Application of convolutional neural networks to detect damaged buildings

Abstract

The paper considers the use of convolutional neural networks for detecting dam-aged buildings in images. The research is aimed at improving the understanding and practical application of convolutional networks in the tasks of recognizing damaged buildings in images. To achieve the goal, the optimal convolutional network architecture, namely U-Net, was determined. The U-Net architecture is used in semantic segmentation tasks, which is an advantage for damaged building recognition, as it introduces higher accuracy at the pixel level, among other advantages is the simplicity of the model and the ability to use less data for training. To train the network, a unique dataset was created using the available Google Earth images of the city of Mariupol in 2022. Different con-figurations of models were considered in the work, such metrics as overall accuracy, IoU (Intersection over Union) separately for two classes of buildings and the average IoU for all classes were used to calculate the quality. The best model among those trained with the optimal parameters for solving the task was determined.