Método usando deep learning para processamento de alvos de uso e cobertura da terra em imagens landsat: estudo de caso da mineração no Brasil

Abstract

Deep learning techniques, specifically convolutional neural networks, have shown great potential in the analysis of remote sensing imagery for detecting changes in the Earth's surface. In this thesis, an automatic methodology based on a U-shaped CNN (U-Net) is proposed for the detection and mapping of mining areas in Brazil, differentiating between industrial mining and artisanal mining (alluvial). Mining, an extractive activity that removes the substrate to access layers rich in mineral sediments, has significant environmental and socioeconomic impacts, especially in the Amazon. Although economically relevant, there are still no automatic methods to map this activity continuously over long periods, distinguishing its different types. The proposed methodology was applied to annual cloud-free Landsat mosaics, covering a period of 37 years (1985-2022). The accuracy of the U-Net was spatially validated by specialists, obtaining an average of 99% overall accuracy, 91% producer accuracy, and 91% user accuracy. The results demonstrate a significant increase in the area occupied by mining, which grew about 10 times between 1985 and 2022, totaling 4500 km². Artisanal mining was the subtype of mining that grew the most proportionally, increasing from 218 to 2627 km². The proposed modified U-Net showed a 30% reduction in the number of trainable parameters compared to the original U-Net. It is concluded that the proposed methodology is effective for mapping mining activity, providing accurate and up-to-date data for environmental management and territorial planning.