CCD-Conv1D: A deep learning based coherent change detection technique to monitor and forecast floods using Sentinel-1 images

Abstract

Floods are among the most common natural disasters affecting human lives and public amenities. In the North-Indian region, the situation is severe as floods continue to create havoc with flood fatalities and huge infrastructure damages every year. To mitigate this risk, flood monitoring based on detecting the changes in land cover and future predictions is required to be developed using Synthetic Aperture Radar (SAR) images. In this paper, a novel DL-based coherent change detection (CCD-Conv1D) model comprising a combination of coherent change detection technique, deep learning (DL) models based analysis, and flood forecasting implementation on the obtained change patterns, which pave the way to generate flood maps and identify the flooded areas has been developed. The proposed coherent change detection technique on Sentinel-1 images using image segmentation generated a log ratio image with statistics creating a changed band. An enhanced accuracy achieved in detecting changes from log-ratio-based temporal composition for Ayodhya and Basti cities shows positive threshold values of 2.96 and 2.01 during and after the crisis which is higher than 2.34 and 1.46 before and during the crisis respectively. The experimental outcomes demonstrated that the inundation concentrated mostly over the vegetation region of these cities. Additionally, the DL-based flood prediction performed through the Convolutional Neural Network (Conv1D) and Naïve Forecast (NF) model demonstrated that the positive changes for Ayodhya city were 31.4 and 31.8 and for Basti city were 30.40 and 35.04 respectively, depicting larger variation inferring that significant area is expected to be inundated. The outcomes from CCD-Conv1D based on the analysis of results, accuracy in change detection, and DL-based flood predictions confirmed that it is more reliable when compared with individual traditional approaches. In the future, more DL models can be explored for a wider insight and for comparative analysis of the outcomes from CCD-Conv1D implementation to develop an efficient flood monitoring and early warning system (FMEWS).