Satellite Image Segmentation of Gold Exploration Areas in the Amazon Rainforest Using U-Net

Abstract

Gold exploration activity in the Amazon rainforest area has been increasing recently. Explorations of this nature have direct consequences on fauna, flora, and the lives of indigenous people living around these areas. In an attempt to map and detect land usage, among other activities using satellite images, several approaches of machine learning and artificial intelligence have been explored, both classical and modern. This paper develops a method for segmentation of gold exploration areas using U-Net, on high-resolution satellite images, estimating the most appropriate loss, optimization function, and training batch size for the task. The results achieved for segmentation obtained the values of 91.29% for precision, 74.64% recall, 76.60% f1-score, and 98.34% accuracy.