Automating disaster detection via remote sensing

The challenge of detecting disasters is driving a wide variety of deep learning approaches. This paper describes recent advances including deep learning models such as SegNet, U-Net, FCNs, FCDenseNet, PSPNet, HRNet, and DeepLab, which have been applied to detect and analyze natural disasters like forest fires, floods, and earthquakes through computer vision analysis of remote sensing images and analysis of seismographic readings. These AI tools can distinguish between different land types (e.g. burned and unburned), detect damaged infrastructure, and identify areas at risk of further danger, significantly enhancing our ability to respond to and manage natural disasters. However, challenges remain in the use of semantic segmentation approaches for earthquake and flood detection, such as limited generalizability, limited representation of all flood types, and difficulties in detecting small flooded regions.

As more deep learning models are developed, disaster response efforts will be able to operate with a clearer understanding of conditions on the ground. However, refining these models to perform well under the diverse variety of conditions disasters present around the world will require considerable additional research.