Mapping tropical forest degradation with deep learning and Planet NICFI data

• DOI: 10.1016/j.rse.2023.113798
• Publication Date: 2023-09-19T23:30:22Z
• Authors: Ricardo Dalagnol; Fabien Hubert Wagner; Lênio Soares Galvão; Daniel Braga; Fiona Osborn; Le Bienfaiteur Sa...; Polyanna da Conce...; Matthew Payne; Celso Silva Junior; Samuel Favrichon; Vinicius Silgueiro; Liana O. Anderson; Luiz Eduardo Oliv...; Rasmus Fensholt; Martin Brandt; Philipe Ciais; Sassan Saatchi
• Keywords: Forest degradation; Understory; Deforestation

Tropical rainforests from the Brazilian Amazon are frequently degraded by logging, fire, edge effects and minor unpaved roads. However, mapping extent of degradation remains challenging because lack frequent high-spatial resolution satellite observations, occlusion understory disturbances, quick recovery leafy vegetation, limitations conventional reflectance-based remote sensing techniques. Here, we introduce a new approach to map forest caused road construction based on deep learning (DL), henceforth called DL-DEGRAD, using very high spatial (4.77 m) bi-annual monthly temporal Planet NICFI imagery. We applied DL-DEGRAD model over forests state Mato Grosso in Brazil with attributions 2016 2021 at six-month intervals. A total 73,744 images (256 × 256 pixels size) were visually interpreted manually labeled three semantic classes (logging, roads) train/validate U-Net model. predicted study area for all dates, producing accumulated maps biannually. Estimates accuracy areas performed probability design-based stratified random sampling (n = 2678 samples) compared it existing operational data products level. significantly better than other logging activities (F1-score 68.9) fire 75.6) when Brazil's national (SIMEX, DETER, MapBiomas Fire) global (UMD-GFC, TMF, FireCCI, FireGFL, GABAM, MCD64). Pixel-based comparison showed highest agreement DETER SIMEX as official derived visual interpretation Landsat The closely trained human delineation logged burned forests, suggesting methodology can readily scale up monitoring regional scales. Over Grosso, combined degrading remaining intact an average rate 8443 km2 year−1 2017 2021. In 2020, record 13,294 was estimated which two times deforestation.