Hybrid-KANet: a hyperspectral remote sensing crop classification method based on the Kolmogorov–Arnold network
Hyperspectral remote sensing crop classification is crucial in precision agriculture management. However, existing studies are usually difficult to adequately model the complex nonlinear feature distributions in hyperspectral data, which often limits the classification performance and affects the re...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Taylor & Francis Group
2025-08-01
|
| Series: | Geo-spatial Information Science |
| Subjects: | |
| Online Access: | https://www.tandfonline.com/doi/10.1080/10095020.2025.2538831 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Hyperspectral remote sensing crop classification is crucial in precision agriculture management. However, existing studies are usually difficult to adequately model the complex nonlinear feature distributions in hyperspectral data, which often limits the classification performance and affects the recognition accuracy. In order to enhance the model’s ability to represent and discriminate complex nonlinear boundaries in hyperspectral images without introducing an attentional mechanism, an improved HybridSN (Hybrid-KANet) model based on the Kolmogorov-Arnold network (KAN) is proposed in this study. The model introduces a 3D fast kernel-activated convolutional layer, replaces the traditional linear activation function with a radial basis function (RBF), and realizes the nonlinear feature representation by spline paths and basis function paths. To enhance the model’s ability to model high-dimensional nonlinear features, a KANLinear layer is integrated into the classifier in place of the traditional fully connected structure. By employing fitting using learnable B-spline basis functions, the model can adaptively adjust to local features and achieve fine-grained approximation of complex decision boundaries in the input space. Experiments are conducted on Indian Pines and WHU-Hi-LongKou hyperspectral remote sensing datasets. The results show that the model achieves overall classification accuracies of 99.22% and 99.87% on the two datasets, which are 1.71% and 0.11% better than HybridSN; the mean intersection and merger ratio (mIoU) is improved by 4.77% and 0.53%, and the Kappa coefficient is improved by 1.96% and 0.15%, respectively. The ablation experiments demonstrate the advantages of RBF kernel function in modeling complex nonlinear relationships by systematically comparing the differences in classification performance and boundary modeling ability of different kernel functions, which improves the classification accuracy and spatial consistency. In conclusion, the Hybrid-KANet model proposed in this study provides theoretical innovation for precision agriculture management and a new solution for hyperspectral remote sensing crop classification. |
|---|---|
| ISSN: | 1009-5020 1993-5153 |