Although numerous satellites have been developed and launched recently, low Earth orbit satellites offer high spatial resolution but long revisit cycles, resulting in low temporal resolution, whereas geostationary satellites offer high temporal resolution but low spatial resolution. As a result, there are still limitations in reliably acquiring satellite images with high spatiotemporal resolution. To overcome these limitations, research on super-resolution fusion using various satellite images is underway. However, challenges such as data loss due to clouds, differences in revisit cycles between satellites, and sensor characteristic mismatches make it difficult to produce fused super-resolution images. Therefore, this study proposes a multi-satellite-based fusion framework that addresses these issues and reliably generates spatiotemporal super-resolution fusion images.
To this end, this study utilized various satellite images, including GK2A (high temporal frequency, 2km resolution), MODIS (high temporal frequency, 500m resolution), GOCI-II (high temporal frequency, 250m resolution), Landsat-8 (30m resolution), Sentinel-2 (10m resolution), PlanetScope (2.8m resolution), and KOMPSAT-3(3m resolution). Each satellite image underwent preprocessing steps, including geometric correction, radiometric correction, BRDF (Bidirectional Reflectance Distribution Function) correction, and normalization, to ensure spatial alignment and radiometric consistency.
Subsequently, a deep learning model based on DeepLabV3+ ResNet101 was used to generate cloud mask label data, creating mask labels for clouds and missing areas in the video. These labels were then used to apply a gap-filling technique to fill in the cloud and missing regions. Finally, a step-by-step resolution enhancement image fusion method based on spatial resolution was employed to produce a spatiotemporal super-resolution fused image.
The final super-resolution fused image will be validated using spectral data collected from a ground observation tower located in Naju, Jeollanam-do, South Korea. The multi-satellite fusion framework proposed in this study can efficiently overcome the limitations of spatiotemporal resolution by utilizing deep learning and physics-based models during various processing stages. The fusion results are expected to be applicable in various remote sensing fields, such as detecting climate change and environmental variations.
Acknowledgements: This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00515357).
How to cite: Han, D., Hahn, S., Ryu, Y., Jeong, S., Ha, J., and Yeom, J.: A Deep Learning and Physics-Based Multi-Satellite Fusion Framework for Spatiotemporal Super-Resolution Image Generation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9172, https://doi.org/10.5194/egusphere-egu26-9172, 2026.
Mountainous regions of the Kumaon Himalayas are particularly prone to landslides due to steep terrain, weak geological conditions, intense monsoon rainfall, and increasing human activity. In such environments, continuous ground-based monitoring is often difficult because of poor accessibility, dense vegetation cover, and frequent cloud conditions. Microwave remote sensing, especially satellite-based Synthetic Aperture Radar (SAR), offers a reliable, weather-independent means of monitoring surface deformation over large areas and long time periods.
This study applies an integrated multi-temporal InSAR (MT-InSAR) and deep learning framework to investigate surface deformation and landslide activity in the Nainital region, Kumaon Himalayas, India. Sentinel-1A SAR data (2020–2025) were processed on the ASF Vertex HyP3 cloud platform using the GAMMA Small Baseline Subset (SBAS) processing chain. The cloud-based workflow automates key interferometric steps, enabling efficient processing of multi-year SAR archives without the need for local high-performance computing facilities.
Time-series inversion and analysis were carried out using MintPy in a GPU-enabled OpenSARLab environment. Weighted least-squares inversion was applied to generate line-of-sight (LOS) deformation time-series and mean LOS velocity maps. In addition, ENU decomposition was performed, and the vertical (Up) component was used for subsequent analysis. The resulting five-year deformation record highlights marked spatial variability across the study area, reflecting deformation associated with slow-moving landslides, slope creep, and other forms of localized instability.
To focus on actively deforming areas, pixels were objectively selected using Otsu thresholding applied to long-term displacement metrics derived from the MT-InSAR time-series. This data-driven approach allowed stable and deforming areas to be separated without relying on subjective thresholds, capturing both known unstable slopes and newly emerging deformation zones. The selected high-deformation pixels were then used for short-term deformation forecasting using two models: a Long Short-Term Memory (LSTM) network and a Temporal Convolutional Network (TCN).
Model performance was assessed using five-fold time-series cross-validation. The TCN model showed consistently better performance, achieving an R² of ~0.95 and an F1-score of ~0.97, compared to the LSTM model (R² ~0.93, F1 ~0.92), indicating improved representation of long-range temporal dependencies and non-linear deformation behaviour.
To examine the spatial evolution of instability, K-means clustering was applied to both five-year historical and twelve-month forecasted displacement time-series, producing deformation cluster maps for each period. Areas showing transitions from lower to higher deformation classes were identified as emerging landslide hotspots. The observed deformation patterns and hotspot distribution show strong agreement with previous MT-InSAR-based landslide studies and regional landslide inventories from the Himalayan region, providing independent validation of the proposed framework for landslide hazard assessment and risk management.
How to cite: Basu, A., Dikshit, O., and Tiwari, A.: Predictive Modelling of Landslide Hotspots in Nainital using MT-InSAR and Deep Learning, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9267, https://doi.org/10.5194/egusphere-egu26-9267, 2026.
High-resolution X-ray microtomography (XMT) imaging of rock deformation experiments at micrometer scale provide valuable insights into the coupled evolution of pores, cracks, and fluid pathways (Noiriel & Renard, 2022). A critical step in XMT data processing is the removal of ring artefacts, which is attributed to malfunctioning detector components within the acquisition system (Vo et al., 2018). These artefacts appear as stripes in the raw acquired sinogram domain and concentric circles in reconstructed images. Ring artefacts can adversely affect downstream analyses such as pore and fracture segmentation, and digital volume correlation (DVC) (Mahdaviara et al., 2025). Advances in GPU computing and ML-based image processing have led the synchrotron community to explore deep learning architectures including ResUNET (Fu et al., 2023), and attention-based variants (Zhang et al., 2022) to suppress ring artefacts. Most ML-based denoisers rely on single-domain, pixel-based loss functions, such as L1, L2, or structural similarity index measure (SSIM), applied either in the sinogram or reconstructed image domain.
