI developments over similar, less studied, but rapidly developing mid-size cities.

Abstract: Tropical cyclone (TC) inner-core maximum wave height (MWH) measurements are virtually absent, forcing reliance on reanalyses such as ERA5. Using a GPM-based dynamic MWH product at 5 km resolution, 165 global TCs (maximum wind ≥ 25 m s⁻¹) are analyzed to evaluate ERA5 MWH in both magnitude and location. GPM reveals an inner-core MWH/significant wave height (SWH) ratio of ~2.0, higher than ERA5’s ~1.9, and—critically—exhibits a steep radial decline that ERA5 entirely misses. The absolute underestimation Δ = GPM_hmax − ERA5_hmax reaches 5.4 m for super typhoons (median 0.44 m). ERA5 also misplaces the MWH point: the GPM MWH point lies 1.3 km inside the radius of maximum wind (median), whereas ERA5 places it 12.5 km outward — a 13.8 km systematic outward bias (p < 0.0001). ERA5 shows an artificially narrow azimuthal spread (circular variance 0.499 vs. GPM’s 0.647, bootstrap p = 0.0005) and a flatter radial gradient. Physical consistency (steeper gradient, rain-rate collocation) supports GPM MWH realism. The bias magnitude is comparable to ERA5’s effective resolution (~28 km), indicating a resolution-related origin.

Abstract: Remote sensing research increasingly depends on heterogeneous satellite, UAV, hyperspectral, multispectral, SAR, and environmental monitoring data to support land, urban, hydrological, and environmental applications. However, these data are often affected by sensor differences, spatial and temporal heterogeneity, missing observations, irregular sampling, noise, and non-stationary environmental processes. This systematic review was conducted within the context of the Romanian Hub for Artificial Intelligence (HRIA) project, which supports the development of strategic artificial intelligence technologies. The review synthesizes current research on adaptive neural networks for remote sensing and Earth observation, with particular attention to Liquid Neural Networks and related continuous-time neural models. A systematic search was conducted across IEEE Xplore, Scopus, Web of Science, ScienceDirect, SpringerLink, Wiley Online Library, Google Scholar, and reference lists. After duplicate removal, screening, and full-text assessment, 61 studies published between 2018 and 2026 were included in the qualitative synthesis. The findings show that adaptive neural networks have gained increasing attention after 2022 and are mainly applied to image-centered remote sensing tasks, including classification, mapping, object detection, segmentation, enhancement, and change detection. Most studies adapt established deep learning architectures through multi-scale processing, adaptive feature fusion, attention mechanisms, graph relationships, or task-specific refinement. Continuous-time models are used less frequently but are relevant for irregular observations and dynamic environmental processes. Liquid Neural Networks remain emerging, yet show promise for temporal, noisy, multimodal, and nonlinear remote sensing applications.

Abstract: Glacier collapse hazard events have become increasingly frequent on the Tibetan Plateau under climate warming. However, the spatiotemporal evolution of multi-source remote sensing-based precursory signals before glacier collapse remains poorly understood. In this study, multi-source remote sensing datasets acquired during 1990–2016 were used to systematically investigate the pre-collapse evolution of Aru 53 and Aru 50 Glacier in terms of glacier geometry, surface elevation, glacier surface velocity (GSV), surface albedo, glacier surface temperature (GST), and surface synthetic aperture radar (SAR) backscatter coefficient. The results show that both glaciers experienced continuous terminus retreat, accompanied by upstream surface lowering and downstream surface thickening within the collapse zone. GSV increased markedly. Prior to collapse, extensive crevasses developed in the central and frontal parts of Aru 53, accompanied by a progressive increase in SAR backscatter coefficient. The enhanced SAR signals spatially coincided with regions experiencing pronounced surface structural changes. In contrast, decreases in surface albedo and increases in GST did not exhibit distinctive pre-collapse signatures for either glacier. Overall, the characteristic pattern of upstream thinning coupled with downstream thickening, anomalous acceleration of GSV, and rapid crevasse development are identified as the key remote sensing precursors of glacier instability.

