Intelligent Disaster Classification in Images Using Attentional Universal Inverted Bottleneck

Earthquakes, hurricanes, and floods are among the most disastrous events that destroy a considerable part of the infrastructure and affect millions, and the total losses in 2023 are estimated at $ 313 billion worldwide. Disaster assessment is a vital process that should be fast and correct to lead to efficient response and deployment of resources. In this paper, the author cites DisasterNN, a convolutional neural network (CNN) trained on disaster image classification, utilizing the Comprehensive Disaster Dataset (13,557 images) of six categories: Damaged Infrastructure, Fire Disaster, Human Damage, Land Disaster, Non-Damage, and Water Disaster. The DisasterNN incorporates Universal Inverted Bottleneck Blocks and Triplet Attention, which improves feature extraction at the discretion of a complex visual data. Trained using a $70-10-20$ division between training, validation and testing data, DisasterNN results in robust hurt on a completion outcome of 92.89 percent accuracy, 83.06 percent precision, 80.17 percent recall, 81.40 percent F1-score, 97.51 percent AUROC and 86.18 percent Matthews Correlation Coefficient. These findings prove that DisasterNN can perform quick crucial analysis, which allows the deployment of resources efficiently and possible lives saved. In our results, we are focused on the idea that the refined CNN network, such as DisasterNN, is going to be the next major step towards automated disaster assessment and response.