This study investigates a dual-domain loss function that combines loss terms in the sinogram domain with those in the corresponding Fast Fourier Transform magnitude (FFT amplitude) domain, aiming to improve generalization of trained U-Net variants. Existing artefact-free XMT images of basalt were used to simulate stripe artefacts in its raw sinogram domain. Stripe artefact generation was controlled using three parameters: pixel thickness, amplitude, and number of stripes per sinogram. A total of 5,000 paired noisy and clean sinograms were generated and split into training, validation, and test datasets. Three UNET-based architectures were evaluated: a baseline U-Net (baseUNET), a residual U-Net (ResUNET), and a residual U-Net with attention gates (AG-ResUNET). Models were trained for 100 epochs using the Adam optimiser and their performances were assessed using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and qualitative inspection of sinograms, reconstructed images, and FFT amplitude spectra. The study compares single- and dual-domain loss functions in terms of ring artefact suppression and generalization beyond the training data distribution, discusses limitations related to the dynamic range of the training data and its implications for denoising experimental XMT datasets.
References
- Fu, T., Wang, Y., Zhang, K., Zhang, J., Wang, S., Huang, W., Wang, Y., Yao, C., Zhou, C., & Qin, Y. (2023). Deep-learning-based ring artifact correction for tomographic reconstruction. Journal of Synchrotron Radiation, 30(3).
- Mahdaviara, M., Mousavi, M., Rafiei, Y., Raoof, A., & Sharifi, M. (2025). Improving numerical fluid flow simulation by ring artifact removal in micro-CT images of porous media using attention autoencoder–decoders. Transport in Porous Media, 152, 57.
- Noiriel, C., & Renard, F. (2022). Four-dimensional X-ray micro-tomography imaging of dynamic processes in geosciences. Comptes Rendus Géoscience, 354(G2), 255–280. https://doi.org/10.5802/crgeos.137
- Vo, N. T., Atwood, R. C., & Drakopoulos, M. (2018). Superior techniques for eliminating ring artifacts in X-ray micro-tomography. Optics Express, 26(22), 28396. https://doi.org/10.1364/oe.26.028396
- Zhang, J., Niu, Y., Shangguan, Z., Gong, W., & Cheng, Y. (2022). A novel denoising method for CT images based on U-net and multi-attention. Computers in Biology and Medicine, 152, 106387. https://doi.org/10.1016/j.compbiomed.2022.106387
How to cite: Dhanapal, S., Cordonnier, B., and Renard, F.: Improving Generalization of Deep Learning–Based Ring Artefact Removal in X-ray Microtomography Imaging of Geomaterials, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10211, https://doi.org/10.5194/egusphere-egu26-10211, 2026.
Abstract：Open-pit mines, typical land-surface features shaped by intensive human activities, require rapid identification of their spatial distribution for effective mineral resource supervision, ecological disturbance assessment, and land inspection. Optical remote sensing imagery, with its wide coverage, convenient acquisition, and rich spatial details and textures, provides intuitive morphological and contextual cues for open-pit mine identification and is therefore widely employed in routine monitoring and rapid assessment. Nevertheless, open-pit mines often bear strong visual similarities to quarries, bare land, construction-disturbed zones, and waste dumps. Meanwhile, slender structures (e.g., pit boundaries, bench slopes, and haul roads) tend to be smoothed out in multi-scale representations, which makes it challenging to balance global shape characterization with precise local boundary localization. To address these issues, we propose GLSNet (Global-Local State-space Network) , a feature-enhancement framework for open-pit mine detection, consisting of three synergistic modules. First, an Adaptive Scale-aware Spatial Pyramid Pooling Fast (A-SPPF) module is introduced to adaptively select effective contextual ranges, suppress confusing background interference, and improve scale robustness. Second, a Low-resolution State-Space Modeling (LS-SSM) module is designed to efficiently model long-range dependencies and scene structural relationships, enhancing discrimination between open-pit mines and visually similar land-surface units. Third, a Scale-adaptive Global–Local Fusion (SGF) module is proposed to jointly strengthen global structural constraints and local boundary details, thereby balancing holistic morphology representation and key boundary localization, and improving detection stability and cross-region generalization.We evaluate our method on the public Open Pit Mine Object Detection Dataset and compare it with Faster R-CNN, YOLOv5, YOLOv8, YOLOv10, RTMDet, RT-DETR, DEIM, and Mamba-YOLO. Results demonstrate that GLSNet achieves superior overall detection performance, with particularly notable advantages in resisting background-induced confusion under complex conditions and in recognizing small-scale targets, while maintaining high inference efficiency, thereby validating the effectiveness and synergy of the proposed modules.
Keywords：open-pit mine detection; state-space models (SSM); multi-scale features; global-local fusion.
How to cite: Chen, Z., Zheng, Y., Yao, D., and Zhao, J.: GLSNet: State-Space-Enhanced Open-Pit Mine Detection With Global-Local Information Fusion, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16478, https://doi.org/10.5194/egusphere-egu26-16478, 2026.
Machine learning has seen widespread adoption across the geosciences. In particular, deep learning methods have proven effective for producing gridded datasets of climate parameters. Convolutional neural networks are commonly used, but their performance can be limited by the availability and structure of data, especially for sparse or irregularly sampled climate observations. Graph neural networks can handle irregular spatio-temporal data, but their reliance on local interactions restricts their ability to capture large-scale climate processes.
An alternative approach is DeepKriging, originally proposed by Chen et al., which embeds the spatial domain using basis functions centered at knot points across the region of interest. By using these basis functions as input features for a neural network, DeepKriging provides an efficient and flexible representation of both spatial and temporal domains, making it suitable for irregular data and capable of capturing both large-scale and local effects. However, DeepKriging requires basis functions to be manually defined before model training, which can require extensive work from the practitioner to fine-tune the model. This also limits the model’s ability to adapt to varying spatio-temporal patterns.
