Conference Agenda

Climate change is intensifying droughts, creating an urgent need to understand these events and take necessary mitigation actions. This work focuses on the Bundelkhand region of Uttar Pradesh, India, an area that frequently experiences severe water stress and is highly susceptible to drought. We used multi-source remote sensing datasets to monitor drought conditions through established drought indices. The analysis period spans from 2000 to 2021. CHIRPS data were used to calculate SPI and RAI, while PKU GIMMS NDVI data were used to calculate VCI. ERA-5 Land was used for soil moisture data to derive SMCI. To track how drought propagates, we performed a correlation analysis between the indices representing meteorological and agricultural drought. The results show that the datasets and the methods are suitable for identifying droughts in the region. Historical drought episodes were accurately detected, and the analysis of the 2015 drought revealed its onset from June to September, which aligns with the monsoon season in Bundelkhand. The datasets and indices used provide a practical and reliable output for sparse ground-based observations for regional drought monitoring and management.

The Near-global Combined Drought Monitoring (NEC-DROMO) dataset provides a comprehensive depiction of drought conditions by integrating multiple agro-climatological variables across global land areas. Spanning 2002–2021 at a monthly time step and 0.25° spatial resolution, NEC-DROMO combines soil moisture, vegetation water content (VWC), rainfall, and temperature to capture both agricultural and meteorological drought signals. A key feature of the dataset is the use of Principal Component Analysis (PCA) to derive dynamic, month-specific weights for each variable, allowing the Combined Drought Indicator (CDI) to reflect seasonal and regional variability in drought drivers.

The dataset is built primarily on inputs from the ECoHydrological Land Reanalysis (ECHLA), which provides soil moisture, VWC, and temperature derived from passive microwave observations assimilated through a land–vegetation model. Rainfall fields are obtained from the ERA5 reanalysis, ensuring consistency across atmospheric and land-surface conditions. Validation against satellite-based indicators, ground observations, and event-based disaster datasets demonstrates NEC-DROMO’s strong capability to reproduce observed drought patterns globally.

With its multi-variable foundation and long-term coverage, NEC-DROMO serves as a valuable resource for drought monitoring, climate analysis, food-security assessment, and agricultural risk management. It supports detailed historical analyses and offers an integrated perspective for users seeking reliable, spatially consistent drought information.

Tropical terrestrial vegetation is critical to the global carbon cycle but faces escalating drought threats. Traditional assessments using fixed climate thresholds often ignore actual physiological responses and non-moisture disturbances. To address this, we developed a novel framework that isolates the true physiological impacts of atmospheric and soil moisture (SM) deficits to identify growing-season vegetation droughts (1982–2019). Results reveal pantropical increases in drought intensity, with tropical forests experiencing significantly sharper intensifications than other biomes. Regionally, African forests exhibit the most severe expansions in drought intensity and area. Interpretable machine learning attributes this intensifying drought predominantly to declining SM (NDVI: 52.1%; LAI: 53%). Finally, while reliable historical reconstruction is vital for future projections, CMIP6 models fail to reproduce these observed trends. These findings highlight mounting drought pressures on tropical forests and underscore the critical need for improved climate models to inform mitigation strategies.

Extreme heat events (EHEs) pose growing risks to densely populated subtropical cities such as Hong Kong, yet there remains a need for lightweight, interpretable tools that can provide multi-day forecasts based on readily available observations. This study develops a multi-source data driven framework that integrates aerosol optical depth (AOD), land surface temperature (LST), precipitable water (PW), and precipitation (Precip), together with ARIMA-based anomaly features, to predict EHEs over Hong Kong. Using a seven-day sliding window, independent XGBoost classifiers are trained to forecast daily EHE occurrence probabilities for the next 1–5 days over ten climate years (March 2015–February 2025). A lead-specific threshold optimization on a validation subset is applied to maximize F1-score. Test results show that AUC values for Lead 1–Lead 5 remain between 0.935 and 0.883, with F1-scores between 0.738 and 0.639, indicating robust predictability up to five days in advance. A process-scale duration inference method based on the leading continuous segment of the predicted sequence achieves 67.08% exact-match accuracy, 77.69% accuracy within ±1 day, and a mean absolute error of 0.75 days. The proposed framework is computationally efficient and operationally relevant, offering practical support for urban heat early warning and risk management.

