Mapping tropical forest degradation with deep learning and Planet NICFI data

Background

Forest degradation, driven by logging, fire, and infrastructure expansion, represents a major yet under-detected source of carbon emissions in tropical forests. Unlike deforestation, degradation involves partial canopy loss and is difficult to capture using conventional remote sensing due to small-scale disturbances and rapid vegetation recovery. Existing global products often underestimate degradation or fail to attribute its causes. Advances in high-resolution satellite imagery and deep learning provide new opportunities to improve detection accuracy.

Goals and Methods

This study develops a deep learning-based approach (DL-DEGRAD) to map forest degradation in the Brazilian Amazon using high-resolution (4.77 meter) Planet NICFI imagery. A U-Net convolutional neural network is trained on 73,744 labeled image samples to classify degradation into logging, fire, and roads. This model produce biannual degradation maps from 2016 to 2021 and is validated using stratified random sampling. The methodology of this study leverages spatial patterns rather than pixel-level reflectance, improving detection of subtle disturbances.

Conclusions and Takeaways

This study shows that deep learning significantly improves degradation detection compared to existing operational products, particularly for logging and fire disturbances. In Mato Grosso, degradation rates exceed deforestation in some years, highlighting its growing importance in carbon emissions. The approach demonstrates strong agreement with human interpretation, indicating scalability for national monitoring systems. For practitioners, this method provides a powerful tool for REDD+ MRV, carbon markets, and forest governance. Future applications should focus on expanding geographic coverage and integrating attribution-based monitoring into policy frameworks.