Here, we propose several extensions to DeepKriging, primarily by allowing basis functions to be updated throughout model training. The resulting model dynamically adapts to diverse spatio-temporal patterns while converging on a basis function representation that is optimal for the current data. We further improve the flexibility of the spatial embedding through a mesh generated via constrained Delaunay triangulation. This approach is applied to multiple climate variables, including precipitation and wind data for Ireland, demonstrating an improved performance compared with the original DeepKriging as well as several state-of-the-art deep learning and geostatistical gridding methods.
Finally, we also show how basis function representations are particularly well suited for datasets with limited availability, such as sparsely sampled climate parameters like relative humidity or soil moisture. This flexibility can be leveraged across a range of machine learning frameworks, including transfer learning with DeepKriging models or more lightweight algorithms such as Random Forests and XGBoost.
How to cite: O'Sullivan, B. and Coonan, B.: Basis Functions Representation for Deep Learning Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16765, https://doi.org/10.5194/egusphere-egu26-16765, 2026.
Detecting wildebeest from very-high-resolution (VHR) satellite imagery enables large-area population monitoring in a single acquisition, avoiding aircraft-induced disturbance and reducing sampling bias caused by transect-based surveys. However, wildebeest appear as extremely small objects in satellite images, and direct application of classical object detectors (e.g., YOLO-style detectors) often yields poor performance. In particular, high-density aggregation areas suffer from severe missed detections due to scale mismatch and limitations in post-processing for densely packed small objects.
To address these challenges, we develop a targeted detection solution that integrates (1) an adaptive sliding-window strategy to better capture local context under varying density conditions, (2) resolution–detector adaptation to mitigate scale mismatch between object size and detector design, and (3) improved post-processing modules, including an enhanced non-maximum suppression (NMS) tailored for dense small-object scenarios. We evaluate the proposed framework using WorldView-2 and WorldView-3 imagery over the Serengeti acquired in 2022 and 2023. The overall F1-score improves from 0.727 to 0.770 in 2022 and from 0.682 to 0.756 in 2023. Notably, in high-density areas in 2022, the F1-score increases from 0.330 to 0.821, demonstrating that our approach effectively reduces missed detections in dense small-object scenarios that commonly lead to substantial omissions in traditional pipelines.
Beyond wildebeest monitoring, our results highlight a generalizable pathway for adapting classical detectors to dense small-object detection in VHR satellite imagery, where objects are tiny and crowded.
How to cite: Xu, Z., Wu, Z., Duporge, I., Lee, S., and Wang, T.: An adaptive window and resolution-aware detection framework for dense small-object mapping from very-high-resolution satellite imagery, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17458, https://doi.org/10.5194/egusphere-egu26-17458, 2026.
The urban-rural fringe (URF) has become the most dynamic area of land use transition and urban-rural factor flows during urbanization, forming a critical focus of sustainable land management. However, existing identification methods have not adequately captured the fine-scale textures and ambiguous transitional boundaries characterizing the URF. Taking Kunming City as the study region, this study developed a lightweight convolutional neural network (UF-Net) to extract spatial textures and boundary features, integrating it with eXtreme Gradient Boosting to construct a hybrid recognition framework. Multisource remote sensing and geospatial datasets were employed to delineate the URF from 2013 to 2023, and stage-specific driving mechanisms were examined using propensity score matching and binary logit models. The results showed that our framework achieved an overall accuracy of approximately 94% for both periods. Over the decade, built-up areas expanded markedly, and the spatial structure evolved from a single-core pattern characterized by fragmented peripheral development to a polycentric configuration with increasingly continuous URF zones. Chenggong and southern Guandu District emerged as major growth frontiers, while URF morphology shifted from linear to ring-shaped and cluster-type forms. Furthermore, the drivers of urban expansion transitioned from dominance by natural terrain and ecological suitability to a regime shaped primarily by human activities and transport accessibility. The proposed hybrid recognition framework, integrating deep feature extraction with ensemble-based classification, establishes a generalizable methodological path for interpreting URF evolution, providing analytical support for optimizing urban spatial structure and sustainable development strategies.
How to cite: Zhu, Y., Hu, T., Li, X., Zhang, D., and Peng, J.: A Deep Learning and Ensemble Decision Method to Identify the Urban–Rural Fringe: A Case Study in Kunming City, China, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5616, https://doi.org/10.5194/egusphere-egu26-5616, 2026.
The research presented in this study addresses the subject of large-scale soil type classification. It is based on multispectral data from the Sentinel-2 satellite along with recent advances in deep learning for tabular data analysis. Initially we created a soil dataset aligned with the World Reference Base (WRB) classification system. This dataset was created by integrating Sentinel-2 spectral bands with different indices regarding vegetation, exposed soil conditions, mineralogical composition, and moisture dynamics. The study assesses the performance of different classification models, and some hybrid approaches using ensemble learning techniques. We applied and assessed several techniques for data balancing and augmentation to address the uneven class distribution that often exists in soil datasets. The results show that combining multispectral satellite features with specific spectral indices and various learning methods offers an effective and scalable way to generate WRB-consistent soil maps from Sentinel-2 data.
Acknowledgment:
This work is supported by the project "Romanian Hub for Artificial Intelligence-HRIA", Smart Growth, Digitization and Financial Instruments Program, MySMIS no. 351416.
How to cite: Bacu, V.: Deep Learning–Based Soil Classification from Sentinel-2 Multispectral Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-8216, https://doi.org/10.5194/egusphere-egu26-8216, 2026.
