GISTAM 2026 Abstracts

Area 1 - Artificial Intelligence and Learning

Full Papers

Paper Nr:	43
Title:	Urban Waste Indicator Derived from Street View Combined with UAV Imagery: Actionable Insights for Dar es Salaam
Authors:	Levi Szamek, Benjamin Herfort, Sven Lautenbach, Iddy Chazua, Benedcto Adamu, Innocent Maholi, Alexander Zipf and Steffen Knoblauch
Abstract:	Mismanaged solid waste increasingly threatens human health and worsens flooding by clogging drainage systems, which is an issue in cities like Dar es Salaam, where waste management lags behind urban growth. However, spatial data on informal dumping is scarce. This study introduces the first city-scale method for detecting outdoor solid waste using image classification with 360° Street View Imagery (SVI), supplemented by detections from unmanned aerial vehicle (UAV) imagery. Two YOLOv11m classification models were trained based on openly available SVI data: one for trash bags and one for waste piles. These were combined with UAV detections into an urban waste indicator, which visually revealed spatial patterns related to informality, population density, and socio-economic differences. The waste detections were integrated with drainage infrastructure data to evaluate flood risks linked to solid waste accumulation. The SVI models achieved high performance, with F1-scores of 0.86 and 0.98, for trash bags and waste piles respectively. Spatial overlap between SVI localisations and UAV detections was limited to just 0.5\% of the area, each modality covered different parts of the environment. UAV imagery detected waste both on streets and in off-road areas inaccessible to SVI, while SVI captured street-level waste even when UAV views were obstructed. This highlights the need for a multimodal approach. Results show that especially the Msimbazi River catchment face elevated clogging and flooding risk due to accumulated waste at critical drainage nodes. The presented multi-modal workflow bridges critical data gaps, making urban waste management more actionable and resilient to flooding risks.

Short Papers

Paper Nr:	58
Title:	LLMs as Agents to Manage Complex Geospatial Workflows in Cloud GIS Platforms
Authors:	Eva Malinverni and Marsia Sanità
Abstract:	This study explores the potential of Large Language Models (LLMs) as "agents" to simplify complex geospatial analyses on cloud platforms such as Google Earth Engine (GEE). Despite the high availability of satellite data, its effective use is often limited to experts with advanced programming skills. This work proposes a new interaction paradigm where the user, even without technical knowledge, can formulate requests in natural language. The LLM acts as an intelligent intermediary, interpreting the user's intent and auto-matically translating it into ready-to-execute Python scripts. The methodol-ogy was validated using the municipality of Pesaro as a case study, analyzing land cover changes between 2018 and 2024. By integrating Dynamic World data and NDVI (Normalized Difference Vegetation Index) calculation, the system generated transparent and reproducible workflows capable of accu-rately quantifying urban expansion and environmental transformations. The results demonstrate that this approach not only reduces the cognitive load for the user but also democratizes access to geographic information. While not replacing necessary expert supervision, LLMs prove to be a powerful tool for making GIS analysis faster, more intuitive, and accessible on a large scale.

Area 2 - Data Acquisition and Processing

Full Papers

Paper Nr:	64
Title:	Assessing Vertical Accuracy of the National LiDAR DTM (DGU, 1 m) and UAS LiDAR DTM (DJI L2, 0.1 m) in a Post-Fire Karst Mountain Environment: Biokovo Nature Park, Croatia
Authors:	Lovre Panđa, Ante Šiljeg and Fran Domazetović
Abstract:	High-resolution digital terrain models (DTMs) are critical inputs for erosion and flood-related modelling in steep terrain. Yet, uncertainty in elevation products can strongly affect derived geomorphic and hydraulic metrics. This study evaluates the absolute vertical accuracy of two DTMs in a post-fire gully system in Nature Park Biokovo, Croatia: the national airborne LiDAR-based DTM from the Croa-tian State Geodetic Administration (DGU), provided as a raster with 1 m spatial resolution, and a UAS LiDAR DTM with 0.1 m spatial resolution derived from DJI Matrice 350 RTK and Zenmuse L2 data. Absolute accuracy was assessed us-ing 27 Trimble R12i GNSS checkpoints (CPs). The L2 DTM achieved centime-tre-class performance (RMSE = 0.0222 m; ME = -0.0025 m), whereas the DGU DTM showed substantially larger errors (RMSE = 0.2113 m; ME = +0.1064 m). In addition, inter-surface discrepancies between the two DTMs were analysed us-ing a DEM of Difference (DoD = DGU − L2) within a 6.48 ha contributing catchment. Negative DoD values dominated (64.68% of the area), yielding an ar-ea-weighted mean DoD of -0.1095 m and a net discrepancy volume of -7,092.39 m³. Because the two DTMs were acquired in different years (DGU 2023; wild-fire 2024; UAS LiDAR 2025), the DoD is interpreted as a combined signal of residual measurement uncertainty, mixed-spatial-resolution representation effects, and real post-fire topographic adjustment, rather than as a pure relative-accuracy metric. The results highlight the importance of recent, high-resolution terrain data, explicit reference-system consistency, and careful co-registration when DTMs are used for post-fire hazard and process modelling.

