Weakly Supervised Learning For Near Real-Time Flood Mapping

Keywords: flood mapping; flood detection; weakly supervised; deep learning; remote sensing; image processing; machine learning;GeoAI; natural hazards
Abstract Type: Paper Abstract

Abstract

Advances in machine learning (ML), particularly deep learning, and computer vision are making significant contributions to disaster management when used in combination with remotely sensed data. While proved to be effective, existing supervised methods require intensive human labeling of flooded pixels to train a multi-layer deep neural network that learns abstract semantic features of the input data, hindering the applications in emergency disaster response. Since the time-consuming hand labeling is infeasible in operational disaster response workflows, automatic label generation is the key to a real-time flood mapping system. To this end, this research proposes a weakly-supervised pixel-wise flood mapping framework by leveraging multi-temporal remote sensing imagery, image processing techniques (e.g., thresholding, k-means clustering, and edge detection), and the fusion of spatial, temporal, and spectral information to create weakly- labelled training and test datasets without the need for time-consuming and labor-intensive human annotations. Using the floods from Hurricanes Florence and Harvey as case studies, the experimental results demonstrate the feasibility of the proposed framework for generating high-quality, large-scale labels for flood mapping, significantly accelerating the flood hazard response.