Flood Extent Mapping using Computer Vision and Satellite Imagery

Challenge Background

Floods are among the most destructive extreme weather events which cause problems for inhabitants, crops, and vegetation. Recently, a third of Pakistan is underwater amid its worst floods in history. More than 33 million people have been affected and 1 million homes have been damaged or destroyed, more than $10 billion in damages to infrastructure, homes, and farms. The consequences of floods vary greatly depending on the location and extent of flooding. Hence, accurate and timely mapping of flood extent to ascertain damages is critical and essential for relief activities. In this context, remote sensing data has repeatedly shown its interest and usefulness during the various phases of the flood management process by providing an overview of the situation on the ground without direct contact with the flooded area and by allowing the decision-makers to follow the water extent during the disaster. Hence the interest in developing new techniques to precisely delimit the extent of flooded areas based on high-resolution satellite images has become of immense importance. This project will aim to train a convolutional neural network (CNN) architecture for flood segmentation to distinguish water and land areas. Also, we will use the segmentation model on satellite images of areas before and during the flood image. Hence, we will generate a flood extent map of the flood-affected area by comparing the segmentation results on a pixel-to-pixel basis. This would be used to create a dashboard to visualize the impact of the current flood in Pakistan, which has been occurring since June 14, 2022, by generating a time series of mapping flood extent throughout Pakistan.

The Problem

It is very critical to acquire timely information about which areas are flooded, so that disaster and relief agencies can speed up emergency response for relief and rescue measures and flood victims can migrate towards non-flooded areas. It is very difficult for people to visit the area and manually assess the impact of the affected areas. Hence, satellite imagery will be used instead, to generate flood extent mapping techniques that can be used to process images quickly, providing near real-time flooding information for damage assessment and relief activities.

Goal of the Project

1. Analysing and comparing the performance of our trained segmentation model against baseline models. 2. Develop a model that can generate a flood extent map using temporally correlated satellite imagery. 3. Provide insights into past performances of flood extent mapping. 4. Dashboard showing real-time flood extent maps on the currently flooded areas of Pakistan and view time series of maps for past data 5. Allow for newly found visualizations to assess the impacts of the currently flooded areas of Pakistan and track water recession over time 6. Show extent maps generated with the segmentation model on the dashboard and possibilities of improvement with further research.

Project Timeline

What you'll learn

1. Collection of Data. 2. Data Cleaning. 3. Data Analysis. 4. Exploring the Data. 5. Data Visualization. 6. Deep Learning of Flood Segmentation . 7. Flood Extent Map Generation 8. Visualizations using trained model on dashboard