Detecting Mining in the Amazon Using Computer Vision and Convolutional Neural Networks

Abstract

The expansion of illegal mining in the Brazilian Amazon poses a major threat to biodiversity and local communities, demanding agile and scalable monitoring mechanisms. This work proposes the development of a computer vision model based on Convolutional Neural Networks (CNNs) for the detection of mining fronts using satellite imagery. The methodology employed the Google Earth Engine platform for data collection and pre-processing, resulting in a balanced dataset containing over 111,000 image chips, labeled and curated to mitigate cloud interference. To validate the approach, a baseline was established using the Random Forest algorithm, which reached an accuracy ceiling of 72% and an Area Under the Curve (AUC) of 0.80, highlighting the limitations of pixel-based statistical methods. In contrast, the implementation of the EfficientNet-B0 architecture, combined with Transfer Learning and Fine-Tuning techniques, significantly outperformed traditional models, achieving an accuracy of 85.92% and an AUC of 0.9371. The results demonstrate that Deep Learning can effectively extract complex spatial features, distinguishing mining scars from other ambiguous visual targets. As a contribution to the scientific community and environmental enforcement efforts, the entire source code and the dataset creation methodology have been made available in an open repository, promoting reproducibility and further research on Amazon biome monitoring.