Abstract: By recording reflectance spectral information of the ground on an aircraft or satellite platform, hyperspectral imagery (HSI), occupying dozens of or even hundreds of contiguous narrow bands, possesses abundant discriminative information for land use. Compared with visible light images and multispectral images, HSI can reveal subtle spectral characteristics, which contribute to a more accurate identification of the materials and classes of land covers. However, most existing methods overly focus on spectral knowledge while neglecting the potential morphological and spatial information within the hyperspectral input. In the classification of complex objects, the capture of morphological differences is much more necessary for searching out the class boundaries of fine-grained classes, e.g., forestry tree species. In this paper, the importance of morphological structure utilization is analyzed, and different feature extractors are designed. Specifically, focusing on fine-grained traits extraction, we propose a coarse-to-fine spatial information integration network, called MS-NET (morphological and spatial information based network), for tree species classification. The morphological operators are effectively embedded with the trainable structuring elements, which contributes to acquiring distinctive morphology representations, enhancing the classification accuracy. We evaluate the classification performance of the proposed method on two tree species datasets, and the results demonstrate that the proposed method provides superior performance when compared with other state-of-the-art classifiers.

Key words: deep learning, hyperspectral image, convolution neural network, forest tree species, morphology