A Human-Trustable Self-Improving Machine Learning Framework for Rapid Disaster Responses Using Satellite Sensor Imagery

Due to the abrupt changes in Earth's climate and the dramatic global rise of urbanisation, natural disasters have become unpredictable and caused great social and economic devastation in recent years. According to one published study, between 2015-2019, there were a total of 1624 reported natural disasters, such as earthquakes, floods, landslides, etc., killing on average 60,000 people each year globally. Although humans cannot prevent natural disasters in most cases, timely responses can play a critical role in disaster relief and life-saving. Rapid and accurate building damage assessment (BDA) is required in humanitarian assistance and disaster response to carry out life-saving efforts. However, current BDAs are mostly based on manual inspection and documentation, which is time consuming and labour-intensive.

Although high-resolution satellite sensor images (HRSSIs) such as GeoEye-1, WorldView-2 and 3, have become the major source of first-hand information for BDA, those images often present a mosaic of complex geometrical structures and spatial patterns. Automatic information extraction from HRSSIs of disaster-affected areas is imperative under time-critical situations, and has the potential to facilitate post-disaster assessment, speed up the life-saving rescue processes. However, this remains an extremely challenging task for the state-of-the-art machine learning (ML) algorithms. In practice, human experts have to manually interpret and examine the captured HRSSIs, which involves significant time and labour.

Conventional ML-based BDA methods leverage mainstream classifiers, such as support vector machine, random forest, to generate a damage map based on hand-crafted features extracted from pre- and post- disaster images. However, the complexity and heterogeneity of HRSSIs hinder the applicability of conventional methods, making feature extraction extremely difficult. Besides, buildings often involve only a few pixels, leaving minimal structural information...