Area 3 - Modelling, Representation and Visualisation

Full Papers

Paper Nr:	86
Title:	HybriDTVis: 3D Building-Road Reconstruction and Visualization of Hybrid Primitives of City-Scale LiDAR Point Clouds for Digital Twins
Authors:	Krishnakumar N, Archis Kulkarni and Jaya Sreevalsan-Nair
Abstract:	Semantic 3D city models used in urban digital twins are required for urban planning, infrastructure monitoring, and geospatial analysis. They require reconstruction and visualization methods for city-scale Light Detection and Ranging (LiDAR) point clouds. In such models, buildings and roads form foundational urban elements, but are reconstructed either separately or exclusively, resulting in a loss of semantic information. We propose to reconstruct these elements in a mutually informed manner, with building geometry leveraging reconstructed road networks as structural priors. This paper presents a visualization tool, HybriDTVis, supporting partial reconstruction of point clouds, that is, road and LOD1 building geometry reconstruction. HybriDTVis integrates semi-automated road reconstruction with fully automated building generation, together with a novel lightweight web-based interactive visualization system that uses hybrid primitives, namely, points and polygons. The tool implementation pipeline includes steps on road network reconstruction and building footprint pose estimation as intermediate steps. We demonstrate HybriDTVis for both airborne and mobile LiDAR point clouds to show its generalizability and robustness. Our code is available at https://github.com/GVCL/HybriDTVis.

Short Papers

Paper Nr:	70
Title:	Position Paper: Community-Run Cartography - Distributed Infrastructure for Open Source Map Services
Authors:	Jannik Breckner, Stefan Funke, Niven Ratnamaheson, Felix Weitbrecht and Mykhailo Zelia
Abstract:	Most existing map services, despite relying on open data sources such as OpenStreetMap (OSM), remain dependent on precomputed tile-based architectures. While these pipelines ensure fast delivery and mature client support, they also impose severe constraints: massive storage requirements, lengthy precomputation cycles, and limited opportunities for customization. This paper argues that open, community-driven map services should move beyond the precomputed-tile paradigm toward dynamic, data-centric rendering. We propose to replace static tiles with data structures capable of generating the required map elements for chosen regions and levels of detail on demand.

Area 4 - Our Changing Planet

Full Papers

Paper Nr:	59
Title:	Flood Risk Mapping and Assessment for Ras Al Khaimah (RAK) Using the Analytic Hierarchy Process (AHP)
Authors:	Hind Yousif Alhammadi, Rami Al-Ruzouq, Ratiranjan Jena and Nezar Atalla Hammouri
Abstract:	Flooding is a growing concern in Ras Al Khaimah (RAK), especially following the heaviest rainfall after 75 years in April 2024. This study produces a comprehensive flood risk map for RAK using the Analytical Hierarchy Process (AHP), which is incorporated into a GIS-based weighted overlay. Twelve factors are selected, divided into susceptibility (Elevation, Slope, TWI, Aspect, Distance from River, Drainage Density, Soil Type, Rainfall) and vulnerability (Population Density, Distance from Roads, LULC, and NDVI). Factors are then reclassified, and respective weights are computed using AHP. The weights are deemed acceptable in terms of reliability, with CRs of 7.06% and 4.86%. The weights are then used, and a weighted overlay is performed using ArcGIS Pro to produce the flood susceptibility and flood vulnerability maps. The two maps are then multiplied using the raster calculator to generate the flood risk map. The results showed that the majority of RAK falls into low, medium, and high risk categories, with 26.3%, 33.6%, and 29.1% in each class, respectively. Validation was performed using SAR imagery acquired before and after the event in April 2024. SNAP is used to perform calibration, multi-looking, speckle filtering, and terrain correction on the data. The flood risk map is reclassified into two classes: non-flooded and flooded. A confusion matrix is constructed to determine that the overall accuracy is 59.4%, recall is 54.4%, and precision is 1.34%. The relatively low precision indicates that while the model successfully identifies flood-prone zones, it may over predict flooding in areas that were not actually inundated. This limitation is likely due to the use of static spatial factors and the absence of dynamic rainfall intensity data. The model accurately identifies flood-prone zones and serves as a helpful tool for risk reduction, planning, and mitigation, even if it does not accurately anticipate the magnitude of a single occurrence.

Short Papers

Paper Nr:	39
Title:	Advancing National Planetary Boundary Assessment through Earth Observation and Machine Learning
Authors:	Bruna Almeida, Pedro Cabral, Sibyll Schaphoff, Fabian Stenzel and Dieter Gerten
Abstract:	The Planetary Boundaries (PB) framework provides a scientific basis for defining a safe operating space for humanity by identifying critical Earth system processes that regulate planetary stability. Among these, the Land-System Change (LSC) boundary captures the transformation of natural landscapes driven primarily by deforestation and land-use change. Translating the PB framework to national scales remains challenging due to data limitations and methodological constraints. This study assesses the national status of the LSC boundary in mainland Portugal by integrating dynamic vegetation modelling with spatially explicit remote sensing and machine learning approaches. Remaining natural forest extent relative to Holocene conditions was calculated using outputs from the Lund–Potsdam–Jena managed Land (LPJmL) dynamic global vegetation model and the “boundaries” software package. Current forest cover was mapped using Sentinel-2 imagery and Random Forest classification, while forest canopy structure was analysed using Tree Cover Density data from the Copernicus Land Monitoring Service (CLMS) High-Resolution Layers. The results indicate that forests currently cover 40.3% of mainland Portugal, with a mean canopy density of 51.5%, whereas the LSC indicator shows that only 30.1% of the potential natural forest extent remains relative to the Holocene-like baseline. This discrepancy highlights the distinction between present-day forest cover and the persistence of natural forest ecosystems and places Portugal near the upper limit of the increasing risk zone for the temperate forest biome. The study demonstrates the potential of combining Earth observation data and PB modelling to support national-scale assessments of LSC and to inform sustainable forest management and land-use policy.

