Disaster Anomaly Detector via Deeper FCDDs for Explainable Initial Responses
share
Urban and rural areas can often be devastated by extreme natural disasters. Towards any disaster event, an initial response is the key to rescuing within 72 hours and prompt recovery. For the stage of initial responses, it is important to quickly recognize the disaster damage over a wide area and determine priority areas. Among machine learning algorithms, deep anomaly detection is effective in detecting devastated features that are different from ordinary vision everyday life. In addition, explainable computer vision applications have been expected to justify the initial responses. In this paper, we propose an anomaly detection application utilizing the deeper fully-convolutional data descriptions (FCDDs), that enables to localize devastated features and visualize damage-marked heatmaps. More concretely, we show numerous training and test results to a dataset AIDER with the four disaster categories: collapsed buildings, traffic accidents, fires, and flooding areas. We also implement ablation studies of anomalous class imbalance and the data scale competing against the normal class. Finally, we discuss future works to improve more robust, explainable applications for effective initial responses.
READ FULL TEXT
