An autoencoder-based model for forest disturbance detection using Landsat time series data

Monitoring and classifying disturbed forests can provide information support for not only sustainable forest management but also global carbon sequestration assessments. In this study, we propose an autoencoder-based model for forest disturbance detection, which considers disturbances as anomalous events in forest temporal trajectories. The autoencoder network is established and trained to reconstruct intact forest trajectories. Then, the disturbance detection threshold is derived by Tukey's method based on the reconstruction error of the intact forest trajectory. The assessment result shows that the model using the NBR time series performs better than the NDVI-based model, with an overall accuracy of 90.3%. The omission and commission errors of disturbed forest are 7% and 12%, respectively. Additionally, the trained NBR-based model is implemented in two test areas, with overall accuracies of 87.2% and 86.1%, indicating the robustness and scalability of the model. Moreover, comparing three common methods, the proposed model performs better, especially for intact forest detection accuracy. This study provides a novel and robust approach with acceptable accuracy for forest disturbance detection, enabling forest disturbance to be identified in regions with limited disturbance reference data.