Paper Nr:	69
Title:	Modeling and Forecasting Open-Space Loss in California, Using Geographic Information Systems (GIS)
Authors:	Navid Shaghaghi, Vidit Agarwal, Toby Asfaw and Ron Bonhagen
Abstract:	California’s open spaces (defined here as undeveloped land that supports biodiversity, carbon storage, recreation, water security, ecosystem services, and climate resilience) have undergone a sustained decline driven by urban expansion, agricultural conversion, and climate disturbance. Urban expansion and farmland conversion continue to shrink open spaces in California, eroding ecosystem services and long-term climate resilience. Although vegetation indices such as the Normalized Difference Vegetation Index (NDVI) are widely used to monitor ecosystem health, NDVI alone is insufficient to capture the structural, spatial, and anthropogenic complexity of land-cover transitions; particularly along suburban and peri-urban frontiers where vegetation may persist despite irreversible change in land-use. This paper presents a framework for modeling and interpreting open-space dynamics across California from 1984 to 2024 using a unified seven-class land-cover schema including Urban, Suburban, Farmland, Forest, Shrub/Grass, Water/Wetlands, and Desert/Barren. The framework unifies Landsat surface reflectance time series, Moderate Resolution Imaging Spectroradiometer (MODIS) land-cover products, and 64-dimensional AI-derived AlphaEarth embeddings to harmonize multi-sensor observations across four decades. The pipeline leverages Google Earth Engine for large-scale preprocessing, LightGBM for spatial modeling, SHapley Additive exPlanations (SHAP) for explainable attribution, and Prophet for long-horizon time-series forecasting. A LightGBM regressor trained on 14,309 spatial samples achieved strong predictive performance for NDVI continuity and land-cover representation (R² = 0.954, RMSE = 0.0021), enabling stable classification across multi-sensor regimes and reducing temporal flicker common in long-term remote sensing analyses. Statewide analysis reveals a net decline of approximately 13.2% in open space between 1984 and 2024, with losses concentrated along major metropolitan corridors and agricultural edges. Forecasts extending to 2035 project an additional 4.3% decline under business-as-usual scenarios, while strengthened protection policies substantially mitigate projected losses. By unifying long-term satellite data, learned spatial representations, explainable machine learning, and forecasting within a single reproducible pipeline, this study provides a scalable foundation for evidence-based conservation and land-use planning in California and beyond.

Area 5 - Sensing

Full Papers

Paper Nr:	33
Title:	On the Classification of LiDAR Point Clouds with Neural Networks Using Local and Global External Features as Auxiliary Inputs
Authors:	Stephane A. Guinard and Thierry Badard
Abstract:	The classification of 3D LiDAR point clouds is a well-studied problem, with recent deep learning approaches achieving impressive accuracy by implicitly learning multi-scale geometric representations. However, such models often rely on complex architectures and large annotated datasets, while offering limited control over early training dynamics. In this work, we investigate how a small, carefully selected set of handcrafted descriptors can be used as auxiliary inputs to guide neural network training without modifying the network architecture. We consider three types of descriptors: sensor, local, and global, and introduce a novel encoding of a geometry-based segmentation as a poitwise feature vectors. Experiments conducted on two public benchmarks using two fundamentally different architectures (RandLA-Net and SuperPoint Transformer) show that the proposed descriptors consistently improve classification accuracy (up to +2.5% mIoU, and nearly +10% on complex classes) and significantly accelerate convergence. These results suggest that explicit contextual descriptors remain a valuable complement to deep learning approaches for LiDAR point cloud classification.

Paper Nr:	46
Title:	Anthropo-Environmental Geospatial Multi-Risk Modelling of Nilotic Archaeological Sites of Saï Island in Sudan
Authors:	Sébastien Gadal and Qacem Khezami
Abstract:	Satellite data, including high-resolution imagery such as Pleiades, Sentinel-2 MSI, Landsat 5 TM and Landsat 8/9 OLI, enable spatial modelling for monitoring and analysing geographical risk exposure at archaeological sites in conflict-affected regions. The methodology combines multi-temporal satellite imagery (1990–2025) at different spatial resolutions with a Digital Elevation Model (DEM) ALOS PALSAR to analyse environmental and anthropogenic dynamics. Five geographical factors are integrated within a geographic information system (GIS) to produce a multi-scalar assessment of risk exposure. The model also incorporates hydraulic simulations of Nile floods based on 115 years of measurements from the Aswan station (southern Egypt) and 72 years from the Dongola station (northern Sudan). Demographic and urban dynamics related to agricultural expansion, as well as climate-driven revegetation processes, are also considered. Saï Island, located in Nubia, therefore provides a relevant case study for modelling and mapping the environmental and anthropogenic risks affecting Sudanese archaeological heritage.

Paper Nr:	51
Title:	Direct Change Classification of Post-Fire Land-Cover Using PlanetScope SuperDove and Machine-Learning Classifiers: A Case Study of Biokovo Nature Park (Croatia)
Authors:	Ivan Marić, Fran Domazetović and Ante Šiljeg
Abstract:	This study evaluates the applicability of machine-learning (ML) classifiers within a Direct Change Classification (DCC) framework for mapping post-fire land-cover change (LCC) using very-high-resolution PlanetScope SuperDove imagery over the Biokovo Nature Park (Croatia). Pre-fire and post-fire surface reflectance images were quality-masked and integrated into a 15-band multitemporal feature stack consisting of pre- and post-fire spectral bands, band-wise differences, and NDVI (Normalized Difference Vegetation Index) descriptors. Reference data were derived from field surveys, UAV (Unmanned Aerial Vehicle) mapping, and web-based sources, delineated as polygon core zones and converted to stratified random point samples. Four LCC classes were defined: stable vegetation, burned (high), burned (low/moderate), and stable non-vegetation. Five classifiers (Logistic Regression - LR, Random Forest - RF, Support Vector Machine with a Radial Basis Function - SVM-RBF, Gradient Boosting Machine - GBM, XGBoost) were tested under three validation strategies: stratified random split, polygon group-based validation, and spatial-block cross-validation, including controlled downsampling. Random splits produced near-perfect scores, indicating optimistic bias from spatial autocorrelation. SVM-RBF and LR showed the highest and most stable performance, particularly under reduced training data. The final LCC map delineated a coherent burn area and estimated 607.2 ha of fire-affected vegetation, consistent with reported impacts. Across all evaluated classifiers, performance remained consistently high, indicating that the DCC workflow is a reliable and generally applicable approach for quick post-fire LCC mapping from high-resolution PlanetScope SuperDove imagery.

