Mapping Tropical Forest Degradation with Deep Learning and Planet NICFI Data
Ricardo Dalagnol,Fabien Hubert Wagner,Lenio Soares Galva,Daniel Braga,Fiona Osborn,Le Bienfaiteur Sagang,Polyanna da Conceica Bispo,Matthew Payne,Celso Silva Junior,Samuel Favrichon,Vinicius Silgueiro,Liana O. Anderson,Luiz Eduardo Oliveira e Cruz de Aragao,Rasmus Fensholt,Martin Brandt,Philipe Ciais,Sassan Saatchi
DOI: https://doi.org/10.1016/j.rse.2023.113798
IF: 13.5
2023-01-01
Remote Sensing of Environment
Abstract:Tropical rainforests from the Brazilian Amazon are frequently degraded by logging, fire, edge effects and minor unpaved roads. However, mapping the extent of degradation remains challenging because of the lack of frequent high-spatial resolution satellite observations, occlusion of understory disturbances, quick recovery of leafy vegetation, and limitations of conventional reflectance-based remote sensing techniques. Here, we introduce a new approach to map forest degradation caused by logging, fire, and road construction based on deep learning (DL), henceforth called DL-DEGRAD, using very high spatial (4.77 m) and bi-annual to monthly temporal resolution of the Planet NICFI imagery. We applied DL-DEGRAD model over forests of the state of Mato Grosso in Brazil to map forest degradation with attributions from 2016 to 2021 at six-month intervals. A total of 73,744 images (256 x 256 pixels in size) were visually interpreted and manually labeled with three semantic classes (logging, fire, and roads) to train/validate a U-Net model. We predicted the three classes over the study area for all dates, producing accumulated degradation maps biannually. Estimates of accuracy and areas of degradation were performed using a probability design-based stratified random sampling approach (n = 2678 samples) and compared it with existing operational data products at the state level. DL-DEGRAD performed significantly better than all other data products in mapping logging activities (F-1-score = 68.9) and forest fire (F-1-score = 75.6) when compared with the Brazil's national maps (SIMEX, DETER, MapBiomas Fire) and global products (UMD-GFC, TMF, FireCCI, FireGFL, GABAM, MCD64). Pixel-based spatial comparison of degradation areas showed the highest agreement with DETER and SIMEX as Brazil official data products derived from visual interpretation of Landsat imagery. The U-Net model applied to NICFI data performed as closely to a trained human delineation of logged and burned forests, suggesting the methodology can readily scale up the mapping and monitoring of degraded forests at national to regional scales. Over the state of Mato Grosso, the combined effects of logging and fire are degrading the remaining intact forests at an average rate of 8443 km(2) year(-1) from 2017 to 2021. In 2020, a record degradation area of 13,294 km(2) was estimated from DL-DEGRAD, which was two times the areas of deforestation.