High-sulfidation (HS) orebodies are typically characterised by advanced argillic alteration and strong structural control. However, their 3D delineation remains challenging due to the inherent complexity of alteration facies and the limitations of discrete drillhole observations. We present a comprehensive 3D machine-learning workflow to delineate HS orebodies at the Čukaru Peki deposit (Eastern Serbia) by integrating drill-core SWIR spectroscopy with geological and geochemical constraints.Alteration mineralogy was characterised from SWIR spectra using The Spectral Geologist (TSG), extracting diagnostic sulfate–clay signatures (e.g., alunite-group minerals) and spectral scalars (e.g., ~2.20 μm absorption depth and white-mica crystallinity). To bridge the gap between discrete samples and a continuous volume, we constructed voxel-scale attribute fields using CatBoost regression. Unlike conventional distance-based interpolation, CatBoost learns nonlinear spatial dependencies conditioned on coordinates and geological context (lithology, alteration facies, and fault proximity), enabling data-driven 3D inference across the entire modelling volume.Subsequently, a Transformer encoder was employed for voxel-wise evidence fusion on the stacked 3D attribute layers. The model captures the nonlinear mapping of "multi-evidence interaction → HS mineralisation probability" to output a probabilistic targeting volume. The model was trained on labelled exploration drilling data (604 samples) and rigorously validated against an independent in-mine dataset (2,850 samples). Performance evaluation using confusion matrices and ROC curves consistently suggests that sulfur enrichment, alteration intensity, and structural proximity jointly govern HS distribution. This approach provides a robust, interpretable basis for 3D orebody modelling and drill targeting in complex porphyry–epithermal systems.
How to cite: Wang, G., Zhang, G., Wang, Z., Yang, S., and Wang, Y.: 3D HS orebody delineation integrating CatBoost modelling and Transformer-based evidence fusion: A case study from Čukaru Peki, Serbia, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15741, https://doi.org/10.5194/egusphere-egu26-15741, 2026.
Change detection from high-resolution remote sensing images lies as a cornerstone of Earth observation applications, yet its efficacy is often compromised by two critical challenges. First, false alarms are prevalent as models misinterpret radiometric variations from temporal shifts (e.g., illumination, season) as genuine changes. Second, a non-negligible semantic gap between deep abstract features and shallow detail-rich features tends to obstruct their effective fusion, culminating in poorly delineated boundaries. To step further in addressing these issues, we propose the Frequency-Spatial Synergistic Gated Network (FSG-Net), a novel paradigm that aims to systematically disentangle semantic changes from nuisance variations. Specifically, FSG-Net first operates in the frequency domain, where a Discrepancy-Aware Wavelet Interaction Module (DAWIM) adaptively mitigates pseudo-changes by discerningly processing different frequency components. Subsequently, the refined features are enhanced in the spatial domain by a Synergistic Temporal-Spatial Attention Module (STSAM), which amplifies the saliency of genuine change regions. To finally bridge the semantic gap, a Lightweight Gated Fusion Unit (LGFU) leverages high-level semantics to selectively gate and integrate crucial details from shallow layers. Comprehensive experiments on the CDD, GZ-CD, and LEVIR-CD benchmarks validate the superiority of FSG-Net, establishing a new state-of-the-art with F1-scores of 94.16%, 89.51%, and 91.27%, respectively. The code will be made available at https://github.com/zxXie-Air/FSG-Net.
How to cite: Xie, Z., Miao, S., Jiang, Y., Zhang, Z., Yao, J., Li, X., Huang, J., and Ghamisi, P.: FSG-Net: Frequency-Spatial Synergistic Gated Network for High-Resolution Remote Sensing Change Detection, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17651, https://doi.org/10.5194/egusphere-egu26-17651, 2026.
Semantic segmentation is the foundation of a wide range of practical applications, such as urban planning, climate modeling, and environmental protection, all of which have direct socio-economic implications. However, the accelerating densification of metropolitan regions in developing countries complicates accurate mapping of fine-scale urban land uses, as three-band optical imagery often fails to capture spectral variability and the restricted capacity of the CNN-based model to establish spatial and inter-band relationships. Therefore, to address these limitations, we propose a multi-modal architecture built on a SegFormer-B2 backbone. The pipeline integrates auxiliary datasets of DEM for surface information, SWIR for capturing water absorption characteristics, and an ancillary dataset of built-up layers for enhanced urban boundary delineation, along with multi-temporal false-color composites from LISS-4 and Sentinel-2 over the Bengaluru region. The proposed framework integrates convolutional feature extraction with transformer attention to jointly learn local spectral–spatial patterns and global cross-band dependencies. Attention-guided up-sampling, a hybrid loss function, and cross-attention modules are incorporated to strengthen feature fusion across heterogeneous modalities by establishing a link between the multi-band synergy of the Auxiliary and Ancillary datasets. Empirical evaluation reveals consistent qualitative improvement and higher overall accuracy, with substantial gains for Barren land when incorporating SWIR and vegetation, and when integrating DEM. These results validate the effectiveness of the proposed framework in overcoming spectral insufficiency and spatial ambiguity, as it outperforms baseline models. Overall, the proposed approach offers a scalable and transferable solution for private developers and government agencies seeking robust, fine-resolution mapping to support a sustainable and structured urban environment.
Keywords: Urban mapping, Deep learning architecture, Spectral Feature extraction, Performance Optimization
How to cite: Kumar, V. and Haridas Aithal, B.: Contextual Aware Hybrid Deep learning framework: Assessment with Auxiliary and Ancillary Data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1347, https://doi.org/10.5194/egusphere-egu26-1347, 2026.
Growing demand for critical raw materials over the coming decades underscores the need for robust, continental-scale frameworks to identify new mineral resources. Volcanogenic massive sulphide (VMS) deposits supply Cu, Zn, Pb, Au, and Ag, and they play a crucial role in meeting Europe’s growing demand for strategic raw materials. Despite Europe’s long mining history and extensive geological datasets, mineral prospectivity assessments remain largely restricted to national boundaries, limiting the ability to evaluate mineral systems that operate across regional tectonic domains. This study develops the first integrated European-scale prospectivity model for VMS by merging harmonised public geoscience datasets within a mineral-system and machine-learning (ML) framework. This work is carried out as part of the EU-funded TERRAVISION project, which aims to enhance the entire critical raw materials value chain towards implementing sustainable mining practices.
