Excellence in Research: Mapping of Natural Disasters by Deep Subspace Learning in Multi-band and Multi-spectral Satellite Images

This proposal develops a suite of techniques for identification of regions affected by natural disasters on satellite images to monitor natural phenomena and map natural disaster events. Such events include flooding, wildfires, earthquakes, faulting, volcanic activity, and landslides. Accurate automated image analysis and machine learning techniques enable the identification of areas most affected by a natural disaster that is critical for saving human lives and assessing damages. Real time monitoring using high performance computing and a large dynamic database of disaster data will be the result of this work. Satellite remote sensing systems produce abundant global coverage and repetitive measurements via simultaneous and/or multi-temporal observations of the same features by different sensors. However, most sensors are hosted by different platforms; therefore, combining this tremendous amount of data becomes a non-trivial task. Human volunteers can perform co-registration and change detection manually; however, it is time-consuming and depends on their level of training, which is often limited. To address these limitations, this project develops a system that detects semantic visual information to guide registration and detection of areas affected by natural disasters. It employs deep network architectures to obtain regional representations for multi-temporal and multi-band image registration and fusion for natural disaster mapping. This methodology uses concepts from self-supervised and semi-supervised learning to reduce the requirement for large training data. It then utilizes deep subspace learning to identify changes caused by natural disasters. For the case of multiple disaster types, it employs supervised learning techniques to classify the disaster type. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.