BULC-D

Overview

A single satellite classification is noisy, and in cloudy tropical regions clear images are rare. BULC (Bayesian Updating of Land Cover) treats each classification as evidence and updates a posterior over land-cover classes at every time step. BULC-D extends it to detect deviation from expected values, such as new deforestation.

What I built

• Implemented the Bayesian update: each observation becomes an event, combined with the prior and normalized, with a probability floor so classes can recover.
• Output class probabilities with a confidence surface, not a single label.
• Combined BULC-D with the Kalman filter, and built a headless Earth Engine runner for batch experiments with parameter validation and dry runs.