The modelling approach builds on and extends existing regional-scale frameworks by addressing several persistent challenges in regional-scale exploration. A positive–unlabelled training strategy was used to mitigate the lack of reliable negative labels, and ML models capable of estimating uncertainty, along with multiple explainability techniques, were applied. To ensure that predictors capture meaningful geological processes, both data-driven and knowledge-based feature selection were implemented. Model explainability was evaluated through three complementary approaches: (i) built-in feature importance from the ML classifier, (ii) permutation feature importance to assess the robustness of predictor influence, and (iii) SHapley Additive exPlanations (SHAP) values to quantify local and global predictor contributions. Together, these methods provide transparent, interpretable insights into the geological and geophysical variables that indicate prospectivity patterns. The model successfully identified over 97% of known VMS deposits and occurrences with spatial patterns showing strong correlation between high-probability areas and established mineralisation. Importantly, they also highlight prospective trends in regions with limited documented exploration.
The analysis highlights several metallogenic zones that exhibit geological and geophysical signatures consistent with favourable mineral-system conditions, but where known deposits are sparse. These areas represent potential greenfield opportunities at a continental scale. The study also illustrates the value of applying the mineral system concept to regional datasets. Harmonised lithological data and spaceborne geophysical data contribute significantly to mapping crustal-scale structures and tectonic domains with a history of submarine seafloor volcanic activity, a key requirement for VMS formation. More broadly, the proposed framework is transferable to other deposit types and illustrates the strategic potential of continental-scale, process-informed and explainable ML approaches to strengthening Europe’s strategic raw-material knowledge base through consistent, process-informed regional assessments.
How to cite: Dekavalla, M., Tenorio Matanzo, S., López Del Río, M., Papathanasiou, C., and Amditis, A.: Continental-Scale Prospectivity Modelling of Volcanogenic Massive Sulphide Deposits in Europe Using a Mineral-System and Explainable Machine-Learning Framework, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1610, https://doi.org/10.5194/egusphere-egu26-1610, 2026.
The aim of this work is to propose a new technique for automatic fault tracking from 3D seismic data using the 2D Continuous Wavelet Transform (CWT) method combined with artificial intelligence. Time slices of the variance attribute, derived from the 3D seismic data and chosen by the user, are analysed using the 2D CWT with the 2D Mexican Hat as an analysing wavelet, and the maxima of the modulus of the 2D CWT are mapped for the full range of scales. The ensemble of mapped maxima for the set of time slices is filtered using a Convolutional Neural Network machine. Machine training is performed with a supervised mode using the manually tracked faults as a desired output. Application to real data shows the efficiency and robustness of the proposed method, which can greatly help seismic interpreters in avoiding manual fault tracking, a difficult and time-consuming task.
How to cite: Aliouane, L. and Ouadfeul, S.-A.: Automatic fault tracking from 3D seismic data using the 2D Continuous Wavelet Transform combined with a Convolutional Neural Network, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1908, https://doi.org/10.5194/egusphere-egu26-1908, 2026.
Remote sensing time series monitoring plays a vital role in capturing the dynamic evolution of the Earth’s surface. Recent deep learning based temporal change detection (TCD) methods have achieved remarkable progress under cloud-free optical image sequences. However, optical imagery is frequently affected by clouds and cloud shadows, resulting in pervasive and irregular data gaps that disrupt temporal continuity and sampling regularity. Consequently, current TCD approaches struggle to cope with highly dynamic surfaces and long-term or irregularly missing observations, often leading to inaccurate change detection results. To address these challenges, we propose UniRT, a unified framework that jointly performs time-series reconstruction and change detection, enabling robust monitoring from image sequences with missing observations. Specifically, a temporal-adaptive module is seamlessly embedded into a spatiotemporal learning framework while maintaining a lightweight architectural design. In addition, a time-aware decoder is introduced to better capture temporal dependencies and enhance robustness and generalization capability under irregular sampling conditions. Extensive experiments conducted on DynamicEarthNet and SpaceNet7 demonstrate that UniRT consistently outperforms state-of-the-art methods in temporal change detection, particularly in challenging scenarios characterized by severe data gaps and highly dynamic surface changes.
How to cite: Huang, H., Shao, Z., Zhong, C., Zhu, D., and Zhang, W.: UniRT: A Unified Framework for Time-Series Remote Sensing Image Reconstruction and Change Detection, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7703, https://doi.org/10.5194/egusphere-egu26-7703, 2026.
GHGSat currently operates 14 methane satellites and has plans to expand the constellation further, acquiring images globally of facilities that could emit methane for monitoring and mitigation. The constellation produces almost 1,000 observations per day in which methane is detected, geolocated and quantified. It is impractical to rely on human inspection alone at such large scale, and so automated solutions are required. However, automation must handle a highly complex classification problem—distinguishing small methane plumes from retrieval artifacts—while operating reliably at very high throughput.
Two common types of automation that help human operators are using machine learning models to detect methane and to propose segmentation masks. The first one helps reduce the total amount of data seen by human operators, and the second helps reduce the operator time spent per observation. While these types of automation are common in methane detection with coarse-resolution public satellites such as Sentinel-2 or EMIT, their applications to fine spectral and spatial resolution satellites have been more limited.
To handle the growing amount of data, we develop transformer-based detection and segmentation models, which can assist operators in processing the observations. We present the models used and performance achieved in terms of precision and recall, both for detection and segmentation, as well as discuss future improvements to further diminish operator time.
How to cite: Piedboeuf, F., Girard, M., Jervis, D., McKeever, J., and Sampson, J.: Automatic Detection and Segmentation of Methane Plumes in GHGSat Imagery , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11947, https://doi.org/10.5194/egusphere-egu26-11947, 2026.
