Classification of Maize Seed Hyperspectral Images Based on Variable-Depth Convolutional Kernels

Yating Hu¹Hongchen Zhang^1,2*Changming Li²Qianfusu Su³Wei Wang¹

• ¹College of Information Technology, Jilin Agricultural University, Changchun, Hebei Province, China
• ²Engineering Technology R & D Center, Changchun Guanghua University, Changchun, China
• ³Institute of Plant Protection, Jilin Academy of Agricultural Sciences(Northeast Agricultural Research Center of China), Changchun, China

The classification of corn seeds plays a crucial role in the rational utilization of germplasm resources and the efficiency of seed selection and breeding. Compared with traditional manual classification methods, machine learning-based automated classification techniques significantly enhance both classification accuracy and efficiency. Due to its superior classification precision, robustness, and generalization capability, convolutional neural networks (CNNs) have become the dominant approach in automated corn seed classification. This study addresses the limitation of traditional hyperspectral data processing methods, which fail to effectively extract both spectral and textural features simultaneously. To overcome this challenge, we propose a variable-depth convolutional kernel structure (VD-CNN), which extracts continuous spectral feature information by adjusting the depth of the convolutional kernels while leveraging the adaptive nature of CNN convolution operations to extract texture information. In the experiment, eight common corn seed varieties were selected, with 100 seeds of each type. A four-layer CNN was designed to extract features from hyperspectral images, and 12 classification models were developed by varying the convolutional kernel depth. The experimental results indicate that when the convolutional kernel depth is set to 15, the proposed model achieves optimal performance, yielding a training accuracy of 0.9865 and a test accuracy of 0.9697. To further validate the generalizability of the model, comparative experiments were conducted using a publicly available rice dataset and corresponding classification models. The results demonstrate that the proposed model outperforms existing methods, achieving the highest classification accuracy and improving the best-performing baseline model on this dataset by 3.14%. These findings confirm that the variable-depth convolutional kernel structure effectively extracts representative features from hyperspectral data, providing a novel approach for hyperspectral data analysis.