Anthropogenic climate change and land use transformations are interactively reshaping environmental risks. This study investigates the critical feedback between Land Use/Land Cover (LULC) change and shifts in Köppen-Geiger (KG) climate zones in Taiwan from 2001–2020, and their combined impact on disaster hotspots.

Using MODIS and CHIRPS data alongside a comprehensive disaster inventory, we quantified the spatial co-occurrence of LULC change and climate zone transitions. Our preliminary results reveal a significant climatic shift, with over 10,500 km² transitioning from tropical monsoon (Am) to a drier tropical savanna (Aw) climate, alongside substantial wetland loss and urban expansion. We hypothesize that these dynamic "climate transition zones" are emerging fronts of heightened disaster risk.

Our analysis tests whether areas undergoing active climate reclassification concentrate a disproportionate share of historical landslides and floods. The expected outcome is a novel, dynamic risk assessment framework that moves beyond static models. By identifying these emerging hotspots, this research provides a critical tool for proactive land-use planning and climate-resilient disaster risk reduction, with methodologies applicable to other complex, hazard-prone regions.

Climate change is amplifying the occurrence and intensity of Extreme Event Induced Disasters (EEID), such as floods and wildfires, which increasingly threaten societies and ecosystems. Fast and accurate monitoring tools are therefore essential for damage assessment and emergency response. Remotely sensed data, particularly from the Copernicus Sentinel-2 mission, provide valuable multispectral information for large-scale environmental monitoring, but their manual analysis remains time-consuming.

Recent advances in Deep Learning (DL) have enhanced classification, segmentation, and change detection of geospatial data. New multimodal Prompt-Based (PB) architectures integrate image and text inputs via Text Encoders (TEs), enabling zero-shot detection of previously unseen objects. These models promise flexible, prompt-driven analysis but often underperform compared to Object-Specific (OS) models optimized for particular tasks.

In Earth Observation (EO), foundation models such as Prithvi-EO and TerraFM mark a major step forward, offering generalized pre-training across vast multi-sensor datasets to support downstream OS tasks with limited data.

While DL traditionally requires coding expertise, commercial GIS platforms now integrate DL tools accessible through Graphical User Interfaces (GUIs), allowing inference and limited fine-tuning of pre-trained models. This democratizes DL access for GIS users but shifts expertise toward model evaluation and interpretability.

This study systematically compares PB and OS models executed through both GUI-based and Python environments using Sentinel-2 flood and wildfire imagery, assessing accuracy, flexibility, and processing efficiency to evaluate the balance between accessibility and performance in the democratization of DL for EEID monitoring.

Multi-modal Satellite Image Time Series (SITS) analysis faces significant computational challenges for live land monitoring applications. While Transformer architectures excel at capturing temporal dependencies and fusing multi-modal data, their quadratic computational complexity and the need to reprocess entire sequences for each new acquisition limit their deployment for regular, large-area monitoring. This paper studies various dual-form attention mechanisms for efficient multi-modal SITS analysis, that enable parallel training while supporting recurrent inference for incremental processing.

We compare linear attention and retention mechanisms within a multi-modal spectro-temporal encoder. To address SITS-specific challenges of temporal irregularity and unalignment, we develop temporal adaptations of dual-form mechanisms that compute token distances based on actual acquisition dates rather than sequence indices.

Our approach is evaluated on two tasks using Sentinel-1 and Sentinel-2 data: multi-modal SITS forecasting as a proxy task, and real-world solar panel construction monitoring. Experimental results demonstrate that dual-form mechanisms achieve performance comparable to standard Transformers while enabling efficient recurrent inference. The multi-modal framework consistently outperforms mono-modal approaches across both tasks, demonstrating the effectiveness of dual mechanisms for sensor fusion. The results presented in this work open new opportunities for operational land monitoring systems requiring regular updates over large geographic areas.

