Selective logging is a major driver of tropical forest degradation and is estimated to span over 400 million hectares of tropical forest. Despite widely available forest monitoring tools that effectively map deforestation, accurate and scalable remote sensing methods to detect selective logging are less advanced. Previous efforts are largely unable to reliably detect the low-intensity selective logging (<10 m3 ha−1) that dominates across much of the Amazon rainforest, the world’s largest remaining stock of tropical timber. Utilising a unique training dataset of high-resolution uninhabited aerial vehicle imagery from logged forests across the Peruvian Amazon, we build random forest models trained to detect selective logging using freely available optical satellite images from Sentinel-2 and Landsat. We find the Sentinel-2 model to be highly accurate (F1 score: 0.88, kappa: 0.85, false detection rate: 6.3%), outperforming the Landsat model (F1 score: 0.77, kappa: 0.74, false detection rate: 21.7%). Both models accurately detected 3- to 20-fold more selective logging activity in our validation data than widely available forest monitoring tools (TMF, GLAD-S2, RADD). We demonstrate novel uses for these logging-detection models in the monitoring of legal timber harvesting inside forest concessions and illegal harvesting of wood inside Protected Areas. These results have the potential to transform our understanding of low-intensity, logging-induced forest degradation at broad scales, demonstrating the clear potential of remote sensing methods to effectively monitor both legal and illegal selective logging in tropical forests.

Read original article

Read original article