Mapping Vegetation Alliances Using Deep Learning and Multi-Source Remote Sensing Data

Mapping vegetation alliances is essential for understanding ecological patterns and supporting sustainable land management in arid and semi-arid areas. However, traditional remote sensing typically only distinguishes grassland boundaries or broad subclasses, failing to differentiate specific vegetation alliances. Furthermore, while traditional vegetation mapping relies heavily on field surveys, manual interpretation, and expert knowledge, this labor-intensive approach hinders efficient large-scale mapping. This study proposes an efficient method for vegetation mapping by integrating field survey data with multi-source and multi-temporal remote sensing variables using deep learning. A series of deep neural network models was designed to systematically characterize and leverage spectral signatures, climatic factors, and topographic habitat features for fine-grained classification of 32 vegetation alliances in Xinjiang, a typical arid to semi-arid region. The resulting vegetation map achieved an alliance-level classification accuracy of 0.5125 on an independent test set, with the five most dominant alliances: Stipa spp. desert steppe, Stipa spp. steppe, Poa spp. meadow, and Anabasis spp. desert, accounting for over 10.45% of the total area of Xinjiang. Compared with traditional approaches, this method significantly improves mapping efficiency and offers a scalable solution for large-area, updatable vegetation classification. The approach provides a valuable reference for ecological assessment and dynamic vegetation monitoring in arid and semi-arid regions.