Abstract: Selective logging often marks the beginning of forest degradation, but remote sensing alert systems typically detect it long after it starts. This type of disturbance produces a subtle canopy signal not easily recognised in the early stages, causing alerts in some areas of the Brazilian Legal Amazon to lag behind the first visible signs by several months. This study explores whether a geospatial foundation model can help reduce this time gap. For each documented disturbance site, we compile Harmonized Landsat and Sentinel-2 (HLS) image time series spanning October 2024 to May 2026 into a datacube and use it to pretrain, without labels, a compact spatiotemporal masked autoencoder (ST-MAE). This model learns to embed the dynamics of forest degradation into latent-space representations, the so-called "embeddings". Following this, a lightweight downstream model — trained with few labelled samples — interprets ST-MAE's embeddings to evaluate every new HLS observation for recent logging activity. To effectively translate the evaluation scores into production-ready alerts, we apply a simple change-detection approach: an edge filter for sudden logging-level changes, paired with a persistence filter for sustained detections. We conducted a leave-one-out cross-validation assessment across 20 well-documented logging sites: yielding no false alarms prior to the first disturbance evidence, the change detector confirmed the onset of the 20 logging activities with a median delay of 15 days after the first post-disturbance image. The downstream model's scores remain low during the pre-disturbance period but increase once logging begins, confirming that it responds to logging-related events rather than to other landscape dynamics or calendar artefacts. This label-free pretraining enables effective performance with minimal annotations, providing a practical solution to reduce the temporal gap in monitoring forest degradation.

Abstract: Reconstructing dense and temporally continuous Landsat time series remains challenging due to low temporal revisit, sparse and irregular clear-sky observations due to clouds and shadows, and sensor-specific artifacts such as the Landsat 7 Scan Line Corrector (SLC) failure. These limitations hinder long-term environmental monitoring and restrict the utility of Landsat’s five-decade archive for applications that benefit from temporally consistent time series. Recent deep learning approaches have advanced Landsat time series modeling; however, their performance degrades under very low temporal densities, and Transformer-based reconstruction methods in particular have primarily relied on temporal dependencies, with comparatively limited integration of spatial context relative to CNN-based and spatiotemporal fusion approaches. To address these challenges, this study developed a spatiotemporal deep learning framework, re-ferred to as a 3D Transformer, that extends the 1D Transformer into a hybrid CNN–Transformer architecture incorporating spectral, spatial, and temporal dependencies. The 3D architecture combines a ResNet-based convolutional encoder to extract spa-tial-spectral features with a Transformer to model temporal dynamics from the encoded sequences. This enables the model to leverage spatial context even when the central pixel is missing, while a quality-aware masking strategy ensures that only valid or partially valid patches contribute to reconstruction. Two complementary training strategies were employed: (1) dense reference time series generated from combined Landsat–MODIS observations, and (2) high-resolution Harmonized Landsat–Sentinel-2 (HLS) observa-tions. More than 95,000 augmented samples were used to train the models through a 5-year moving-window scheme. Results from 1,500 independent test time series and 319 hold-out Landsat images demonstrate that the 3D transformer consistently outperforms the 1D model across all spectral bands, seasons, land-cover types, and observation den-sities. Average RMSE reductions of ~10% were observed across spectral bands in the GPR-based evaluation, with the largest gains occurring during spring and fall when the time series changed the most rapidly. Image-based evaluations further show average improvements of 29%, 4%, 20%, and 6% across spectral bands in spring, summer, fall, and winter, respectively. The 3D model also showed improved performance under SLC-off conditions, reducing the effect of SLC-off data gaps to a greater degree compared to the 1D model. Spatial error analyses confirm that the 3D model reduced reconstruction er-rors across heterogeneous landscapes, particularly in cropland regions. These findings highlight the value of integrating spatial context into Transformer architectures and demonstrate the potential of spatiotemporal deep learning for generating continuous, analysis-ready Landsat time series spanning multiple decades for a wide range of remote sensing applications.

