Predicting the spatial distribution of vegetation alliances with ecological knowledge under sample-limited conditions

Abstract Accurate spatial distribution of vegetation types is fundamental to understanding ecosystem structure, biodiversity patterns, and environmental responses. However, predicting the distribution of lower-level vegetation classification units such as alliances remains challenging due to limited and uneven sample availability, particularly for rare or narrow-niche communities. To address this issue, this study proposes the KnowSim method that integrates expert-defined ecological knowledge to evaluate environmental similarity between locations. Vegetation types are predicted by assigning each site the type of its most ecologically similar sample. The method was tested in two regions of the Tibetan Plateau (Bome and Zoige), which exhibit contrasting yet complementary climatic and topographic conditions, together representing the Plateau’s typical environmental settings. Results demonstrate that KnowSim consistently outperforms statistical methods (Random Forest, eXtreme Gradient Boosting, Support Vector Machine, and Logistic Regression) in both accuracy and type diversity. The improvement is particularly evident for alliances with sparse samples, achieving up to 24.6% higher accuracy in Zoige for alliances with fewer than five training samples. Moreover, the predicted vegetation maps better align with ecological gradients and field observations, demonstrating both ecological interpretability and predictive robustness under sample-limited conditions.