Satellite-Scale Eucalyptus Disease Recognition Using Semi-Supervised Domain Adaptation with Limited Target-Domain Labels

Remote sensing-based forest disease monitoring is important for forest health assessment, early warning and precision management. However, the deployment of deep semantic segmentation models across regions is constrained by domain shifts arising from differences in imaging conditions, forest backgrounds and disease distributions, as well as by the limited availability of annotated disease samples in target regions. This study performs patch-level binary semantic segmentation of eucalyptus disease stress using Sentinel-2 multispectral imagery under limited target-domain supervision. The objective is to identify and delineate the spatial extent of disease-stressed forest patches. To address these challenges, we propose a dual-branch semi-supervised domain adaptation method for remote sensing-based eucalyptus disease recognition under limited-label conditions. Separate source- and target-domain mapping layers first perform domain-specific feature mapping, after which a lightweight Attention U-Net extracts multiscale disease-related semantic features. A domain-fusion branch then learns transferable cross-domain semantics, whereas a domain-separation branch preserves discriminative region-specific information to improve disease delineation in the target domain. Unlabelled target-domain samples are incorporated through conditional adversarial domain alignment, while a small number of labelled target-domain samples impose class-conditional semantic constraints that enhance the consistency of same-class features across domains. The results show that the domain-fusion and separation branches learn complementary representations of cross-domain commonality and regional variation. On the eucalyptus disease dataset, the proposed method achieved a F1fg score of 0.8683, exceeding the Source-only baseline by 0.2962. These findings demonstrate that the proposed method can effectively exploit abundant unlabelled imagery when disease annotations are scarce, providing a practical approach for satellite-scale delineation, cross-region monitoring and precision management of eucalyptus disease.