MultiEarth 2022 Deforestation Challenge -- ForestGump

TL;DR

This paper presents a deep learning approach combining optical and SAR satellite imagery to accurately estimate deforestation in the Amazon, overcoming challenges of large area size and weather conditions.

Contribution

It introduces a novel method using a conventional UNet with comprehensive data processing of multi-source satellite data for deforestation estimation.

Findings

High accuracy deforestation estimation achieved

Effective use of Sentinel-1, Sentinel-2, and Landsat 8 data

Method applicable to novel queries

Abstract

The estimation of deforestation in the Amazon Forest is challenge task because of the vast size of the area and the difficulty of direct human access. However, it is a crucial problem in that deforestation results in serious environmental problems such as global climate change, reduced biodiversity, etc. In order to effectively solve the problems, satellite imagery would be a good alternative to estimate the deforestation of the Amazon. With a combination of optical images and Synthetic aperture radar (SAR) images, observation of such a massive area regardless of weather conditions become possible. In this paper, we present an accurate deforestation estimation method with conventional UNet and comprehensive data processing. The diverse channels of Sentinel-1, Sentinel-2 and Landsat 8 are carefully selected and utilized to train deep neural networks. With the proposed method, deforestation status for novel queries are successfully estimated with high accuracy.