Evaluating Remote Sensing Foundation Model Embeddings for Cross-City Thematic Mapping of Eucalyptus Plantations in Guangxi, China

Accurate thematic mapping of plantation types remains challenging when labeled samples are unavailable in geographically unseen target regions and spectrally similar woody vegetation complicates class discrimination. This study evaluated whether remote sensing foundation model embeddings can improve cross-city eucalyptus plantation mapping. Annual AlphaEarth Foundations (AEF) and Tessera embeddings were integrated with 2024 Sentinel-2-derived spectral, index, texture, and topographic predictors, and seven feature configurations were assessed using XGBoost with 12,000 balanced reference samples from Liuzhou, Laibin, and Yulin, Guangxi, China. Performance was examined through 50 repeated within-city experiments and three leave-one-city-out evaluations in which all target-city labels were excluded from feature selection and model development. The three-source configuration achieved the highest mean within-city Matthews correlation coefficient (MCC = 0.911), exceeding the conventional-feature baseline by 0.138. In cross-city tests, MCC increased by 0.078, 0.093, and 0.130 for Liuzhou, Laibin, and Yulin, respectively. SHAP analysis showed that AEF and Tessera supplied most of the repeatedly retained discriminatory information, while selected spectral, textural, and terrain variables retained complementary value. In the three representative-site comparisons, the fusion model reduced scattered false positives and excessive patch expansion relative to the conventional-feature baseline. Within the 2024 Guangxi setting evaluated here, these results show that complementary foundation model representations can improve cross-city predictive performance for eucalyptus plantation mapping when the target city is excluded from model development.