Detecção de Desmatamento no Estado de Mato Grosso do Sul utilizando Segmentação Semântica em imagens bi-temporais dos satélites Landsat 8 e Sentinel 2

Abstract

The unauthorized suppression of native vegetation in the state of Mato Grosso do Sul has threatened the conservation of local biomes such as the Cerrado and the Atlantic Forest. In this context, it is relevant to emphasize that the Brazilian Constitution of 1988 established as duty of the State and Society the defense of a balanced environment, assigning to the Public Prosecutor's Office functions of action for environmental protection. For example, the Public Prosecutor's Office of the State of Mato Grosso do Sul, through the DNA Ambiental Program, has been monitoring and identifying points of deforestation without environmental authorization in its Geotechnologies Center (NUGEO), promoting actions and measures for the punishment and reparation of any damages caused. Currently, this mapping has been carried out through visual, non-automated analysis of images from various satellites, especially Landsat-8 and Sentinel-2 satellites, requiring considerable hours of work and specialized labor. However, with the evolution of Deep Learning techniques, new algorithms may be able to automate the multitemporal analysis process of satellite images, promoting agility, efficiency gains, and enabling the allocation of human resources to other services. This work aimed to present a proposal for automating the process of identifying deforestation using the deep neural networks DeepLabv3+, U-Net, and Multi-Scale Attention, for semantic segmentation in bi-temporal images of satellite, providing a trained artificial intelligence model capable of mapping deforested areas in any scene of the Cerrado and Atlantic Forest biomes. For this purpose, an appropriate dataset was sought. The first dataset was generated from a shapefile with predefined deforested areas polygons, associated with the manual clipping of two Landsat-8 satellite scenes, resulting in a dataset with few samples and much noise. The second dataset was obtained from a new deforestation shapefile and Sentinel-2 satellite scenes, with downloads and clippings performed automatically using a framework developed with APIs and cloud infrastructure from the Planetary Computer project, associated with specialized curation, producing a dataset with a larger number of samples and reduced noise. Notably, this dataset yielded the best results in preliminary training using the U-net neural network, established as the baseline, and subsequently adopted for training other evaluated artificial neural networks in this study. Additionally, an experiment was conducted to address data imbalance issues, employing various loss functions. Therefore, the main contributions of this dissertation are: (1) a labeled and curated public deforestation dataset; (2) an experimental evaluation using the U-Net, DeepLabv3+, and Multi-Scale Attention neural networks for semantic segmentation in bi-temporal Sentinel-2 satellite images; and (3) an experimental evaluation of the loss functions Binary-Cross-Entropy, Weighted Binary-Cross-Entropy, and Focal Loss. The most favorable outcomes were achieved with the Multi-Scale Attention for Semantic Segmentation architecture, utilizing the Weighted Binary Cross-Entropy loss function. Finally, the most effective model has been made accessible to the MPMS for deployment, implementation of validation procedures, and further refinement of samples and model evolution.