Optimizing Disaster Impact Detection Pipelines

Started Aug. 1, 2022

Abstract or project description

Natural disasters have seen an uptick in recent years, most recently major earthquakes in Turkey and Syria. Real-time data is crucial to first responders, a key part of which is monitoring building damage. This research aims to reduce model complexity of current existing research to achieve faster detection of damaged buildings. Data is sourced from the xView Challenge containing satellite images, with the intention of detecting all buildings in various states of damage. Model complexity is a key concern due to the standard approach of semantic segmentation labeling each pixel; this research proposed to simplify outputs into bounding boxes along with classes. This becomes especially relevant on satellite images, which tend to have extremely high resolution. Major contributions included loading and preprocessing data, transforming dataset-provided polygons into bounding boxes, and architecture design, in addition to training and testing pipelines. Resultant model accuracy achieved 62.6%. In conclusion, the bounding box plus semantic segmentation approach has potential to become accurate and effective if improvements are made. Future work should focus on improving this model, such as having variable output length (including polygonal shapes and polygon count) and utilizing pre-existing models like YOLO and SSD.