Domain Adaptation for Deforestation Detection in Remote Sensing: Addressing Class Imbalance and Performance Estimation

Abstract Deep learning approaches applied to remote sensing imagery have become essential for large-scale monitoring and quantification of global deforestation. Their performance, however, is strongly constrained by the limited availability of extensive annotated datasets. Domain adaptation has emerged as a promising strategy to mitigate this issue by transferring knowledge from related domains where labeled data are abundant. The success of domain adaptation depends, however, on the level of (dis)similarity between the source and target domains. Although there are some statistical techniques that may be used to measure relative discrepancies between domain data distributions, anticipating the outcome of domain adaptation is an open issue. Another major difficulty arises from class imbalance, which is particularly severe in deforestation detection, as only a minor portion of extensive forest areas are deforested within the monitored periods. This work proposes new solutions for both of these issues. To estimate domain adaptation performance without requiring labeled samples from the target domain, we present a method that quantifies prediction uncertainty, providing insight into the expected generalization ability of the adapted model. To mitigate class imbalance, we incorporate an unsupervised debiasing module that computes sampling probabilities for batch selection during training, based on the distribution of samples across the entire dataset. This module increases the sampling likelihood of underrepresented classes. We assess the proposed solutions through a set of experiments conducted across four distinct domains within the Amazon rainforest. These domains correspond to different geographic regions, each with unique vegetation characteristics and deforestation dynamics. The results demonstrate that the proposed uncertainty measure serves as a highly reliable indicator of the adapted models’ generalization performance, showing strong correlation with actual accuracy drops. Furthermore, integrating the debiasing strategy into domain adaptation frameworks promotes consistent predictions, helping to mitigate class imbalance effects in complex scenarios.