Sensitivity Analysis of Image Augmentation Methods to Improve Flooded Area Detection Performance

Abstract

Recently, detecting flooded areas in CCTV images was performed based on semantic segmentation models (e.g., U-Net, FCN, etc.). However, these flooded area detection techniques are based on large-scale manually annotated images, which consume manpower and time. Image augmentation is one of the ways to overcome the limitations mentioned above. Some previous studies have used image augmentation to improve the performance of flooded area detection by combining two or more methods. However, there has been no study quantifying which augmentation methods are reasonable. This study aims to verify which image augmentation method is reasonable to improve the performance of urban flooded area detection techniques. First, this study develops a flood area detection technology corresponding to training images augmented with five different methods (Brightness, Blur, Contrast, Rotation, Crop). Subsequently, the performance changes for each technique were quantified, and characteristics related to the performance variations of each method were examined.