Paper Nr:	61
Title:	Assessment of UAV Georeferencing Accuracy Using the Centipede RTK Network
Authors:	Muhammad Ali Sammuneh, Mojtaba Eslahi, Rani El Meouche and Elham Farazdaghi
Abstract:	Accurate and efficient mapping of shallow coastal waters remains challenging due to limited accessibility and the weak geometric constraints of optical imagery over homogeneous marine surfaces. In previous research, shallow water depths were derived from UAV-based optical imagery acquired during two seasonal surveys. Image processing and georeferencing were performed in Pix4D using shoreline-distributed Ground Control Points (GCPs) measured with Real-Time Kinematic (RTK) positioning. However, confining control points to a single coastal boundary introduced geometric instability, resulting in degraded image matching and reduced block robustness with increasing offshore distance. To overcome these limitations, this study presents an alternative methodology tai-lored to nearshore environments, based on direct georeferencing using a UAV equipped with an external GNSS receiver. Image coordinates were corrected us-ing data from the Centipede GNSS network, minimizing reliance on shoreline-based GCPs. Three GNSS mount points were employed to assess the internal consistency of the photogrammetric block, while independently surveyed ground points served as check points for accuracy evaluation. The Results demonstrate a substantial improvement in georeferencing robustness and positional accuracy over low-texture marine areas. The proposed workflow enhances the reliability of UAV-based surveys in shallow coastal waters and provides a practical frame-work for bathymetric mapping and coastal monitoring applications.

Short Papers

Paper Nr:	26
Title:	Towards Onboard Continuous Change Detection for Floods
Authors:	Daniel Kyselica, Jonáš Herec, Oliver Kutis and Rado Pitoňák
Abstract:	Natural disaster monitoring through continuous satellite observation requires processing multi-temporal data under strict operational constraints. This paper addresses flood detection, a critical application for hazard management, by developing an onboard change detection system that operates within the memory and computational limits of small satellites. We propose History Injection mechanism for Transformer models (HiT), that maintains historical context from previous observations while reducing data storage by over 99\% of original image size. Moreover, testing on the STTORM-CD flood dataset confirms that the HiT mechanism within the Prithvi-tiny foundation model maintains detection accuracy compared to the bitemporal baseline. The proposed HiT-Prithvi model achieved 43 FPS on Jetson Orin Nano, a representative onboard hardware used in nanosats. This work establishes a practical framework for satellite-based continuous monitoring of natural disasters, supporting real-time hazard assessment without dependency on ground-based processing infrastructure. Architecture as well as model checkpoints is available at https://github.com/zaitra/HiT-change-detection.

Paper Nr:	27
Title:	Signal Recovery with Convolutional Neural Networks for Airborne LiDAR Bathymetry
Authors:	Dandi Wang, Shuai Xing, Chaoqun Yu, Pengcheng Li, Guoping Zhang, Yang Zhou, Dongqing Zhao, Linyang Li, Yong Deng and Yan Wang
Abstract:	Airborne LiDAR bathymetry (ALB) is an advanced and effective technology for the coastal zone mapping. However, the signal detection is challenging due to the varying water conditions. Previous studies have used waveform decomposition methods to accurately extract the signal. Few studies have fo-cused on deep learning methods due to the lack of ideal training samples. In this article, we proposed a convolutional neural network architecture design for bathymetry signal recovery (BSR-CNN) to improve the signal detection rate by enhancing the signal and denoising. For training samples generation, a mathematical model for bathymetric waveforms is constructed, and the probability distributions of the parameters are estimated based pulse history. Two BSR-CNNs, BSR-CNN1 and BSR-CNN2, one for the deconvolution of shallow water waveforms and the other for the denoise of deep water wave-forms, are trained with simulated waveforms. They are validated on two field data sets in the South China Sea, and compared with the conventional methods and two existing decomposition methods. In extremely shallow wa-ters, BSR-CNN1 detects 20% more points compared to the conventional method and performs better than the decomposition method. In extremely deep waters, BSR-CNN2 supplies at least 10% more points with comparable accuracy compared to the other methods. Moreover, BSR-CNNs are signifi-cantly less time-consuming than the decomposition methods. Without the mass changes in environment and system parameter setting, this study pro-vided an effective way to fast obtain signals from ALB waveforms.

Paper Nr:	48
Title:	TPSC+P: Precision Agriculture through Satellite Image Time Series for Large-Scale Crop Mapping in Brazil
Authors:	Flavio Barbosa, Jose Salazar, Jhon Erazo, Jorge Benitez, Gustavo Mourao, Jonathan Tenorio de Lima, Kenia Mourao Santos, Moises Pereira Galvao Salgado and Marcelo Ricardo Stemmer
Abstract:	Large-scale crop monitoring requires models that can exploit irregular, multi-sensor satellite image time series while preserving parcel boundaries and producing operationally useful outputs. This paper presents TPSC+P, a three-stage pipeline for crop mapping in Brazil using Sentinel-2 and Landsat 7/8 imagery over more than 1.4 million hectares. The workflow combines preprocessing of multi-temporal patches, U-TAE module with a temporal self-attention encoder for semantic segmentation, U-TAE+PaPs module for panoptic segmentation, and a post-processing fields module (PPF) for crop typing, harvest-event detection, area estimation, and GeoJSON generation. On the Brazilian datasets used in this study, the best Sentinel configuration achieved 96.00% semantic accuracy and 90.93% IoU, while the best Landsat configuration achieved 94.12% accuracy and 83.92% IoU. For panoptic segmentation, the best models reached 47.5% PQ on Sentinel and 31.69% PQ on Landsat. The harvest classifier obtained overall test accuracies of 90.16% on Sentinel and 92.51% on Landsat. These results indicate that the proposed pipeline is effective for large-scale crop delineation and downstream field-level analytics, although challenges remain with rectangular field geometries and cross-patch field merging.

