Biodiversity Change: A Spatiotemporal Machine Learning Approach to Detect Forest Canopy Height Changes across the U.S. West Coast

Monitoring forest structural changes is critical for understanding biodiversity dynamics and responding to natural disturbances such as wildfires. In this study, we present a machine learning framework that combines optical remote sensing data from HISTARFM Landsat product with spaceborne LiDAR measurements from the GEDI mission to generate forest canopy height maps and detect structural changes at 30-meter resolution across the U.S. West Coast. We developed ConvLSTM models trained using monthly composites of Landsat data and annual GEDI-derived RH98 canopy height metrics. Additionally, we explored two approaches for incorporating uncertainty information into ML modeling process. Our best-performing model achieved an RMSE of 7.536 and a Pearson’s r of 0.847. Using the trained model, we generated canopy height maps for 2019 and 2020 and performed change detection by differencing these maps. Evaluation against 135 wildfire events yielded a moderate ROC AUC of 0.65. We analyzed detection errors and outlined potential improvements from both data and modeling perspectives.