Sensor harmonization is required to produce consistent Landsat imagery for long-term change detection. This study investigated the effect of seasonality and aerosol optical thickness (AOT) on linear harmonization functions, which are frequently used to create consistent Landsat 7 ETM+ and Landsat 8 OLI time series data. We found that training harmonization functions with pixels that have low or average AOT can greatly reduce the difference between near-coincidental Landsat 7 and Landsat 8 observations, and that seasonally trained harmonization models outperform models trained on year-round data. We assessed the effect of ETM+/OLI sensor harmonization on forest type classification with a Random Forest model, and found that seasonally harmonized imagery provided more consistent classification maps than the alternatives. This study illustrates important details related to the creation of harmonized datasets and is a significant step toward creating more consistent Landsat 7 and Landsat 8 data for long-term change detection analysis.

Urban growth in Ecuador's inter-Andean valleys has accelerated the territory's transformation, driven by the expansion of road infrastructure and the occupation of environmentally fragile areas. In this context, the Ruta Viva highway has reconfigured urbanisation patterns in the parishes of Cumbaya and Tumbaco, advancing the urban frontier into agricultural areas and moderate slopes. The objective of this study is to evaluate the dynamics of urban expansion in the parishes of Cumbaya and Tumbaco during the period 2002-2032, using a multitemporal and predictive approach to project future urbanisation scenarios and generate inputs for sustainable territorial planning and land management. The methodology integrated multitemporal analysis of land use and land cover data from MapBiomas (2002-2022), predictive modelling using CA-Markov-MOLUSCE, and urban expansion analysis. The results show a 3% increase in urban coverage during the 2002-2022 period and a projected 12% growth by 2032, concentrated south of the Ruta Viva corridor and within the agricultural mosaic. Simulations show that slopes below 25° are more susceptible to urbanisation, while vegetation cover loss reaches 30% on the slopes of Ilalo Hill. This study provides a robust, replicable tool for anticipating urbanisation scenarios in Andean environments, guiding land management and environmental conservation strategies in regions of high urban pressure.

The Area-Based Approach (ABA) for modelling forest biomass with ALS data assumes perfect co-registration, but operational inventories often have spatiotemporal misalignments. This study isolates and quantifies the independent error contributions from temporal gaps and spatial co-location errors. The analysis uses a unique dataset from the Taiga Plains, Canada, featuring 163 re-measured field plots paired with repeated ALS acquisitions from the same sensor.

To assess temporal effects, we constructed scenarios with varying time-gap distributions. Symmetrical time gaps (SD 1.1 vs 2.5 years) increased RMSE by ~1 percentage point but did not add bias. In contrast, skewed distributions introduced significant systematic biases of 8.0 % (6.8 Mg ha⁻¹).

To assess spatial effects, we linked co-location uncertainty directly to plot-level neighbourhood heterogeneity. This was done by shifting the 20x20m ALS footprint over a 1m lattice and recalculating predictors. The resulting predictor variability (RMS(CV) 12.7%) was propagated through the model, implying a positional sigma of 10-15%. Monte Carlo simulations confirmed this spatial component is the dominant error source, contributing 2–4 percentage points to the ~22% baseline %RMSE.

Our findings show that while balanced temporal gaps are manageable, spatial co-location affected by the local heterogeneity is the most critical factor for robust ABA models.