Paper Nr:	49
Title:	Automatic Band Selection for Crop Classification Using Feature-Attributions to Reduce Inference Cost
Authors:	Berk Ülker and Sander Stuijk
Abstract:	Sentinel-2 (S2) time series enable accurate crop type mapping, but multi-band and multi-date inputs quickly become high dimensional, increasing storage footprint, I/O pressure, and inference latency in large scale deployment. This paper studies how to automatically reduce the number of S2 spectral bands for time series crop classification while preserving accuracy. We estimate band importance using six complementary feature attribution methods implemented in PyTorch Captum [9], spanning gradient based, reference based, perturbation based, and model agnostic families [18, 19, 17, 12]. We then aggregate method specific rankings with an Ensemble Mean Rank (EMR) to select top-K band subsets for K ∈ {1, 2, 3}. We evaluate seven datasets from six agricultural regions in Turkey (2019-2022) and benchmark robustness across three deep time series architectures: TempCNN [14], OmniScaleCNN [20], and a Transformer encoder [22]. Selection quality is validated against the exhaustively searched best subset over all K band combinations. For K = 3, EMR achieves near optimal performance (97-99% of the exhaustive best accuracy across region-architecture pairs), while reducing input storage by approximately 70% and consistently accelerating disk→RAM→GPU transfer, yielding 1.18-1.70× end-to-end speedup. Compute savings are architecture dependent: models that preserve channel sensitivity deeper in the network exhibit substantial FLOPs reductions, whereas architectures dominated by later stage mixing primarily benefit from reduced I/O. Overall, attribution guided band selection provides a practical and transparent approach to compress S2 inputs for scalable crop mapping.

Paper Nr:	53
Title:	Bridging the Expertise Gap: A Service-Oriented Architecture for Complex InSAR Workflows within DestinE
Authors:	José A. Navarro, Àlex Flores, Oriol Monserrat and Pavel Pavlovsky
Abstract:	Advanced Interferometric Synthetic Aperture Radar (InSAR) is not an easy discipline. It is not only the level of expertise required but also the high computational power needed to process data that keeps non-specialist stakeholders away from it. This paper presents the Global Ground Motion Service (GGMS), a project still under development within the framework of the Destination Earth (DestinE) initiative. It is a cloud-native service that has been designed to build a bridge closing the gap for those non-expert potential users. Three well-seasoned tools, such as the PSIG software chain---Persistent Scatterers (PS)---, ADAfinder---Active Deformation Areas (ADA)---, and DDM---Differential Deformation Maps---, are being moved to the cloud as the components of an integrated service. Said service consists of a front-end hosted by the Destination Earth Service Platform (DESP) interacting with the final users and a specialized InSAR processing back-end. This setup shields non-expert users from technical complexity and makes it possible for them to ask for high-level products without having to deal with the complexities of InSAR processing or satellite geodesy.

Paper Nr:	62
Title:	Depression-Based Topographic Control Index and Scenario-Based Flood Activation Framework for City of Zadar (Croatia)
Authors:	Rina Miloševic
Abstract:	Pluvial flooding is an increasingly significant hazard in urban areas, driven by short-duration, high-intensity rainfall interacting with highly modified terrain. This study evaluates the applicability of the depression-based Topographic Control Index (TCI) as an indicator of pluvial flood predisposition in the City of Zadar (Croatia) and extends the static TCI framework by integrating rainfall–runoff activation scenarios. High-resolution LiDAR data (1 m) were used to identify 298 hydrologically valid surface depressions and delineate their contributing catchments. Results indicate that 93% of depressions exceeded their storage capacity under at least one rainfall scenario, with the 3-hour, 10-year event representing the dominant activation threshold. Presence-only validation showed that the top 1% of TCI values captured 78% of recorded flood events. The findings further revealed a divergence between class-based thresholds and percentile-based ranking of TCI, underscoring the continuous nature of topographic control in low-relief urban terrain. Overall, the results clarify the role of depression-based topographic structure in shaping overflow activation patterns under defined rainfall conditions.

Paper Nr:	66
Title:	Spatially Independent Land Use and Land Cover Mapping in an Arid Coastal City Using Multi-Sensor Sentinel Data
Authors:	Moustapha Mohamed Mahmoud, Nomane Nouamane, Fatimetou Taleb, Ahmed Hemmeyen and Mewloud Saghir
Abstract:	Accurate land use and land cover (LULC) mapping in arid urban environments remains challenging due to strong spatial autocorrelation, limited reference data, and the spectral similarity of surface materials. As a result, reported model performances are often overestimated when spatial dependence between training and testing samples is not explicitly controlled. This study proposes a reproducible LULC mapping workflow based on multi-sensor Sentinel-1 and Sentinel-2 data, with an explicit focus on spatially independent evaluation. The approach is applied to Nouakchott, Mauritania, a rapidly expanding coastal city characterized by extensive bare and mixed soil surfaces. Five land cover classes are defined to reflect local physical conditions, including a dedicated Mixed Bare Soil class. A systematic comparison of feature configurations and machine learning models shows that a Random Forest classifier combined with a correlation-reduced feature set provides the best balance between performance, robustness, and interpretability. Using a strictly spatial train--test split, the selected model achieves a mean F1-score of approximately 0.91. Diagnostic analyses further highlight class-specific confusions and remaining sources of uncertainty. Finally, a high-resolution LULC map for 2025 is produced, providing a robust baseline for future multi-year land cover analyses in this rapidly evolving urban context.

