Urban pattern recognition from multi-spectral satellite images and flood prediction using machine learning models

Abstract

Bangladesh suffers from the cumulative effects of floods brought on by water flashing from nearby hills, the accumulation of the inflow of water from upstream catchments, and locally heavy rainfall made worse by drainage congestion because it is located in such a basin and is less than 5 meters above mean sea level. Additionally, the rapid landscape changes in river areas brought on by the strong water flow make them more vulnerable to flooding. The purpose of this study is to create a system of detection of flooding in a given area using satellite-collected Multi-Spectral satellite imagery and numerical data collected from Bangladesh Water Development Board. It can then forecast if a flood will occur in the area shortly based on the landscape presented and its current shape. Additionally, it can show the likelihood of a flood as well as whether there is a chance of one. Five classification algorithms, VGG19, GoogleLeNet, UNET, ResNet, and Inception which represent various machine learning concepts, have been chosen and implemented on a free and open-source basis on the image datasets and MLP was used on the numerical dataset and output of the models was feed forwarded to an FCNN model to detect the likelihood of a flood. The multi-spectral image datasets and numerical datasets used for this study’ s foundation date from 2015 to 2023.