Earth Observation data (EO) can support food-security decision making in Sub-Saharan Africa, yet operational crop-type mapping in dryland smallholder systems remains challenging under rainfed-season cloud cover, heterogeneous cropping calendars, and small, irregular fields. Model performance is further degraded by scarce and noisy labels, mixed-cropping and intercropping practices, and strong domain shift across agro-ecologies. A particularly consequential failure mode is confusion between cropped and fallow parcels, where vegetated fallows can mimic crop spectral–temporal signatures and bias cropland statistics and downstream indicators used for early warning, input targeting, and program planning. A hierarchical, stage-wise sequential transfer-learning framework built on Convolutional Recurrent Neural Networks (ConvRNNs/ConvLSTMs) to improve robustness in data-scarce smallholder landscapes is proposed. The approach learns reusable spatiotemporal representations in a coarse-to-fine curriculum and transfers them across tasks of increasing label granularity. Stage 1 produces a cropland mask by classifying cropland versus other land uses (explicitly including fallow), targeting the crop–fallow confusion that dominates errors in dryland settings. Stage 2 refines cropland into agronomic family groups (e.g., cereals, legumes, vegetables), preserving interpretable subclass structure that is often sufficient for operational monitoring when fine labels are sparse. Stage 3 resolves fine-grained crop types and mixed-dominant intercropping states. The ConvLSTM backbone is trained stage-wise: parameters learned at a coarser stage initialize the next stage, while stage-specific classification heads are optimized for the current hierarchy level. The framework is demonstrated in Senegal using Planet NICFI monthly composites (~5 m; RGB+NIR) and in situ polygon labels collected during the 2020 and 2023 rainfed seasons. Training samples are built as ~0.5 ha image patches (14×14 pixels) extracted from interior points within polygons, with sampling density scaled by polygon area to better represent large fields while maintaining coverage of small parcels. The dataset, 6,978 labeled polygons in 2020 and 5,827 in 2023 generate 18,380 and 18,378 patches for September and October 2020 (no August imagery), and 13,733/13,623/13,524 patches for August/September/October 2023. To address severe long-tail imbalance typical of regional crop inventories, offline quota-based corpus curation, online weighted sampling, and consolidate ultra-rare fine-grained labels into an “OTHER” class at Stage 3 to stabilize training, are combined. The staged framework is benchmarked against machine and deep learning baselines (Random Forest, XGBoost, CNN, and single-stage recurrent models) using macro-averaged metrics and precision–recall behavior, selecting operating points that favor higher precision for operational mapping. Results show robust cropland maps with stable accuracy under limited labels and small, irregular fields, while preserving subclass structure; cereals and legumes remain identifiable at Stage 2 (validation accuracy ≈ 0.59). At Stage 3, precision is highest for major crops, Groundnut 0.83, Millet 0.72, and Maize 0.69, and moderate for Cowpea 0.51 and Rice 0.42. Remaining errors are primarily driven by data imbalance, mixed-cropping systems, and spectral confusion, highlighting priority areas for improving long-tail supervision and intercropping representation.
How to cite: Gelaye, K. K., Sarr, M. A., Gumma, M. K., Traore, P. C. S., Bassene, C. B. E., Mbemgue, F., and Mutuku, J. M.: Stage-wise ConvLSTM Sequential Transfer Learning for Hierarchical Crop Type Mapping in Senegal’s Smallholder Systems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12575, https://doi.org/10.5194/egusphere-egu26-12575, 2026.
Stratigraphic forward modeling (SFM) is a geological modeling framework that simulates depositional processes in sedimentary systems (e.g., deep water, fluvial, delta), enabling the generation of stratigraphic architectures and reservoir property distributions. This approach is particularly effective in reproducing the realistic non-stationarity and geological heterogeneity of deep-water reservoirs, which are difficult to capture using conventional geostatistical methods such as two-point and multipoint statistics. However, SFM results are highly sensitive to small variations in initial geological input parameters, making the integration of observational data such as well logs and seismic data challenging and thereby limiting its application at an industrial scale.
In this study, we propose a novel geological model characterization framework that combines SFM with a generative artificial intelligence approach capable of achieving both high generation efficiency and robust geological realism (Fig. 1). First, an SFM-based geological model is constructed and then preprocessed to make it suitable for neural network training. A single-image diffusion model, SinFusion, is then applied to learn the geometric and property distributions of the geological model and to enable multiple equivalent generations. Furthermore, a well data integration strategy is developed using the aforementioned trained SinFusion. By infusing well data during the reverse diffusion process, the proposed method allows seamless conditioning on well data regardless of the number or spatial locations of wells. This enables immediate model updates when new well data become available in the field with no further need for costly model retraining, ensuring high flexibility.
The validity of the proposed method is evaluated through quantitative comparisons of spatial continuity, property distributions, and geometric pattern similarity with the original SFM model. The results demonstrate that the proposed method can efficiently generate multiple geological realizations and is well suited for ensemble-based uncertainty assessment. Moreover, the proposed method has the potential to expand the applicability of SFM toward industrial-scale geological modeling workflows.
Fig. 1. Overview of the SinFusion framework for geological model augmentation and well data integration based on stratigraphic forward modeling
This work was supported by Korea Gas Corporation (RD2025-0071).
How to cite: Park, E., Lee, H., Yoon, J., and Jo, H.: SinFusion-based Geological Model Augmentation and Well Data Integration, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4683, https://doi.org/10.5194/egusphere-egu26-4683, 2026.
Development of effective urban climate adaptation and mitigation strategies requires comprehensive spatial information of rooftops and buildings. The aforementioned information is important for assessing the ecosystem services provided by green and blue infrastructure in urban areas, especially for urban heat island (UHI) mitigation and energy conservation. While green roofs are widely acknowledged as a promising solution for enhancing thermal comfort in urban climate, most existing research tends to focus either on mapping current green rooftops or the potential rooftops to implement green rooftops.
This study presents a modified deep convolutional neural network-based rooftop classification framework, based on the Roofpedia framework originally created by the Urban Analytics Lab at the National University of Singapore (NUS). The model leverages high-resolution aerial imagery and incorporates slope of the rooftop to assess green roof suitability. The proposed model uses publicly available geospatial datasets from Swisstopo such as aerial images from SwissImage dataset, elevation data from the swissALTI3D digital terrain model, and building footprints from the swissTLM3D vector dataset.