Paper Nr:	17
Title:	Investigating Associations between Land Use Land Cover, Temperature and Rainfall in the Western Cape Province, South Africa
Authors:	Sebastian Rogalski and Masengo Ilunga
Abstract:	This study investigates the interrelationships between Land Use Land Cover (LULC), Land Surface Temperature (LST), and rainfall in the Western Cape Province, South Africa. A multi-source dataset spanning 2017 to 2023 was utilised, incorporating South African Weather Service (SAWS) rainfall data, satellite-derived LST, elevation data, and LULC datasets. A Random Forest (RF) regression model was implemented within Google Earth Engine (GEE) to predict LST using rainfall and elevation as predictor variables. Model performance was quantitatively evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the coefficient of determination (R²). The model achieved an MAE of 1.5°C, RMSE of 2.0°C, and R² of 0.714, indicating good predictive capability in capturing spatial temperature variability across the region. The results reveal clear spatial relationships between climate and land use. High LST and air temperature conditions are associated with arid inland regions dominated by rangeland (approximately 73–76% of the province), while cooler, moderate-temperature zones with higher rainfall support more intensive agricultural activity. A general inverse relationship is observed, whereby low-rainfall areas (<200 mm/year) correspond with higher LST values (>30°C), while higher rainfall regions (>500 mm/year) are associated with cooler LST values (<25°C). Tree-covered and mountainous areas consistently exhibit lower surface temperatures, reflecting the influence of elevation and moisture availability. This study demonstrates the value of integrating machine learning with geospatial datasets to improve understanding of climate–land interactions and provides a foundation for supporting sustainable land-use planning and environmental management in the Western Cape.

Paper Nr:	44
Title:	Integrated Remote Sensing and Biophysical Modelling Framework for Multi-Temporal Forest Canopy Density Stratification in Sabah, Malaysia
Authors:	McKreddy Yaban, Zuraimi Suleiman, Shimatun Jumani Ibrahim, Muhamad Zulfazli Zakaria, Muhammad Shafiei Badrunnahar, Siti Muazah Md Zin, Muhammad Fikri Razali and Maizaitoldura Mohd Isa
Abstract:	Forest canopy density is an important measure of forest health, carbon storage, and habitat quality. Accurate canopy assessment is needed for effective forest management and climate change studies. Traditional field-based methods are limited by small coverage areas, time requirements, and observer bias. Remote sensing offers a practical solution for monitoring canopy density over large areas, although challenges remain in combining satellite data with biophysical models for accurate classification across different forest types. This study uses remote sensing techniques and biophysical modelling to improve forest canopy density classification. High-resolution satellite imagery and vegetation indices, including the Advanced Vegetation Index (AVI) and Normalized Difference Vegetation Index (NDVI), were used to create forest canopy density maps. The results show that multi-temporal Forest Canopy Density (FCD) mapping provides reliable and consistent results. Strong agreement was found between Google Earth Engine (GEE) and ArcMap from 2020 to 2024, with differences in canopy density class areas below 1% for all years. Dense forest areas remained mostly stable during the study period, while a slight increase over time indicates localized forest regeneration and natural canopy recovery. These findings show that both cloud-based and desktop GIS platforms are effective tools for long-term forest canopy monitoring and support sustainable forest management.

Paper Nr:	47
Title:	A Domain-Specific Language for the Automatic Generation of Optical Satellite Image Processing Chains Using Model-Driven Engineering
Authors:	P. F. Antenaina Hery, H. Rakotonirainy, Rabearimalala and A. R. Hajalalaina
Abstract:	With the rapid growth of satellite image processing, processing chains have become essential for transforming raw data into actionable information. They integrate key steps, from pre-processing to analysis, to support environmental monitoring and resource management. Despite advances in formalizing these workflows, their implementation remains complex for non-expert users due to the diversity of data and methods. To address this issue, this paper proposes a domain-specific language (DSL) based on a model-driven engineering (MDE) approach to enable the automatic generation of image processing chains. The proposed solution allows users to define workflows at a high level of abstraction, which are then translated into executable scripts. The approach is applied to forest cover monitoring using Sentinel-2 imagery. The results demonstrate promising performance, highlighting the effectiveness of the proposed method while indicating potential improvements for handling spectral similarities and extending the approach to other application domains.

Area 6 - Spatial Data Analysis and Management

Full Papers

Paper Nr:	29
Title:	A Custom Jaccard Index Tool to Measure Similarity between Boolean and Categorical Values in Geospatial Land Cover Data
Authors:	Timothy Mulrooney, Chima Okoli and Daniel Nduka
Abstract:	Polygonal enumerations units within the confines of a Geographic Information System (GIS) such as counties, ZIP codes and census tracts encapsulate a wide array of data, ranging from population and crime rates stored as numbers to Boolean and categorical data. Many geostatistical tools utilize quantitative attribute values and/or distance tied with location to derive descriptive, inferential, correlative and predictive statistics. There is a void in the space for all-in-one tools to compare two thematic or Boolean maps. In this paper, we developed a custom Python tool utilizing the Jaccard Index and calculated it for two disparate datasets across the National Land Cover Database (NLCD) and Cropland Data Layer (CDL) programs. Results highlight 1) Custom drop-down tools utilizing existing input parameters to create a Jaccard Index Tool are indeed possible 2) Jaccard Index comparisons for vector NLCD and CDL cross-walked to NLCD categories for a quadrangle in Durham, North Carolina, resulted in a Jaccard Index of .756 and 3) NLCD and CDL raster data representing developed land for the year 2004 in Wabash, Indiana, highlight a Jaccard Index of .384 using the binary set as the base for the Jaccard Index calculation, but using the categorical (developed/non-developed) classification, the Jaccard Index is .890. Further work on the application, education and dissemination of these different use cases is required to ensure this tool is not misapplied as this work can have distinct policy and decision-making ramifications.

