Flood classification using satellite images

Completed as part of QMSS5055 GR - Practicum in Large-Scale Data Analysis and Processing with Professor Charlie Riemann. The client was KPMG Digital Lighthouse (New York Office).

This project that aims to use deep learning image classification models to predict flood events and mitigate losses. The final CNN model achieved an accuracy of 81.06%. The data for the project was collected from satellite images provided by Planet and flood event records from the United States National Oceanic and Atmospheric Administration (NOAA).

The data were preprocessed to create a balanced structure. The model was improved through standard hyperparameter tuning, and techniques such as pseudo labelling and confounder control. The project’s final outcome is a business application that measures the economic impact of flooding, with potential application to historically low-flood regions that may experience rapid increases in flooding activity in future years as a result of climate change related threats.

See the presentation slides here.