Deep learning methods based on remote sensing data can play a critical role in monitoring and quantifying deforestation globally. However, their quality depends on the availability of large annotated datasets. Domain adaptation is an emerging technique that addresses the scarcity of annotated training data by leveraging knowledge from application domains for which there are abundant labeled data. 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 a particular domain adaptation method is an open issue. Additionally, class imbalance is a significant problem for domain adaptation. The deforestation detection application is often characterized by a high level of imbalance, as only a minor portion of extensive forest areas are deforested within the monitored periods. This work proposes novel solutions for both of these issues. In order to forecast domain adaptation performance without target labeled samples to assess the adapted model accuracy, we propose a strategy to measure uncertainty in its predictions, gaining insights into its generalization capacity. Regarding class imbalance, we apply an unsupervised debiasing module that determines sampling probabilities for the selection of batches used in the training iterations, considering the distributions of samples across the whole training dataset. The module assigns higher sampling probabilities to underrepresented samples. To evaluate the proposed solutions, several experiments were carried out considering four distinct domains within the Amazon rainforest. The domains correspond to different geographical locations, characterized by different vegetation types and deforestation patterns. The experimental results demonstrate that integrating the debiasing technique into domain adaptation methods improved classification performance, and that the estimated uncertainty is a valuable indicator of the generalization ability of the adapted models.