Advancing Intelligent Interpretation of Remote Sensing Imagery for Disaster-related Applications Using Deep Learning with Limited Labels
Abstract
In recent years, the frequency and intensity of natural disasters have increased significantly due to global climate change. Remote sensing (RS) combined with cutting-edge artificial intelligence (AI) techniques, such as deep learning (DL), has been proven effective in rapidly acquiring ground information in disaster-related applications. However, the lack of annotated data limits the usability of DL in disaster scenarios. This thesis investigates the utilization of the latest DL algorithms based on limited labels for RS-based disaster tasks using high-resolution optical images, aiming to enhance the applicability of DL in real-world scenarios.
Firstly, given the advancements of semi-supervised learning (SSL) algorithms that leverage a mass of unlabeled data, a consistency regularization (CR)-based SSL framework is developed for RS image semantic segmentation. Encouragingly, based on five datasets with diverse tasks, e.g., road extraction, building detection, and land cover classification, the proposed SSL method using only 5% labeled data achieves a relative accuracy ratio over 89% compared to the fully supervised learning method using 100% labeled samples.
Secondly, to extract floodwater rapidly and accurately in urban areas with dense shadows, a modified fully convolutional network model is designed and integrated with a novel SSL framework incorporating CR, RandMix, and test-time augmentation techniques. Experiments on aerial images of the 2013 Calgary flood demonstrate that the presented method achieves an impressive F1-score of 96.34% for flood mapping, utilizing only 4.47% of the total labeled data.
Thirdly, to meet the urgent need for timely and accurate building damage assessment, a novel SSL framework is proposed, combining multitask semantic segmentation with a perturbed dual mean teachers’ scheme. Experiments on three datasets indicate that even with a small fraction of labeled samples (e.g., 5%), the proposed method can generate meaningful results for building damage assessment, offering a potential solution for timely disaster response in emergencies.
Lastly, this thesis presents an operational workflow for rapid urban flood mapping, including a novel weak training data generation strategy and an end-to-end weakly supervised learning (WSL) framework with structure constraints and cross self-distillation. The proposed workflow better balances timeliness and accuracy in flood mapping, exhibiting promising operability in response to urban floodings.