ENHANCING DEFORESTATION DETECTION THROUGH DOMAIN ADAPTATION WITH UNCERTAINTY ESTIMATION

Deforestation monitoring is crucial for biodiversity, climate regulation, and environmental impact assessment. In Brazil’s Amazon, the PRODES project, run by INPE, tracks deforestation through satellite imagery but depends on costly and time-consuming human analysis. Moreover, high accuracy is crucial due to the significant repercussions involved. To help achieve this goal, this work aims to enhance deforestation detection for semantic segmentation task by integrating a deep learning model based on Deeplabv3+ with a domain adaptation method and uncertainty estimation, supported by human review, in a manually assisted approach. The model extends the Domain Adversarial Neural Networks (DANN) to help a classifier trained on one domain generalize to unlabeled images from another. In a second phase, we use an DeepLabv3+ instances ensemble to measure prediction uncertainty, flagging high-uncertainty areas for visual audit.
Experiments across three Amazon and Cerrado domains using Landsat-8 images show that this approach significantly improves detection accuracy.