When the study applies the implemented model to Bern, Switzerland, the model provides the output with labelling the rooftops into four categories: (1) existing green roofs, (2) rooftops suitable for green roof installation, (3) rooftops with solar panels, and (4) flat rooftops which are unsuitable for roof greening. To improve the accuracy and practicality of the classification, roof slope thresholds derived from terrain model were integrated alongside spectral analysis to reflect real-world installation conditions.
The model demonstrated high predictive performance with training loss of 0.0134, mean Intersection over Union (mIoU) of 0.908, and Matthews Correlation Coefficient (MCC) of 0.901. Validation metric demonstrated the robustness with validation loss of 0.0292, mIoU of 0.843, and MCC of 0.822. Comparison with the original Roofpedia framework, the modified model shows significant improvements in multi-class rooftop classification, particularly in identifying realistic opportunities for green roof expansion.
The inclusion of potential green rooftop class, combined with slope-based constraint, allows for a practical and realistic assessment of rooftop suitability for green roof installation. The modified Roofpedia model assists urban planners and decision makers with evidence-based information to support future green infrastructure deployment in Bern and other Swiss cities. Furthermore, the proposed framework is transferable and can be readily replicated in cities worldwide.
How to cite: Ko Ko, H. Y. and Rast, M.: From Rooftops to Ecosystem Services: Deep Learning–Driven Green Roof Potential Assessment in Bern, Switzerland, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4820, https://doi.org/10.5194/egusphere-egu26-4820, 2026.
Self-supervised learning has become a prominent technique for representation learning in Earth observation, largely due to the vast volumes of unlabelled data available in observation archives. However, apart from masked auto-encoding (MAE) techniques and contrastive learning, the diversity of geospatial self-supervised learning schemes in the existing literature remains limited.
This work explores the task of structure and texture disentanglement as an alternative route to self-supervised learning in the domain of Earth observation. Inspired by the Swapping Autoencoder architecture, this pipeline involves an encoder tailored to extract disentangled textural and structural information from an image and reconstruct it back to the image domain. Crucially, it includes an augmentation step that swaps texture and structure embeddings from different samples. This synthetic generation is driven by adversarial training, employing two discriminators: one responsible for assessing the likelihood of the image as a whole being real, and the other for assessing whether individual patches in the image are consistent with the source texture vector.
The texture embedding extracted from the image acts as a global vector describing the aggregated statistics of local features, while the structure embedding represents how these features are distributed in space. This preliminary work explores the potential of this approach on a domain where image labels are particularly scarce: cloud formation types in high-resolution optical imagery. The pipeline is tested on a large collection of cloudy Sentinel-2 images with the goal of identifying observational clusters of cloud formations that share similar properties, as part of the Clouds Decoded project. This work introduces a foundational architecture for this framework along with several methods of analysis that leverage the resulting deep neural network.
How to cite: Czerkawski, M., Francis, A., Borne--Pons, P., Bertozzi, B., and Campbell, J.: Disentanglement of Structure and Texture Representations as a Method of Self-Supervision for Earth Observation Data: A Case Study on Cloud Type, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-7605, https://doi.org/10.5194/egusphere-egu26-7605, 2026.
SAR-to-optical image translation (S2OIT) aims to transform the complex backscattering characteristics of Synthetic Aperture Radar (SAR) into more interpretable optical appearances. However, existing methods often suffer from over-smoothed structural details, generation of pseudo-textures caused by inconsistencies between generated textures and real optical images, and insufficient global consistency in complex scenes. To address these challenges, we propose a Physics-Aware and Frequency-Regularized Generative Adversarial Network (PAFM-GAN) for SAR-to-optical translation. Specifically, we extract local statistical and edge structural cues from SAR images and inject them into the generator as additional guidance, which enhances structural authenticity and mitigates the impact of speckle noise. To mitigate spectral misalignment and suppress high-frequency artifacts, we further transform both the generated and real optical images into the Fourier frequency domain and perform spectral distribution alignment between them. We also introduce a frequency-domain discriminator to suppress unrealistic high-frequency components, thereby effectively reducing spurious details in the synthesized results. In addition, to capture long-range dependencies under high-resolution scenarios with low computational overhead, we integrate a Mamba-based state space module (SSM) into the generator for efficient global context modeling, improving scene-level style coherence and overall consistency. Extensive experiments on the SAR2Opt, SEN1-2, and QXS-SAROPT demonstrate that PAFM-GAN consistently outperforms representative SAR-to-optical baselines across five metrics, including PSNR, SSIM, FID, LPIPS, and FSIMc. In addition, the results of multiple ablation experiments validate the effectiveness of the proposed method.
How to cite: Wang, L., Liu, Y., Zhao, J., and Lu, Z.: PAFM-GAN: Physics-Aware and Frequency-Regularized GAN for SAR-to-Optical Image Translation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10786, https://doi.org/10.5194/egusphere-egu26-10786, 2026.
GRACE and GRACE-FO satellite missions offer an observation-based perspective on terrestrial water storage anomalies (TWSA), which is valuable for assessing climate variability and human influence on large-scale water resources. In practice, however, the short duration of the GRACE record and its coarse spatial resolution make it difficult to build long, spatially consistent storage information that are needed to study basin-scale responses to hydrologic extremes such as droughts and floods. To address this limitation, we develop a multi-model hindcasting framework that reconstructs monthly GRACE TWSA from hydro-climatic predictors and evaluates both predictive performance and hydrologic plausibility using independent evidence related to extremes.
