Post-harvest hyperspectral inspection of maize ear mechanical damage with a dual-attention 3D–2D network

Accurate recognition of mechanical damage on maize ears is essential for post-harvest quality control, yet high-dimensional hyperspectral image (HSI) cubes make efficient and reliable classification challenging. We compile a maize-ear HSI dataset with seven annotated categories and present a lightweight spectral–spatial network that couples a multiscale 3D convolutional module with a compact 2D refinement head. Triplet Attention(TA) enhances directional cross-dimensional interactions, while Double Attention(DA) aggregates global context, yielding a parameter-efficient model that balances accuracy with efficiency. Evaluated against representative convolutional, transformer, 3D-CmT, and YOLO-HSI baselines, the approach attains high accuracy with low-latency inference and a moderate memory footprint. Ablation studies show complementary gains from the 3D convolutional module and the two attention modules, and qualitative maps and class-wise heatmaps indicate consistent improvements across categories, including fine-grained damage types. Given its spectral–spatial design and computational efficiency, the framework could be extended to other crops and sensor configurations in smart agricultural inspection, and it can be integrated into workflows that require reliable, low-latency HSI analysis.