Abstract:

Accurate retrieval of crop structural and physiological traits from remote sensing data remains challenging due to limited field observations and poor cross-platform generalization of data-driven models. This study develops a physics-informed transfer learning framework to quantify the contributions of improving simulated data fidelity and increasing model complexity to retrieving winter wheat leaf area index (LAI) and canopy chlorophyll content (CCC) from hyperspectral observations. Two PROSAIL-D datasets with default and physically optimized leaf angle distributions were generated to represent different levels of simulation fidelity. Four dual-branch deep learning architectures (CNN, CNN–SE, CNN–Transformer, and CNN–SE–Transformer) integrating spectral bands and vegetation indices were pretrained on simulated datasets and transferred to real observations using progressive fine-tuning. Model performance was assessed using ground-based and unmanned aerial vehicle (UAV) hyperspectral datasets, and SHapley Additive exPlanations (SHAP) analysis was applied to interpret feature contributions. Results demonstrated that transfer learning substantially improved cross-domain generalization, while enhancing simulation fidelity provided greater performance gains than increasing network complexity. The CNN–Transformer model pretrained on physically optimized simulations achieved the highest accuracy and robustness for both LAI and CCC retrieval. At ground and UAV scales, it achieved LAI estimation accuracies of R² = 0.55 (RMSE = 0.63) and R² = 0.53 (RMSE = 0.62), respectively. For CCC estimation, the model obtained R² = 0.59 at both scales, with RMSE values of 36.12 μg cm⁻² and 37.56 μg cm⁻² for ground and UAV observations, respectively. SHAP analysis indicated that physically optimized simulations shifted model attention toward physiologically relevant vegetation indices, whereas default simulations induced stronger dependence on unstable visible wavelengths. Physically informed simulation design combined with transfer learning effectively reduces simulation to reality discrepancies, whereas increasing deep model complexity alone provides limited improvement. The proposed framework offers an accurate, interpretable, and scalable solution for cross-platform crop trait retrieval from hyperspectral observations.

Abstract: Accurate crop mapping provides spatially explicit information for agricultural planning, monitoring and resource allocation, while supporting agribusiness decisions related to supply chains and risk management. However, mapping crops in smallholder landscapes remains difficult because fields are often fragmented, small, and irregularly shaped, with varying planting dates. This study evaluated Sentinel-2 Normalised Difference Vegetation Index (NDVI) time series and Dynamic Time Warping (DTW) for maize identification and multi-crop field-boundary classification in a smallholder farming area. A maize reference trajectory was developed from known maize fields and tested using an independent set of maize fields. The maize reference fields showed low DTW distances (0.106 - 0.281) to the reference trajectory, while independent validation fields produced comparable distances (0.074 - 0.325), indicating that maize followed a recognisable NDVI trajectory despite variable planting dates. DTW distances (0.151 to 0.565) of candidate fields suggested that some fields had maize-like phenological behaviour while others were less similar to the maize reference. Days After Planting (DAP) alignment produced a clear maize curve, with NDVI increasing after planting, peaking between 90 and 120 DAP, and later declining. The approach was then extended to multi-crop classification using maize, soybean, potato and tea reference trajectories. The multi-reference DTW classifier achieved an overall validation accuracy of 80.0%, with maize and tea classified most reliably. When applied to 128 unlabelled fields, most were predicted as maize, although DTW confidence varied. The findings show that Sentinel-2 NDVI time series and DTW provide an interpretable and data-efficient approach for maize detection and field-level crop classification in smallholder systems.