We compare four models representing three methodological families: (i) tree-based machine learning (Extreme Gradient Boosting and Extra Trees Regressor), (ii) spatio-temporal deep learning (Convolutional LSTM), and (iii) an efficient Transformer architecture for long-sequence forecasting (Informer). All models are trained and tested over the observational GRACE period (2002–2025) using strict time-block splits, and then applied to hindcast historical monthly TWSA. Predictor co-variables include precipitation, evapotranspiration, temperature, soil moisture, and land data assimilation–based storage components from GLDAS products (Noah and CLSM), enabling the models to learn storage persistence and hydro-climatic controls beyond what can be inferred from GRACE alone. The performance of the models is assessed using standard error and agreement metrics (RMSE and correlation) as well as hydrologically oriented measures (Nash–Sutcliffe Efficiency and Kling–Gupta Efficiency), with additional diagnostics targeting the representation of seasonal and inter-annual variability.
Transformer model shows the best performance on testing periods with R2 of 0.81, CC of 0.89, NSE of 0.86 and KGE of 0.88, and RMSE of 61.3 mm, while ETR showed the least R2 of 0.713, CC of 0.76, NSE of 0.68 and KGE of 0.73. To move beyond statistical agreement with GRACE, we evaluate physical reliability through multi-source validation across all major sub-basins of Türkiye using (1) groundwater table observations and (2) independent flood information, including a flood potential indicator and mapped flood extents where available. All model families capture key groundwater storage variability, while Informer generally provides the highest predictive skill and better preserves persistence and the seasonal cycle than ConvLSTM and the tabular learners. Periods of elevated reconstructed storage are consistently associated with historical record of higher flood potential, while the lower extremes of TWSA record identifying historical droughts, supporting the hydrologic realism of the hindcast products. At the same time, the tree-based models—particularly XGBoost—remain attractive due to their low computational cost and, their stronger ability to reproduce observed flood-extent spatial patterns in some basins while maintaining extreme behavior comparable to transformer architecture.
Overall, the inter-comparison highlights practical trade-offs among accuracy, robustness to extremes, and computational efficiency, and provides guidance for scalable GRACE TWSA hindcasting on cloud platforms. The validation approach is transferable and supports the use of reconstructed storage fields for drought–flood assessment and basin-scale water resources analysis.
How to cite: Karam, W., Yüksel, K., Gümüş, A., and Gündüz, O.: From conventional machine learning to Transformers: multi-model hindcasting of GRACE terrestrial water storage anomalies (TWSA) with multi-source validation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14810, https://doi.org/10.5194/egusphere-egu26-14810, 2026.
Cyanobacterial harmful algal blooms (cHABs) pose serious concerns to drinking-water safety, aquatic ecosystem health, and recreational water use across the globe. cHABs in situ data collection relies on sparse and irregular measurements and hinders the reliable learning of complex ecological processes. Although recently data-driven models have improved bloom prediction skills, the explicit reliance on the black-box nature of these models undermines scientific trust and restricts the actionable value of model outputs. In this study, we developed a mechanism-aligned deep learning framework that embeds ecological process structure directly into the learning architecture using diverse data sources. We incorporated detailed remote-sensing-based atmospheric and environmental variables, including aerosol derived deposition, nutrient wet deposition, and meteorological data. We temporally aggregated this data to reflect both short-term forcing and cumulative conditions over space and time. We evaluated our framework for 2,200 lakes across the continental United States from 2018 - 2023, with a one-week-ahead bloom prediction task. Our model is trained on 2018–2021, validated in 2022, and tested on 2023 dataset. Our preliminary results show stable generalization under diverse spatiotemporal domain shifts (R2 = 0.54, RMSE 0.59) with reduced seasonal bias relative to conventional deep learning baselines. In addition to predictive accuracy, our architecture demonstrates high explanation faithfulness (OTA = 0.83) and positive alignment with independent physical proxies (auxiliary physical proxies, R2 = 0.36). This further demonstrates that architecture learned representations remain physically consistent despite the absence of direct mechanism labels. Our work advances a new paradigm for trustworthy environmental predictions and provides a novel foundation for actionable bloom management and policy decision support in data-limited inland water systems.
How to cite: Alamdari, N., Imtiaz, S. U., and Nasr Azadani, M.: Deep Learning for Trustworthy Prediction of Cyanobacterial Blooms across CONUS Inland Waters, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3574, https://doi.org/10.5194/egusphere-egu26-3574, 2026.
Posters virtual: Wed, 6 May, 14:00–18:00 | vPoster spot 1b
EGU26-3080 | ECS | Posters virtual | VPS22
Multibranch Adaptive Feature Fusion for Hyperspectral Image ClassificationWed, 06 May, 14:03–14:06 (CEST) vPoster spot 1b
Hyperspectral image (HSI) classification often struggles with feature interference across different scales and the inherent challenges of data imbalance and sample scarcity. While deep learning models have significantly advanced the field, traditional single-branch architectures often suffer from scale-related noise, where features from different receptive fields interfere with one another. To address this, we propose the Multibranch Adaptive Feature Fusion Network (MBAFFN). Our approach utilizes three parallel branches to independently extract scale-specific features, effectively decoupling the multiscale information to prevent interference. This architecture is enhanced by two specialized modules: Global Detail Attention (GDA) for capturing broad contextual dependencies and Distance Suppression Attention (DSA) for refining local pixel-level discrimination. Furthermore, a pixel-wise adaptive fusion mechanism is introduced to dynamically weigh and integrate these features, prioritizing the most relevant scales for final classification. The performance of MBAFFN was validated on four benchmark datasets: Indian Pines (IP), Pavia University (PU), Longkou (LK), and Hanchuan (HC). Compared to current state-of-the-art methods, our model improved Overall Accuracy (OA) by 0.91%, 1.71%, 0.86%, and 3.16% on the IP, PU, LK, and HC datasets, respectively. The significant improvement on the HC and PU datasets underscores the model’s robustness in scenarios with limited training samples and complex class distributions. These results, supported by detailed ablation studies, demonstrate that adaptive fusion and scale-specific branching are effective strategies for mitigating feature interference in hyperspectral analysis.
How to cite: Li, C. and Du, B.: Multibranch Adaptive Feature Fusion for Hyperspectral Image Classification, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3080, https://doi.org/10.5194/egusphere-egu26-3080, 2026.
