DACNet: depth-aware convolutional network for corn hyperspectral image classification

Abstract With the rapid advancement of agricultural monitoring technologies, hyperspectral imaging has shown considerable potential in seed classification tasks. However, accurate classification of hyperspectral corn seed images remains a challenge due to the high dimensionality of the data, insufficient feature utilization, and limited classification performance. To address these issues, we introduce a depth-aware convolutional network (DACNet) designed for corn hyperspectral image classification. The network integrates multi-layer 3D and 2D convolutions to extract spectral, spatial, and textural features, with fully connected layers used for the final classification. Specifically, principal component analysis is applied to reduce spectral dimensionality while preserving essential information. Then, 3D convolutions are used to extract joint spectral–spatial features, which are further enhanced by multi-head attention mechanisms. These features are reshaped and processed through 2D convolutions to better capture spatial and textural details while keeping computational complexity manageable. Finally, fully connected layers are used to classify corn seed categories accurately. Experiments conducted on a hyperspectral corn seed image dataset and two publicly available remote sensing datasets validate the effectiveness and robustness of DACNet.