| dc.description.abstract | Advances in deep learning and computer vision are making significant contributions to disaster management 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 transferable to other floods due to the highly variable image background for different flooded events, limiting its performance in real-time for upcoming disasters. This thesis proposes a weakly-supervised pixel-wise flood mapping approach by leveraging multi-temporal RS imagery and automatically generated labels for model training. Specifically, the proposed method utilizes ground truth data (i.e., labels) generated from traditional remote sensing techniques (e.g., NDWI thresholding) to train the bi-temporal UNet model for flood detection to improve the performance of current pixel-wise flood mapping approaches without the need for human labels. In addition, various image processing methods, including histogram thresholding, k-means clustering, and edge detection for noise removal, are applied to the NDWI difference image to generate a weakly-labeled ground truth flood mask over different flood events, to further optimize the model performance. Using the floods from Hurricanes Florence and Harvey as case studies, the proposed weekly-labeled bi-temporal UNet model achieved a higher performance of around 3.3% on average, compared to baseline machine learning algorithms including decision tree, random forest, gradient boost, and adaptive boosting classifiers. | en_US |
