Periodic Reporting for period 1 - STURM (SmarT sub-pixel URban flood Mapping from open earth observation and crowdsourcing)
Berichtszeitraum: 2023-06-12 bis 2025-06-11
Urban flooding represents one of the most pressing and frequent climate-related hazards, driven by rapid urbanization, land use change, and the intensification of extreme weather events.
Timely and precise mapping of flood extent and depth is essential to reduce risks, support emergency response, and inform climate adaptation strategies.
However, traditional flood mapping approaches often suffer from limitations in spatial resolution, data availability, and the timeliness of information, particularly in urban contexts where detailed, localized insights are needed.
The STURM project (SmarT sub-pixel URban flood Mapping from open earth observation and crowdsourcing) addressed these challenges by developing an innovative framework for urban flood mapping using open satellite imagery, crowdsourced visual data, and state-of-the-art deep learning techniques.
Funded under the Marie Skłodowska-Curie Actions, the project aimed to improve the spatial granularity and usability of flood extent and depth maps at city scale, using only globally accessible and non-commercial resources.
The research was designed to be ethically robust, transparent, and reproducible, contributing to scientific knowledge while aligning with EU goals in climate resilience, open science, and digital innovation.
Timely and precise mapping of flood extent and depth is essential to reduce risks, support emergency response, and inform climate adaptation strategies.
However, traditional flood mapping approaches often suffer from limitations in spatial resolution, data availability, and the timeliness of information, particularly in urban contexts where detailed, localized insights are needed.
The STURM project (SmarT sub-pixel URban flood Mapping from open earth observation and crowdsourcing) addressed these challenges by developing an innovative framework for urban flood mapping using open satellite imagery, crowdsourced visual data, and state-of-the-art deep learning techniques.
Funded under the Marie Skłodowska-Curie Actions, the project aimed to improve the spatial granularity and usability of flood extent and depth maps at city scale, using only globally accessible and non-commercial resources.
The research was designed to be ethically robust, transparent, and reproducible, contributing to scientific knowledge while aligning with EU goals in climate resilience, open science, and digital innovation.
The STURM project delivered two main significant technical and scientific contributions to advance urban flood monitoring through open Earth observation and crowdsourced imagery.
First, the project developed STURM-Flood, a large-scale open-access dataset for urban flood extent mapping using Sentinel-1 and Sentinel-2 imagery and Copernicus Emergency Management data as reference. The dataset comprises more than 20000 labeled tiles derived from 60 flood events worldwide. Based on this dataset, a dual-stream U-Net architecture was trained to segment flood extent from either radar or optical inputs. The models achieved consistent and transferable performance across diverse urban and anthropized settings, demonstrating the feasibility of using free satellite data and deep learning for near-real-time flood extent mapping. The dataset and source code are publicly released to support reproducibility and future research.
Second, the project introduced STURM-FloodDepth, a modular pipeline to estimate and geolocate urban flood depth from single street-level and oblique aerial images. Vehicles were selected as reference objects due to their standardized dimensions and presence in urban scenes. The workflow consists of vehicle detection using YOLO-World and SAHI, contextual cropping, super-resolution with EDSR, and flood depth classification with a fine-tuned ResNet-50 model. The classifier was trained on 3,367 image patches and achieved AUC values from 0.78 to 0.98. A cross-view geolocation module using SuperGlue and RANSAC projected detected flood depths onto orthophotos, enabling spatialization. The annotated dataset, trained model, and code are released as open-source resources.
All results were published in high-impact peer-reviewed journals and presented at international conferences and workshops.
Together, these outputs form a comprehensive toolkit for urban flood mapping from open data, with practical relevance for disaster management, urban resilience planning, and climate adaptation efforts.
The project completed a case study on the 2021 Luxembourg flood event. The methodologies were qualitatively validated and discussed with institutional stakeholders, supporting their practical relevance for urban resilience planning.
First, the project developed STURM-Flood, a large-scale open-access dataset for urban flood extent mapping using Sentinel-1 and Sentinel-2 imagery and Copernicus Emergency Management data as reference. The dataset comprises more than 20000 labeled tiles derived from 60 flood events worldwide. Based on this dataset, a dual-stream U-Net architecture was trained to segment flood extent from either radar or optical inputs. The models achieved consistent and transferable performance across diverse urban and anthropized settings, demonstrating the feasibility of using free satellite data and deep learning for near-real-time flood extent mapping. The dataset and source code are publicly released to support reproducibility and future research.
Second, the project introduced STURM-FloodDepth, a modular pipeline to estimate and geolocate urban flood depth from single street-level and oblique aerial images. Vehicles were selected as reference objects due to their standardized dimensions and presence in urban scenes. The workflow consists of vehicle detection using YOLO-World and SAHI, contextual cropping, super-resolution with EDSR, and flood depth classification with a fine-tuned ResNet-50 model. The classifier was trained on 3,367 image patches and achieved AUC values from 0.78 to 0.98. A cross-view geolocation module using SuperGlue and RANSAC projected detected flood depths onto orthophotos, enabling spatialization. The annotated dataset, trained model, and code are released as open-source resources.
