Incorporating high-resolution spatiotemporal data for forecasting LULC trends in the Chesapeake Bay

Abstract

This study investigates a new approach to predicting land cover transitions in the Chesapeake Bay watershed, a region highly vulnerable to pollution due to its high development activity. We built a convolutional neural network model that ingests satellite time-series data from the Landsat satellite. The model was trained to predict land cover transitions into 4 classes: Residential, Commercial, No Change, and Non-classed change. Performance was evaluated on the state of Delaware using Mean Intersection over Union (Mean IoU) and categorical accuracy. Preliminary results showed the model achieving a mean IoU of 0.55. While the model demonstrated a strong capacity to learn discriminative features, further work needs to be done to define the complex, spatial relationships between pixels and study areas.