Generalized few-shot learning for crop hyperspectral image precise classification

• Published in: Computers and electronics in agriculture Vol. 227; p. 109498
• Main Authors: Yuan, Hao-tian, Huang, Ke-kun, Duan, Jie-li, Lai, Li-qian, Yu, Jia-xiang, Huang, Chao-wei, Yang, Zhou
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2024
• Subjects: Crop precise classification; Deep learning; Few-shot learning; Hyperspectral image classification; Deep learning; Few-shot learning; Crop precise classification; Hyperspectral image classification
• ISSN: 0168-1699
• DOI: 10.1016/j.compag.2024.109498

Hyperspectral remote sensing technology, with its advantage of acquiring a substantial amount of spectral information across different bands, has provided a robust tool for crop monitoring and management in the agricultural field. However, a prevalent challenge persists, namely the limited number of labels required for crop classification. We propose a new method named Generalized Few-Shot Learning (GFSL) to address the small-sample problem and get better classification of crops. The proposed GFSL first maps the embedding features extracted by a convolutional neural network to a Hilbert space by an implicit nonlinear mapping with a kernel trick. Then, GFSL maximizes the kernel similarity between each sample and its class mean as much as possible, and minimizes the kernel similarity between each sample and the means of other classes as much as possible at the same time. To give a more meaningful balance between intra-class similarity and inter-class similarity, GFSL defines the negative of intra-class similarity plus the logarithm of the sum of exponential functions of inter-class similarities as the loss function. We conducted experiments on three publicly available crop hyperspectral datasets: WHU-Hi-HanChuan, Salinas, and Indian Pines, and results show that the proposed approach exhibits an improvement in classification accuracy of 11.46%, 6.86%, and 14.49% on the three datasets, respectively, in comparison to some state-of-the-art methods, which demonstrates the superiority of the proposed method for crop hyperspectral image classification with limited training samples. The Python source code is available at https://github.com/kkcocoon/GFSL. •GFSL uses CNN-extracted features mapped to Hilbert space to solve HSI small-sample issues.•The loss function of GFSL balances intra-class similarity and inter-class similarity logarithmically.•GFSL improves classification accuracy over state-of-the-art methods for crop HSI.