All results were published in high-impact peer-reviewed journals and presented at international conferences and workshops.
Together, these outputs form a comprehensive toolkit for urban flood mapping from open data, with practical relevance for disaster management, urban resilience planning, and climate adaptation efforts.
The project completed a case study on the 2021 Luxembourg flood event. The methodologies were qualitatively validated and discussed with institutional stakeholders, supporting their practical relevance for urban resilience planning.
STURM advanced the state of the art in flood monitoring by introducing novel open-access tools tailored for urban environments and grounded in globally available data sources.
- Development of STURM-Flood: a FAIR, accessible, global, deep learning-ready dataset for deploying broadly applicable models that can be particularly beneficial for vulnerable and data-scarce areas. The dataset was curated to ensure zero no-data pixels and a minimum 1% water-presence threshold per tile.
- Development of STURM-FloodDepth, a pipeline that pioneers the use of in-the-wild street-level and oblique imagery to estimate flood-depth levels based solely on vehicle submersion, eliminating anthropometric bias and ensuring ethical, inclusive modelling.
- The modular pipelines reached near real-time inference capability on consumer-grade hardware, making it suitable for rapid post-event analysis by civil protection agencies.
- Novel cross-view feature matching integration to georeference flood-level observations from arbitrary viewpoints onto high-resolution orthophotos.
These outcomes provide scientific and technological advances by delivering benchmark datasets and open architectures that accelerate disaster‐mapping research, operational benefits through rapid, low‐cost flood monitoring workflows for emergency responders and urban planners, commercial and societal value by underpinning insurance‐risk modelling services, smart‐city dashboards, and community‐driven flood reporting networks, and policy and standardization support by supplying empirical evidence of AI‐based solutions for natural‐hazard resilience.
To fully realize STURM’s potential, future efforts must proceed cautiously along three complementary tracks. First, targeted research should secure reference data with high vertical and horizontal resolution, capturing flood depth and extent and in volumes sufficient for automated model training to enable full sub-pixel mapping and robust fusion with hydrodynamic simulations, DEMs, and ancillary layers. Second, carefully scoped demonstration pilots with national civil-protection agencies and water authorities, including live trials during forecast flood events, are needed to validate near-real-time performance and refine operational workflows under realistic conditions. Third, sustainable financing through partnerships with insurance and catastrophe-modelling firms and engagement with EU funding programmes will be essential to improve, scale and maintain the technology. Each step should involve close collaboration with data providers, end users, and regulators to ensure technical feasibility, ethical compliance, and alignment with evolving standards.
- Development of STURM-Flood: a FAIR, accessible, global, deep learning-ready dataset for deploying broadly applicable models that can be particularly beneficial for vulnerable and data-scarce areas. The dataset was curated to ensure zero no-data pixels and a minimum 1% water-presence threshold per tile.
- Development of STURM-FloodDepth, a pipeline that pioneers the use of in-the-wild street-level and oblique imagery to estimate flood-depth levels based solely on vehicle submersion, eliminating anthropometric bias and ensuring ethical, inclusive modelling.
- The modular pipelines reached near real-time inference capability on consumer-grade hardware, making it suitable for rapid post-event analysis by civil protection agencies.
- Novel cross-view feature matching integration to georeference flood-level observations from arbitrary viewpoints onto high-resolution orthophotos.
These outcomes provide scientific and technological advances by delivering benchmark datasets and open architectures that accelerate disaster‐mapping research, operational benefits through rapid, low‐cost flood monitoring workflows for emergency responders and urban planners, commercial and societal value by underpinning insurance‐risk modelling services, smart‐city dashboards, and community‐driven flood reporting networks, and policy and standardization support by supplying empirical evidence of AI‐based solutions for natural‐hazard resilience.
To fully realize STURM’s potential, future efforts must proceed cautiously along three complementary tracks. First, targeted research should secure reference data with high vertical and horizontal resolution, capturing flood depth and extent and in volumes sufficient for automated model training to enable full sub-pixel mapping and robust fusion with hydrodynamic simulations, DEMs, and ancillary layers. Second, carefully scoped demonstration pilots with national civil-protection agencies and water authorities, including live trials during forecast flood events, are needed to validate near-real-time performance and refine operational workflows under realistic conditions. Third, sustainable financing through partnerships with insurance and catastrophe-modelling firms and engagement with EU funding programmes will be essential to improve, scale and maintain the technology. Each step should involve close collaboration with data providers, end users, and regulators to ensure technical feasibility, ethical compliance, and alignment with evolving standards.