Melhorando a Detecção de Desmatamento por meio de Adaptação Multidomínio com Estimativa de Incerteza

Abstract

Deep Learning (DL) models have shown great potential in scientific research, particularly in Remote Sensing (RS) for monitoring natural resources, environmental changes, and land use. Semantic segmentation techniques enable land cover classification, change detection in multi-temporal images, object identification, and vegetation health assessment. However, their effectiveness relies on large labeled datasets, which are costly and time-consuming to obtain. Domain Adaptation (DA) techniques address this challenge by transferring knowledge from a labeled source domain to one or more unlabeled target domains. While most DA research focuses on a single target domain, multi-target and multi-source scenarios remain underexplored. This dissertation proposes a deep learning approach using Domain Adversarial Neural Networks (DANN) for deforestation detection in multi-domain settings. Additionally, an uncertainty estimation phase is introduced to guide human review in high-uncertainty areas, enhancing manual auditing. The model is evaluated on Landsat-8 images from the Amazon and Brazilian Cerrado. In multi-target experiments, a single source domain provides labeled data, while target domains remain unlabeled. In multi-source scenarios, multiple source domains train the model for a single target. Results show significant accuracy improvements over lower-bound baselines, as indicated by F1-score and mean Average Precision (mAP). Furthermore, uncertainty-based review enhances performance, reaching upper-bound baselines in certain biome combinations.