Post Flood Damage Assessment Using Aerial Imagery

Visual Scene understanding is one of the most efficient methods to assess the damage caused by natural disasters. Many datasets exist that contain aerial images from satellites or UAVs. Floodnet is one of those datasets containing images useful to assess the damage done by a flood. Different network architectures have been developed for classification and semantic segmentation using supervised learning to achieve reasonable accuracy. However, they cannot address the problem of considerable class imbalance in the Floodnet dataset. This imbalance arises due to the few labeled images, as labeling an aerial image is an expensive and high-effort task. This imbalance in classes prevents the model from achieving good accuracy. This work uses a semi-supervised learning approach to address the classification and semantic segmentation problem. The classical Pseudo Labeling approach has been implemented on the best-performing model in classification and semantic segmentation objectives. Pseudo-labels are assigned to unlabeled images by predicting classes from labeled images based on the probability distribution. These pseudo-labels are then trained with ground-truth data to achieve better results. Classification models train themselves in a semi-supervised manner based on the pseudo-labels' generation algorithm. Semantic Segmentation models also generate pseudo-label masks for pixel-level identification. Classification model Resnet50 trained in a semi-supervised manner achieves accuracy and Fl-score of 94.97% and 0.971, respectively. Semantic Segmentation model Segformer achieves a Mean Intersection over Union (MIoU) score of 0.596.