Learning to disentangle forest health: A Conditional Variational Autoencoder approach to satellite time-series

Outbreaks by the European spruce bark beetle ( Ips typographus ) cause massive ecological and economic damage across Europe. Detecting these attacks early is a significant challenge because ground-truth data of which trees are attacked and when are often scarce or unavailable. This study presents a Conditional Variational Autoencoder (CVAE) designed to detect outbreaks by combining Sentinel-2 time-series data with static environmental features. Unlike traditional machine learning models that require thousands of labeled examples, the CVAE operates in a semi-supervised setting, meaning it does not require temporally aligned same-year labels. We use static environmental features to identify the susceptibility (risk levels) of specific forest areas. Then, by conditioning a Variational Autoencoder model on these risk levels, we train it to reconstruct forest dynamics for healthy and attacked situations. This allows us to effectively separate healthy and attacked areas within the model’s latent space. Our model was trained on Sentinel-2 data from 2018 and evaluated on data from 2018 and 2019. Our results demonstrate that the model effectively separates healthy and attacked pixels, highlighting its potential as a powerful tool for early detection of bark beetle outbreaks in data-scarce environments.