Paper Nr:	34
Title:	BIM and GIS-Based Procedure to Increase the Number of Risks Identified in Safety Planning for Road Network Construction Projects in Peru
Authors:	Kevin Wilver Tenorio Felices, Grace Graciela Javier Gallegos and Karem Graciela Ulloa Román
Abstract:	Safety planning in road construction projects executed by direct administration in high-Andean regions, such as Ayacucho, presents critical deficiencies due to the use of traditional and reactive methodologies. Currently, management is based on 2D drawings and manual Hazard Identification, Risk Assessment, and Control (IPERC, for its acronym in Spanish) matrices that overlook spatial and constructive risks, increasing accident rates and construction change orders. This paper presents a preventive management procedure structured in 8 steps that integrates BIM (Building Information Modeling) and GIS (Geographic Information Systems) methodologies to improve hazard identification. The proposal was validated through a pilot case study on the "Limonchayocc" road project, utilizing the interoperability of QGIS, InfraWorks, Civil 3D, and Navisworks to simulate the environment and the construction sequence. The results demonstrated that this technological integration allowed for the detection of hidden risks not visible in the original technical file, such as regulatory geometric incompatibilities and drainage interferences, achieving a 100% increase in the detection of engineering hazards compared to the conventional method. It is concluded that the proposed procedure transforms occupational safety from a document-based approach to a technical-predictive one, proving to be a viable tool for reducing uncertainty in complex public infrastructure projects.

Short Papers

Paper Nr:	81
Title:	Spatial Robustness Assessment of Hybrid-Weighted MCDM Models for Photovoltaic Site Suitability Mapping
Authors:	Kamran Ali, Eliseo Clementini, Roberto Patrizi and Carlo Villante
Abstract:	Identifying suitable locations for utility-scale photovoltaic (PV) deployment requires spatial decision frameworks capable of integrating heterogeneous environmental and infrastructural criteria while reducing subjectivity in criteria weighting. This study proposes a hybrid weighting strategy that integrates expert judgment from the Analytic Hierarchy Process (AHP) with machine-learning-derived feature importance within a GIS–MCDM framework. Ten spatial evaluation criteria representing solar resource potential, terrain characteristics, land-use constraints, and infrastructure accessibility were integrated to assess PV suitability in the Province of L’Aquila, Italy. The hybrid weights were applied using three aggregation algorithms: Simple Additive Weighting (SAW), Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), and VIKOR to generate suitability rankings. The robustness of the resulting spatial patterns was evaluated through Spearman rank correlation and Jaccard similarity analysis across multiple land-selection thresholds. Results reveal strong consistency among the models, with correlation coefficients exceeding 0.97 and spatial overlap above 60–70 at the top suitability threshold. The most suitable zones are consistently concentrated in the Avezzano plain area of the Province of L’Aquila, which coincides with existing PV installations in the region. These results indicate that the spatial suitability pattern is primarily determined by the hybrid weighting scheme rather than by the choice of aggregation algorithm. The proposed hybrid-weighted GIS–MCDM framework provides a robust and transferable decision-support approach for renewable energy planning, improving the reliability and efficiency of spatial planning for PV deployment.

Paper Nr:	60
Title:	Mapping Photovoltaic Installations in the Province of Tarragona Using GeoAI and High-Resolution Orthophotography: A Methodological Proposal
Authors:	Lluís Salvat-Garcia, Benito Zaragozí and Antoni Domènech
Abstract:	Over the last decade, Spain's rapid photovoltaic (PV) growth has underscored the need for accurate databases for energy and territorial planning. Publicly available PV datasets often lack the spatial accuracy or level of detail needed to link installations with other territorial information (e.g., cadastral data, solar radiation potential models, or urban planning layers) and to support fine-grained local analyses. We illustrate this limitation in the province of Tarragona, where public PV registries are valuable for tracking adoption but insufficient for detailed, parcel-level assessments. To address this, we propose a geospatial artificial intelligence (GeoAI) pipeline to automatically detect PV systems from high-resolution orthophotos. The goal is to validate and enrich public PV registries, producing a quality-assured spatiotemporal geodatabase of PV diffusion. We curated a ground-truth dataset of 18,577 PV installations, where PV footprints were delineated as segmentation masks through semi-automated methods (e.g., GeoSAM) and quality-controlled via Raster Vision. A deep learning semantic segmentation model is trained and validated iteratively, using standard performance metrics (e.g., intersection over union (IoU), F1-score, precision, recall). Once validated, the model is applied to multi-year orthophotography mosaics to detect unrecorded installations and support the reconstruction of the spatiotemporal evolution of PV deployment. The model allows for retraining when new imagery becomes available, or performance drops due to domain shift. The resulting multidimensional PV database enhances energy and spatial planning, improves interoperability with territorial datasets, and can be annually updated as new public orthophotography is released.

Area 7 - Spatial Data Mining and Computation

Full Papers

Paper Nr:	30
Title:	Improving Locality-Sensitive Hashing for Fine-Grain Skyline Geolocalization
Authors:	Jesse Lew
Abstract:	This paper discusses the application of locality-sensitive hashing (LSH) techniques to the skyline geolocalization problem. New methods for skyline encoding and querying are analyzed for optimal results. We explore approaches to overcome significant challenges including storage, processing, obscurity, and scalability. Our experiments on the CH1 skyline dataset improve query resolution by a factor of 1,600:1, allowing us to geolocate image capture locations within 25m^2 over all of Switzerland without prior knowledge of an image's field of view (FoV). To our knowledge, we are the first to perform such fine-grain skyline geolocalization with LSH, and we do so while also improving both Top 1 geolocation accuracy and query speed over current state-of-the-art.