Bathymetry information around islands, reefs, and shallow-water regions is critical for both navigation safety and environmental management. However, these areas often feature diverse substrate types and strong spatial heterogeneity, which makes it challenging to accurately retrieve fine-scale bathymetry from traditional medium-resolution satellite imagery. High-spatial-resolution (HSR) sensors, such as PlanetScope SuperDove (~ 3.7 m spatial resolution), offer the potential to capture more detailed spatial features, yet their relatively low signal-to-noise ratio (SNR) can lead to noisy retrievals, particularly over low-reflectance waters. To mitigate this issue, incorporating the spatial context of neighboring pixels while jointly utilizing the spectral information offered by low- and high-resolution sensors can enhance the stability and accuracy of HSR-based bathymetry retrievals. In this study, a UNet++ neural network with the spatial and channel squeeze & excitation (scSE) attention mechanisms (BathyUNet++) was employed to retrieve bathymetry from SuperDove imagery. To satisfy the patch-based input requirement of UNet++, the model was fully trained using two sources of data: clear-sky SuperDove image patches paired with Landsat-8-derived bathymetry and a limited set of ALB data. Validation results demonstrated that the model accurately retrieved bathymetry in regions independent of the training set.The proposed model and framework can be readily adapted to other HSR sensors, offering a promising approach for global HSR shallow-water bathymetry retrieval using multi-source satellite observations.

Despite its potential in forestry, agriculture, environmental monitoring, safety surveillance, and defence, real-time hyperspectral imaging (HSI) remains challenging in practice because of the high dimensionality of the data and limited onboard computational resources.

This work introduces a distributed HSI classification framework that integrates federated learning, a Self-learning Minimal Learning Machine classifier (SL-MLM), adaptive Kalman filter-based model fusion, and lightweight MQTT-based communication on Raspberry Pi edge devices and a laptop serving as the base station. Acting as local nodes, Raspberry Pis process HSI data row by row, update their models recursively, and only exchange compact model parameters and classification results with the base station. HSI data in its raw form remains local.

The findings suggest that the proposed local learning workflow can be implemented on Raspberry Pi devices, and Kalman-based fusion improves stability and consistency in comparison to individual local models. The method is feasible in scenarios where the number of labelled data points is restricted, as the SL-MLM classifier can be initialized with a mere handful of class-specific reference points. The research demonstrates a feasible, low-cost approach to distributed

embedded HSI classification and sensing.

Monitoring inland Water Quality Parameters (WQPs) is essential for managing freshwater ecosystems and assessing anthropogenic impacts (Mishra et al., 2017). Satellite remote sensing provides a cost-effective and large-scale approach for monitoring inland WQPs. However, most existing satellite sensors have limited spectral resolution, restricting their ability to capture subtle optical variations expressed by inland WQPs, and/or insufficient spatial resolution to yield valid water-only pixels in narrow rivers or nearshore zones (Ansari et al., 2025).

Recent advances in hyperspectral satellite technology have created new opportunities for inland WQP monitoring. The Wyvern Dragonette-001, launched in April 2023, provides hyperspectral imagery with a spatial resolution of 5.3 m and 23 spectral bands within the visible to near-infrared range (500–800 nm) (Ansari et al., 2025; Wyvern Dragonette, 2023). Given its novelty, the potential of such imagery for assessing WQPs in inland water remains largely unexplored. A recent review (Ansari et al., 2025) evaluating the sensor’s spectral resolution and signal-to-noise ratio for retrieving inland WQPs indicated that Dragonette-001 is suitable for estimating non-algal particles (NAP) and shows potential for chlorophyll-a mapping, although it is likely unsuitable for retrieving Colored Dissolved Organic Matter (CDOM).

This study reports on a practical test that assessed the feasibility of using Wyvern Dragonette-001 imagery to retrieve turbidity, Suspended Sediments (SS), and Dissolved Organic Carbon (DOC) in a portion of the St. Lawrence River, Québec, Canada.

Hyperspectral imaging produces vast data volumes that often exceed storage and transmission capacities on airborne and satellite platforms. This study systematically investigates the effects of lossy hyperspectral data compression on the scientific usability of the resulting data products. Using UAV-based HySpex acquisitions from the HyperThun’22 campaign, several state-of-the-art learning-based compression models were evaluated, including spectral, spatial, and spatio-spectral architectures. The analysis quantifies how compression-induced reconstruction errors propagate through the full pre-processing workflow, from raw digital numbers through radiometric calibration, geometric correction, and atmospheric correction to the final surface reflectance domain.

