Mapping Tropical Forest Tree Cover and Deforestation with NICFI Planet Imagery and Deep Learning
Abstract
Deforestation of tropical forests represents about 20% of total carbon emissions to the atmosphere per year, while reducing the extent of intact forests and causing habitat fragmentation and loss of biodiversity. Monitoring changes of tree cover for rapid assessment of deforestation is considered the key component of any climate mitigation policies for reducing emissions from deforestation and restoration of forests for biodiversity and carbon sequestration.
Here, we map tropical tree cover and deforestation, by using high resolution (< 5 m) satellite imagery from the Planet NICFI program over the state of Mato Grosso (MT) in Brazil. We develop a U-net deep learning model for the segmentation of Planet images acquired at RGB-NIR bands at semi-annual to monthly time series. The U-net convolutional network generated in this study to segment tree cover was trained with 6663 images of 256 x 256 pixels RGB-NIR images and their labelled masks. 80% were used for training and 20% for validation. Our results shows that the designed U-net model presents a high level of segmentation accuracy (> 98%) with an F1-Score of 0.982 on the validation sample. We applied the U-net model to develop a wall-to-wall map of tree cover for the entire state of Mato Grosso from 2015 to 2022 (85056 Planet images). The area of tree cover for the state was 556510.8 km² in 2015 (58.1 % of the MT State area) that reduced to 141598.5 km² (14.8 % of total area) in early 2022. After reaching a minimum tree cover in December 2016 with 6632.05 km2, the bi-annual deforestation area showed a slight increase between December 2016 and December 2019 due to policies for reducing deforestation in the state. A year after, the areas of tree cover almost doubled from 9944.5 km2 in December 2019 to 19817.8 km2 in December 2021. The high-resolution data product showed relatively consistent agreement with official deforestation map from Brazil (67.2%) but deviated significantly from forest cover change from University of Maryland (IMD), mainly due to large area of fire degradation in UMD data. Our result indicates that the deep-learning model with high-resolution imagery from Planet NICFI program and significantly improve mapping deforestation extent in tropics- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.B55B..05W