Physics-Informed Weakly Supervised Learning for Near Real-Time Flood Mapping

Keywords: remote sensing, machine learning, disaster mitigation
Session Type: Virtual Paper Abstract
Day: Tuesday
Session Start / End Time: 3/1/2022 05:20 PM (Eastern Time (US & Canada)) - 3/1/2022 06:40 PM (Eastern Time (US & Canada))
Room: Virtual 18

Abstract

Advances in deep learning and computer vision are making significant contributions to disaster mitigation when used in combination with remotely sensed data. Although existing supervised methods proved to be effective, they require intensive human labeling of flooded pixels to train a multi-layer deep neural network that learns abstract semantic features of the input data. Moreover, training a deep neural network on a single human-annotated ground truth flood mask may not make the model generalize well for other floods due to the highly variable image background for different flooded events, limiting its performance in real-time for upcoming disasters. This study proposes a weakly supervised pixel-wise flood mapping approach by leveraging multitemporal data. The proposed method utilizes ground truth data generated from traditional remote sensing techniques (e.g., NDWI thresholding) to train the multitemporal UNet model for flood detection to improve the performance of current pixel-wise flood mapping approaches without the need for human labels. Otsu’s method for thresholding and edge detection for noise removal is applied to the difference NDWI image to generate a weakly-labeled ground truth flood mask over different flood events. Considering the properties of multitemporal images, this study will also verify the effectiveness of incorporating pre- and post-disaster images compared to the post-disaster image only. The effectiveness of the proposed framework and model are tested against traditional machine learning algorithms including SVM, RF, logistic regression, and AdaBoost.