Results show that spectral models such as the Adaptive 1D Convolutional Autoencoder (A1D-CAE) achieve the highest fidelity, maintaining sub-degree spectral deviations and near-perfect structural similarity. In contrast, purely spatial or 3D convolutional models exhibit severe distortions that persist across all pre-processing levels. The findings demonstrate that lossy compression can be applied at the raw stage without compromising the integrity of reflectance products, provided that spectral correlations are explicitly modeled.

This work highlights the importance of selecting compression architectures consistent with sensor characteristics and pre-processing workflows and provides a quantitative foundation for future operational implementations of onboard hyperspectral compression in Earth observation missions.

Efficient nitrogen (N) management remains a major challenge for sustainable potato production, particularly on coarse-textured soils prone to nutrient leaching. This study investigates the use of Visible–Near Infrared to Short-Wave Infrared (VNIR–SWIR, 350–2500 nm) hyperspectral spectroscopy for non-destructive, in-season estimation of petiole nitrate nitrogen (PNN) under both field and laboratory conditions. Spectral data were collected using an ASD FieldSpec Pro spectroradiometer and processed through Savitzky–Golay smoothing, Standard Normal Variate normalization, and first-derivative transformation. Variable Importance in Projection (VIP) analysis was employed to identify N-sensitive wavelengths, and three predictive approaches—One-Dimensional Convolutional Neural Network (1D-CNN), Support Vector Regression (SVR), and Partial Least Squares Regression (PLSR)—were compared for their predictive accuracy. Calibration transfer using Piecewise Direct Standardization (PDS) was applied to harmonize field spectra with laboratory measurements. Results showed that the 1D-CNN achieved the highest predictive performance (R² = 0.90, RMSE = 0.22%), outperforming SVR and PLSR. PDS improved field-based predictions by reducing spectral discrepancies caused by illumination and canopy variability. The findings highlight the potential of hyperspectral spectroscopy combined with deep learning and calibration transfer techniques to provide accurate and scalable diagnostics of plant nitrogen status. This research supports the integration of proximal sensing and data-driven models for precision nutrient management in potato systems and broader agricultural applications.

This contribution presents a novel method for infrared image stripe remover that addresses the limitations of current approaches in difficulty of extracting structural information of stripes. The proposed framework integrates strip convlution layers with multi-size kernels in a dense connection to enhance stripe structural information expression in challenging conditions and provides more reliable results for practical applications. Experimental evaluations on multiple datasets demonstrate significant improvements compared to state-of-the-art methods. The method is designed to be computationally efficient and suitable for real-world deployment in fields.

This paper aims to evaluate the performance of ISODATA clustering for tree species classification using ultra-high-resolution multispectral data collected with Unmanned Aerial Vehicle. The study focuses on two sites in Żednia forest district near the city of Bialystok, northeastern Poland. The input data consist of 10-band multispectral orthomosaics with a resolution of 10 cm, acquired from an UAV platform equipped with a MicaSense RedEdge-MX dual camera and image-based Canopy Height Model. The classifications were conducted at two levels of forest detail: forest types, including two classes (broadleaf and conifer), and tree species, comprising four classes in Study Area 1 and ten species in Study Area 2. Multiple classifications were generated, testing different input parameters such as the number of clusters and various combinations of input data. For the first level of classification (forest type), overall accuracies range from 84,09% to 97,57% in Study Area 1 and from 82,31% to 92,74% in Study Area 2. At the second level of classification (tree species), overall accuracies vary from 70.73% to 91.77% in Study Area 1 and from 36,51% to 72,33% in Study Area 2. Overall, ISODATA demonstrates robust performance in classifying forest types in both study areas. However, performance in classifying tree species varies across different classes, with relatively high accuracies observed for certain species such as spruce, pine, oak, larch, and birch. The results underscore the potential of multispectral UAV data and unsupervised classification methods for accurately classifying tree species.