Integrating Digital Elevation Models for Enhanced Flood Segmentation from UAV Imagery

Michiel De Baets; Brian Booth; Hiep Luong; Michiel Vlaminck

doi:10.5281/zenodo.17435314

Integrating Digital Elevation Models for Enhanced Flood Segmentation from UAV Imagery

Michiel De Baets; Brian Booth; Hiep Luong; Michiel Vlaminck

Publisher: Zenodo
DOI: 10.5281/zenodo.17435314
Repository: zenodo.org
Language: en

Abstract

Semantic segmentation models are crucial for flood detection from UAVs, yet their performance is often limited because of the complex appearance of flood water. Shadows, reflections, suspended particles, ground colour and partial occlusions introduce large visual variability. This abstract of work describes a novel approach to enhance flood water detection by integrating Digital Elevation Models (DEMs) to post-process the output of classical deep learning methods for semantic segmentation. Our methodology is built upon two core concepts that leverage terrain data to enhance model performance. First, we utilize DEMs to partition the landscape into distinct watershed basins, which serve as local domains for analysis. This approach allows us to assume a single, flat water level within each basin, a geophysical sound simplification that traditional techniques overlook. By aggregating the activation values from a classical RGB convolutional neural network (CNN) within these basins, we create a more refined representation of the potential flood extent. Secondly, we introduce a custom cost function designed to maximize the model's support for a specific flood level within each basin. This function optimizes the water level for the summed output activations of the CNN to align with the physically defined flood boundaries. We demonstrate the effectiveness of our approach on a novel dataset of 4 drone videos from flooding events in Flanders. The results show a significant increase in Intersection over Union (IoU) scores, with gains of over 20% compared to conventional deep learning models. These findings highlight the critical role of incorporating physical domain knowledge into deep learning for flood mapping.

Full text

[email protected] Integrating Digital Elevation Models for Enhanced Flood Segmentation from UAV Imagery IPI: IMAGE PROCESSING AND INTERPRETATION Michiel De Baets, Brian Booth, Michiel Vlaminck and Hiep Luong Problem statement ①Oblique UAV video High resolution images can be captured with UAVs by flying close to the disaster area. Because of the oblique camera angle, it is possible to distinguish the flood water edge on near vertical slopes. Deducing the flood height from a nadir view, would be impossible. ②Fast and cheap deployment Opposed to techniques that require satellite data, UAV deployment can be fast and relatively cheap. RGB cameras are present on practically all commercially available drones. ③Stakeholders •Quickly assess disaster area for first responders •Detailed severity study post-event for insurance companies •Realtime flood prediction & offline model verification Custom cost function A core concept in this approach is the process of finding the optimal water level per watershed basin, defined as follows: 𝑐 𝑡 =෍ xቊ𝐴𝑤𝑥 − 𝐴𝑜(x) if𝐻 𝑥 >𝑡 𝐴𝑜𝑥 −𝐴𝑤𝑥if𝐻 𝑥 ≤𝑡 The cost for a certain flood level is the sum of the rescaled, normalized activation values, with H(x)the height at pixel x and Aw(x),Ao(x)the normalized activation values. The time complexity of this minimization problem comes down to sorting the activation values by height and a single pass to find the minimum, so O(n log(n)). Method Camera poses Video frames DSM RGB CNN flood water segmentation Mesh renderer Normalized activations Height maps Basin maps Basin cost/height plots Two core concepts Masked output N: number of images in the sequence B: number of watershed basins xN xN xN xB xB Activation / height Cost function 𝑐𝑡𝑜𝑡𝑎𝑙 =𝑐𝑤𝑎𝑡𝑒𝑟 +𝑐𝑜𝑡ℎ𝑒𝑟 Watershed basins We use DEMs to partition the landscape into distinct watershed basins, which serve as local domains for analysis. This approach allows us to assume a single, flat water level within each basin, a geophysical sound simplification that traditional techniques overlook. Results IoU % of CNN IoU % of CNN IoU % of CNN IoU % of CNN RGB CNN 0.223 0.344 0.574 0.483 DSM > DTM 0.452 ± 0.027 202.9% ± 12.2% 0.464 ± 0.012 135.0% ± 3.6% 0.595 ± 0.006 103.7% ± 1.1% 0.546 ± 0.025 113.2% ± 5.3% DTM > DTM 0.281 ± 0.004 126.4% ± 1.9% 0.523 ± 0.000 152.0% ± 0.0% 0.598 ± 0.007 104.1% ± 1.2% 0.630 ± 0.010 130.6% ± 2.0% Masked output Algorithm -> Render Preprocessing settings IMEC ICON PROJECT: FLOODIFY Tackling urban flash floods through real-time data integration Gaverbeek Erpe-Mere Ijzer Londerzeel RGB CNN Our Method Annotated Flood Mask [1] Drone video footage used by permission of the Flemish Environment Agency. Compressed video available on waterinfo.Vlaanderen.be [2] Agency Digital Flanders, “Digitaal hoogtemodel vlaanderen ii,” 2014-03-11. acquired under the model license for free reuse Vlaanderen v1.0. Conclusion The results show a significant increase in IoU-scores, with gains of over 20% compared to a conventional CNN. These findings highlight the critical role of incorporating physical domain knowledge into deep learning for flood mapping.