Integrating Phenological Priors with Deep Spatio-Temporal Features for tree species mapping

Abstract. Mapping large-scale tree species distributions is essential for accurately estimating forest carbon storage. Previous studies have shown that Satellite Image Time Series (SITS) can be effective for classifying tree species. However, many of these studies rely heavily on manual feature engineering or overlook critical geoscientific and forestry knowledge. Such domain-specific insights are particularly important in Earth observation because the same species can exhibit diverse spatio-temporal behaviors across different regions, leading to lower accuracy and limited model robustness. In this work, we propose a novel model, PTSViT, which integrates phenological information with deep spatio-temporal features to address these limitations. Our model’s loss function incorporates phenological priors, utilizing ground-based phenological data and tree species labels as supervisory signals to guide the learning of spatio-temporal encoders. We evaluate PTSViT on a newly created dataset, GXData, which includes 11 major tree species in GuangXi. Our model surpasses previous approaches across all evaluation metrics, demonstrating the value of integrating prior knowledge for automated, accurate tree species mapping.