Skip to main content
To KTH's start page To KTH's start page

Supervised and Unsupervised Deep Learning Models for Flood Detection

Time: Wed 2023-11-15 14.00

Location: Bora Bora, Teknikringen 10B, Stockholm

Video link:

Language: English

Subject area: Geodesy and Geoinformatics, Geoinformatics

Doctoral student: Ritu Yadav , Geoinformatik

Opponent: Associate Professor Nicolas Audebert, Conservatoire National des Arts et Métiers (CNAM), France

Supervisor: Yifang Ban, Geoinformatik; Associate Professor Andrea Nascetti, Geoinformatik

Export to calendar

QC 20231030


Human civilization has an increasingly powerful influence on the earthsystem. Affected by climate change and land-use change, floods are occurringacross the globe and are expected to increase in the coming years. Currentsituations urge more focus on efficient monitoring of floods and detecting impactedareas. Earth observations are an invaluable source for monitoring theEarth’s surface at a large scale. In particular, the Sentinel-1 Synthetic ApertureRadar (SAR) and Sentinel-2 MultiSpectral Instrument (MSI) missionsoffer high-resolution data with frequent global revisits that are widely usedfor flood detection.Current solutions such as Copernicus Emergency Management Services(CEMS), MODIS (Moderate Resolution Imaging Spectroradiometer) globalflood product, and many others use data from Sentinel and multiple othersatellites to detect floods. Although existing solutions are helpful, they alsohave several limitations. For instance, solutions like MODIS global floodproduct detect floods solely on optical images causing poor or no detection incloudy areas. In addition, these solutions are threshold-based and often requirecriteria-based adjustments. Furthermore, these solutions do not leveragerich spatial information between neighboring pixels and don’t use temporalfeatures of time series data. Therefore, advanced processing algorithms areneeded to provide a reliable method for flood detection.This thesis presents three Deep Learning (DL) models for flood detection.The first two models are supervised segmentation models proposed todetect floods on uni-temporal Sentinel-1 SAR data. The study sites containfloods from Bolivia, Ghana, India, Mekong, Nigeria, Pakistan, Paraguay,Somalia, Spain, Sri Lanka and USA. The third model is an unsupervised spatiotemporalchange detection (CD) model that detects floods on time series ofSentinel-1 SAR data. The study sites contain floods from Slovakia, Somalia,Spain, Bolivia, Mekong, Bosnia, Australia, Scotland and Germany.The two supervised segmentation models propose improving flood detectionwith the help of self-attention mechanism and fusion of Sentinel-1 SARwith more contextual information. The first network is ’Attentive U-Net’. Ittakes Sentinel-1 channels VV (vertical transmit, vertical receive), VH (verticaltransmit, horizontal receive), and the ratio VV/VH as input. The networkuses spatial and channel-wise self-attention to enhance feature maps resultingin better segmentation. The second network is a dual-stream attentive ’Fusionnetwork’, where the global low-resolution elevation data and permanent watermasks are fused with Sentinel-1 (VV, VH) data. The ’Attentive U-Net’ yields67.2% Intersection over Union (IoU), and the ’Fusion network’ gave 69.5%IoU on the Sen1Floods11 dataset. The performance gain is 3 to 5% IoUwith respect to the existing supervised models like FCNN (49.3% IoU score),U2Net (62% IoU score), and BASNet (64% IoU score). Quantitatively, thetwo proposed networks show significant improvement over benchmark methodsdemonstrating their potential. The qualitative analysis demonstrates thecontribution of low-resolution elevation and a permanent water mask in enhancingflood detection. Ablation experiments further clarify the effectiveness of ratio, self-attention, ratio and various design choices made in proposed networks.Furthermore, to improve across-region generalizability of the flood detectionmodel and to eliminate the dependency on labels, a novel unsupervisedCD model is presented that detects floods as changes on SAR time seriesdata. The proposed model is trained to learn spatiotemporal features of theSAR time series data with the help of unsupervised learning techniques, reconstruction,and contrastive learning. The change maps are generated witha novel algorithm that utilizes the learned latent feature distributions of preand post-flood data. The model achieved an average of 70% IoU score, outperformingexisting flood detection models like RaVAEn (45.03% IoU score),cGAN (51.49% IoU score) and SCCN (54.87% IoU score) with a significantminimum margin of 15% IoU score. The proposed model is tested for generalizabilityand outperformed supervised models ADS-Net and DAUSAR whentested on unseen CEMS flood sites. In addition, an automatic change monitoringand change point detection framework is proposed. The framework isbased on the proposed unsupervised CD model where time series data is processedthrough the model to identify percentage change at each time stampand the change point is detected by identifying the date on which significantchange started to reflect on SAR data. When integrated with high temporaldata i.e. daily images from ICEYE, the framework can help in continuousflood monitoring and early detection of slowly proceeding disaster events,giving more time for response.Overall, this thesis contributes supervised and unsupervised flood detectionmodels, enabling comprehensive and widely applicable flood mapping andmonitoring capabilities. These advancements facilitate near-real-time disasterresponse and resilient urban development, thus contributing to SDG 11 -Sustainable Cities and Communities.