Amazon forest cover change mapping based on semantic segmentation by U-Nets

Abstract

Deforestation remains a major concern with regard to climate change and the maintenance of biodiversity. Meanwhile, the development of new image processing techniques and the broad availability of high spatiotemporal resolution satellite imagery provide an unprecedented setup for the development of effective and scalable forest cover change monitoring systems. This is especially relevant in regions with large forested areas with high rates of deforestation, such as in the region of the Brazilian Amazon rainforest. In this context, existing forest cover change monitoring methods are based on a combination of visual inspection, spectral profiles, statistics, and machine learning techniques, which offer alternative backbones to deal with deforestation monitoring. Given the recent advances in the field of image processing by Fully Convolutional Neural Networks (FCNs), the objective of this study is to evaluate the performance of the U-Net architecture for the mapping of forest cover aimed at identifying deforestation polygons in multi-temporal satellite imagery. To this end, 10-m resolution imagery from the Sentinel-2 satellite covering portions of the Legal Amazon region were employed. The U-Net could identify and draw polygons of forest areas and forest fragments with high accuracy (0.9470), precision (0.9356), recall (0.9676), and F1-score (0.9513), thus outperforming largely applied and well-know supervised and unsupervised image classification methods. The results indicate and we further discuss that U-Nets have the potential to run as the backbone for efficient forest cover change monitoring initiatives and support the deployment of near real-time deforestation warning systems.