Post-Disaster Damage Assessment in Urban Infrastructure using Computer Vision Techniques

Abstract

Flood is growing problem in countries like Pakistan where thousands of people lose their lives and properties. Various disaster events are affecting people worldwide, especially in South Asia country like Pakistan where heavy rainfall have caused major fatalities and danger to infrastructure across the country since June 2022. Due to heavy rain fall and flooding approximately 33 million people have been impacted, it also include 7.9 million people who have been displayed, around 598,000 are currently residing in relief camps [1]. It is important that timely distribution of resources is carried out to help people. Satellite images cover a large area of disaster compared to other sources. Optical images are affected by different conditions related to the weather however Sentinel 1 images are not affected by weather condition and can work at night. In this study we used Sentinel data called Sen1Floods11 [2] which covers 11 flood events including United States, India, Pakistan, Sri Lanka, Bolivia, Paraguay, Ghana, Nigeria, Somalia, Spain, and the Mekong region covering area of 120,406 sq km [2]. It has two polarization band VV and VH. The polarization bands are signals reflected back from the surface of the Earth, for VV, the signal sent and received is vertical, while for VH, the signal sent is vertical and received horizontal. Semantic segmentation is where pixel level classification of satellite images occurs. Different semantic segmentation techniques for pixel level classification are applied to find flood area in image, named as U-Net [3], U-Net++ [3], MAnet and Segformer. The image pass through encoder-decoder architecture like U-Net++ [4] where segmentation map is created for satellite imagery with radar data consisting of 2 polarization bands. For U-Net++ it is tested with Resnet34, Resnet50 [5], densenet121 [6], Resnet50 with imagenet weights, efficientnet-b4 [7], efficientnet-b4 with imagenet weights. U-Net++ [4] with EfficientNet-B4 [7] as encoder choice, efficient model with compound scaling [7], while increasing depth and model layers improves accuracy but it may not always the case with lot of layers so this method is scaling in resolution, depth and width. It is also lightweight and efficient, has IOU of 54.64 and F1 score of 68.87 which is better compared to other models such as U-Net [3], MANet, Segformer. Using Segformer [8], did not give good results. After applying pre-processing methods results are further improved such as using lee filter [9] to solve the problem related to radar imagery such as speckle achieving highest score of IOU 56.72 and F1 score which is 70.54. vi