Flood detection in UAV images using PSPNet and uncertainty quantification with Monte-Carlo Dropout technique

Floods are one of the most frequent natural disasters that affect the society, impact human lives, and make costs for governments. Utilizing new technologies helps managers and first responders to decrease the damaging effect of floods and save time. Unmanned Aerial Vehicles equipped with accurate sensors along with powerful computer vision and deep learning techniques can act as potential platforms for surrveilance, mapping and detection of flooded regions. In this study, PSPNet as the main architecture enhanced by ResNeSt as the encoder, are utilized for semantic segmentation of very high resolution drone imagery acquired from urban flooded regions. Furthermore, in order to interpret and study the performance of the method, Monte-Carlo Dropout (MCD) technique is used as a Bayesian estimator for uncertainty quantification of the results. Comparing the results of our method with other models indicated that increasing the complexity and number of parameters of the model would increase its performance during training and testing by 10% and 3%, respectively, and the certainty of the models will increase in inference time. The Accuracy of semantic segmentation is 97.93% and F1-score is about 89%.