Hyperspectral crop classification is often challenged by substantial intra-class spectral variability, high inter-class similarity, and the scarcity of high-quality labeled samples. These issues frequently lead to insufficient feature fusion or excessive computational complexity in conventional classification methods. To address these problems, this study proposes MDPC-Net, a limited sample hyperspectral crop classification method that couples a multi-dimensional pyramid with a Transformer architecture. The model extracts crop features from spectral, spatial, and joint spectral–spatial dimensions to capture fine-grained characteristics. A feature reorganization strategy is further incorporated to effectively reduce dimensional redundancy, while the Transformer modules enhance global dependency modeling, thereby improving the discrimination of crop features in complex environments. Comparative experiments with six classical models on three datasets—Matiwan Village, WHU-HongHu, and WHU-LongKou—demonstrate that MDPC-Net achieves competitive accuracy with substantially lower computational complexity, effectively balancing the trade-off between classification performance and efficiency. The proposed approach provides a promising solution for fine-grained hyperspectral crop classification under limited sample conditions.
Mingchao Yang (Wed,) studied this question.