Flood Identification Model Design with Deep Learning

Abstract

Deep learning has been widely used in various areas, such as detecting materials, or estimating natural disasters. Especially, generative adversarial network (GAN), which is one of the deep learning models, is enhanced to CycleGAN for generation and discrimination of images even with unpaired datasets. In this paper, we design a model to generate real-like fake flood models, and we confirm that we distinguish between real and fake images by mixing them with real images. Based on this metric, a deep learning model is designed, and a dataset is generated using CycleGAN. We further perform data augmentation to assist in the dataset generation process. The program used to design the model is Python, which uses data from Sentinel-l. Input data is a collection of data from floods during 2019 in West Africa, Southeast Africa, Middle East Asia, and Australia. To determine the accuracy of the generated data, we compare the image using several indicators. The used indicators judge the accuracy and similarity of images such as SSIM and MSE, and PSNR. SSIM, MSE, and PSNR averaged 0.7192, 2014.0066, and 15.5745, respectively. Comparing images with these indicators, we confirm that the actual flood image and the generated flood image are similar. And using generated images, we use different deep learning model, to confirm how similar the real flood image is to the flood image produced in this paper.