Short Papers

Paper Nr:	56
Title:	Probabilistic Route Assignment for Public Transport Carbon Footprint Estimation in Seoul: Using Smart Card Data and the I-RAPTOR Algorithm
Authors:	Suyun Lee, Juheon Lee and Chulmin Jun
Abstract:	This study proposes a probabilistic route assignment methodology for the Seoul public transport system by integrating smart card data with the I-RAPTOR (Improved Round-bAsed Public Transit Optimized Router) algorithm, as a foundational step toward link-level carbon footprint estimation. Unlike conventional approaches that rely on single optimal route assignment, the proposed method employs multi-path search to generate k candidate routes and applies a Multinomial Logit (MNL) model for probabilistic traffic assignment, enabling more realistic estimation of link-level ridership and node-level dwell population. A carbon foot-print framework that accounts for both vehicle operation emissions and facility dwell emissions at urban rail stations and bus stops is also presented. Validation using a 10% sample of smart card data demonstrates that the MNL method achieves a Top-1 route match rate of 65.54% and a Top-k route inclusion rate of 84.28%, confirming the effectiveness of multi-path probabilistic assignment in capturing diverse passenger route choice behavior. These route assignment results provide the basis for subsequent carbon footprint calculation and spatial analysis, which are outlined as a framework and will be implemented in future work.

Paper Nr:	63
Title:	Examining the Scale Effects on Machine Learning Prediction Error across Models with Varying Flexibility for Carbon-Neutral Urban Development Modeling
Authors:	Youngchul Cho
Abstract:	This study examines the impact of spatial unit size on prediction error composi-tion across machine learning models with varying structural flexibility. Grid-based spatial data at four resolutions (0.5km, 1km, 2km, and 4km) are constructed for a metropolitan region in South Korea to predict urbanized area ratios. Mean squared error is decomposed into bias squared and variance components through bootstrap resampling for a linear model (LM), random forest (RF), and extreme gradient boosting (XGB). Results indicate that bias squared constitutes the domi-nant component of prediction error across all models and scales, and that predic-tion performance degradation at coarser scales is driven primarily by rising bias squared rather than variance. The magnitude of scale-induced bias squared in-crease differs across models: XGB exhibits the largest increase, followed by RF, while the LM shows the smallest. The optimal tree depth of XGB decreases sub-stantially at coarser scales, indicating reduced learnable structure in aggregated spatial data. These findings suggest that both spatial resolution and model flexi-bility should be considered jointly in carbon-neutral urban development model-ing, as flexible models are more sensitive to information loss from spatial aggre-gation. The bias-variance decomposition framework can provide diagnostic capa-bilities beyond aggregate accuracy metrics, enabling attribution of scale-induced performance changes to specific error sources.

Paper Nr:	65
Title:	First Steps towards an Open-Source Framework to Support Public Transport Data Management
Authors:	Leonardo Monteiro-Fialho, Carlos Granell, Aaron Gutiérrez-Palomero, Benito Zaragozí, Daniel Miravet and Sergio Trilles
Abstract:	Transportation agencies increasingly generate massive streams of operational, ridership, and sensor data, yet the technical effort and human resource costs required to process and analyse these heterogeneous sources often exceed the capacities of small- to medium-sized city operators. To address this gap, this paper presents the first steps of a work in progress towards an open-source, modular, three-layer framework that (i) ingests raw datasets while preserving their native formats, (ii) enforces schemas, supports versioning, and performs incremental transformations to produce clean, curated, queryable tables, and (iii) enables interoperable access for downstream modelling, visualisation, and decision-support workflows through SQL engines and programmatic client libraries. The pipeline leverages widely adopted open-source building blocks, including object storage, distributed processing engines, and transactional table formats, and is orchestrated through standard cloud deployment practices, enabling rapid deployment on commodity cloud or on-premises hardware. A real-world evaluation on ten years of public transport data from a Spanish municipal agency, combining Automated Fare Collection and geospatial sources, demonstrates that the framework can support advanced passenger behaviour analytics, including passenger profiling, while remaining scalable, cost-effective, and easy to operate. By providing a reusable, standards-based stack, the framework lowers implementation barriers, mitigates vendor lock-in, and empowers agencies to transform raw feeds into actionable analytics, fostering data-driven decision-making for improved service quality, resource optimisation, and sustainable public transport planning.

Area 8 - Domain Applications

Short Papers

Paper Nr:	57
Title:	Analyzing the Trade-off between Operational Efficiency and Carbon Emissions in Ride-Sharing Services: An Experiment Using Late-Night Taxi Data in Seoul
Authors:	Juheon Lee, Suyun Lee and Chulmin Jun
Abstract:	This study addresses the dual challenges of supply–demand imbalance and carbon emissions in Seoul's late-night taxi services. To evaluate the potential of ride-sharing as a sustainable solution, we conducted a simulation using real-world taxi origin–destination (OD) data. Specifically, we analyzed the trade-offs among Service Rate, Mean Detour Ratio, and carbon emissions by varying the pooling intensity (α). The results indicate that while increasing α significantly improves Service Rate and reduces Total Demand emissions achieving a reduction of up to 49% compared to the no-sharing baseline it concurrently leads to higher Mean Detour Ratio. Notably, the improvement in Service Rate saturates beyond a certain threshold (α≥0.6), whereas passenger inconvenience continues to rise. Consequently, this study identifies an optimal operational range (α=0.3 to 0.5) that balances system efficiency, passenger convenience, and environmental sustainability, providing quantitative guidelines for urban mobility policies.