Domain Generalization in Deep Learning for Forest Health Monitoring Using Multispectral UAS Data

Abstract. Deep learning has significantly advanced forest health monitoring by enabling automated analysis of high-resolution aerial imagery. However, the generalization of these models across ecologically diverse regions remains limited due to domain shifts, which are variations in environmental conditions between training and testing locations. In this study, we propose a domain generalization (DG) framework that disentangles domain-invariant, task-relevant features from domain-specific environmental variations in multispectral UAS imagery. Our approach extends a baseline 2D convolutional neural network by incorporating parallel domain-specific and shared feature extractors, along with a domain classifier trained via adversarial learning. We evaluate the model using a leave-one-site-out strategy across three Finnish forest sites with diverse ecological characteristics. Results show that the DG model improves classification accuracy in previously unseen environments, with performance gains of up to 27% compared to the baseline. These findings highlight the effectiveness of feature disentanglement in enhancing the robustness and transferability of deep learning models for forest canopy health assessment, supporting more scalable and reliable forest monitoring solutions.