Land Surface Water (LSW) such as streams is vital for human and ecological systems. Efficient stream type classification is crucial for sustainable resource management. However, it traditionally relies on costly field surveys. This paper proposes a framework that leverages diverse geospatial features and domain knowledge to curate representative datasets and design hierarchical classification models for stream type identification. Empirical results demonstrate that our framework outperforms standard multi-class models, offering great promise for exploring data representation and structure in environmental classification tasks.