Abstract: To solve the problem that a single-source observation is hard to balance between the spatial and spectral properties of natural resources, this paper proposes a data-fusing classification method based on deep learning. Using limited samples of multi-source data, the method accomplishes the detailed classification of vegetation coverage areas, completes the extraction of natural resource types and the investigation of vegetation quantity in a pilot area, the Haizhu National Wetland Park in Guangdong province. The results show that this method can effectively extract natural resource types and classify vegetation. Its accuracy of forest number detection is better than 87%, which significantly improves the fineness of vegetation classification, and explores the way of detailed investigation of natural resources for ownership confirmation and registration.

Key words: natural resources, wetland park, unified rights verification and registration, deep learning, ownership survey