Abstract: Landslides triggered by heavy rainfall have become increasingly frequent and severe, creating a need for the rapid and accurate detection of damaged areas for post-disaster response and recovery planning. This study developed a ChangeFormer-based landslide damage detection model using single-channel differenced Normalized Difference Vegetation Index (dNDVI) imagery derived from pre- and post-event Sentinel-2 data. Landslide reference data were used to construct a patch-based training dataset, and model generalization was evaluated in Sancheong-gun and Hapcheon-gun, Gyeongsangnam-do, Republic of Korea, where landslide damage was reported following heavy rainfall in 2025. As available reference data differed between these regions, region-specific validation strategies were applied. In Sancheong-gun, polygon and point reference data were used for quantitative validation. The model achieved a producer’s accuracy of 100% in the polygon-based validation, detecting all 12 reference damaged sites, and point-based validation showed increasing accuracy with buffer radius, reaching 87.9% accuracy at 80–100 m. In Hapcheon-gun, where quantitative reference data were unavailable, qualitative assessment based on drone imagery indicated that the detected areas were generally consistent with locations believed to represent actual landslide damage. These results suggest that the proposed framework is effective for the post-disaster spatial assessment of landslide-damaged areas.

Abstract: Land subsidence amplifies the impacts of sea-level rise (SLR) in low-lying deltas, yet most Coastal Vulnerability Index (CVI) assessments omit it or rely on global estimates, drastically understating deltaic vulnerability. This study presents the first CVI assessment along Ghana’s coast to integrate validated, spatially resolved land subsidence derived from Persistent Scatterer Interferometric Synthetic Aperture Radar (PS-InSAR). Nine geological, geomorphological, hydrodynamics and anthropogenic variables were quantified across 72 contiguous grid cells spanning ~150 km of the Volta Delta coastline and ranked on a 1–5 vulnerability scale. Two composite indices were computed using an unweighted square root of product mean formulation, one incorporating subsidence (CVI_(+sub)) and one excluding it (CVI_(-sub)), classified against a common quantile distribution for direct comparability. By incorporating measured subsidence, ranging from -2.44 to -4.23 mm/yr (grid cell means), 75% of grid cells indicated High to Very High vulnerability under CVI_(+sub), compared with 28% under CVI_(-sub); a Wilcoxon signed-rank test confirmed the increase as significant (p < .001). Vulnerability peaked along the Keta and Songor lagoonal margins. Omitting measured subsidence understates deltaic vulnerability, and this approach offers a transferable method for data-sparse deltas.

Abstract: We report the installation and first two years of operation of an EM27/SUN Fourier transform spectrometer at the CNR-ISAC facility in Bologna, Italy, the first instrument of its kind operating in the Po Valley, one of the European region mostly affected by anthropogenic pollution. The instrument is integrated into the COCCON network and it retrieves total columns and column-averaged dry air mole fractions of CO₂, CH₄, CO, and H₂O using the PROFFAST processing chain as per standardized COCCON protocols. The data is filtered with the fitted solar gas shift in the O₂ absorption band for quality controls, and here we study the seasonality of the retrieved products. To assess the relationship between column-integrated and near-surface measurements, we present a case-study comparison against a co-located Cavity Ring-Down Spectroscopy analyzer for CO₂ and CH₄ over a single day, using ceilometer-derived Mixed Aerosol Layer height to interpret the differing diurnal variability between the two techniques, with CRDS showing substantially larger amplitude swings driven by boundary-layer dynamics and surface emission/sink processes. We further validate the EM27/SUN against TROPOMI satellite retrievals of XCH₄ and total-column CO over the full observing period, finding low systematic biases (-0.13% and 0.60% respectively) consistent with satellite mission requirements and comparable to TCCON stations at similar latitudes. As a case-study, we show that both instruments jointly detected an anomalous CO enhancement in July-August 2024 attributable to long-range transport of Canadian wildfires smoke, illustrating the capability of ground-based FTIR observations to identify high-altitude pollution intrusions undetectable by in-situ instrumentation alone. These results establish the EM27/SUN in Bologna as a robust reference for satellite validation and a valuable tool for characterizing atmospheric transport events in the Po Valley.