Detection of External Defects in Seed Potatoes Using Spectral–Spatial Fusion of Hyperspectral Images and Deep Learning

To improve the accuracy of detecting external defects in seed potatoes and address the reliance of current hyperspectral imaging methods on single-dimensional data, this study proposes a multi-dimensional spectral–spatial information fusion approach via concatenation based on a one-dimensional convolutional neural network (1DCNN) within the framework of deep learning. Hyperspectral three-dimensional data were acquired for normal seed potatoes and for samples presenting six types of external defects—decay, mechanical damage, wormhole, common scab, black scurf, and frostbite—across a wavelength range of 935–1721 nm. From the hyperspectral images, one-dimensional spectral data and two-dimensional spatial data were extracted. The one-dimensional spectral data were preprocessed using six methods: Savitzky–Golay smoothing (SG), standard normal variate (SNV), multiplicative scatter correction (MSC), first derivative (FD), second derivative (SD), and orthogonal signal correction (OSC). Feature wavelengths were subsequently selected through the successive projections algorithm (SPA) and competitive adaptive reweighted sampling (CARS), serving as inputs for traditional machine learning models. Two-dimensional spatial data were first subjected to dimensionality reduction via principal component analysis (PCA). Texture features were then extracted from each principal component using the gray-level co-occurrence matrix (GLCM). Following normalization, all spatial texture data were fused with the preprocessed spectral data to form the inputs for the deep learning models Basic1DCNN and Stacked1DCNN. The results demonstrate that the fusion data with the Stacked1DCNN model yielded the best performance in identifying normal seed potatoes and six types of external defects. The overall accuracy, precision, recall, F1 score, and mean average precision reached 98.77%, 98.77%, 98.93%, 98.73%, and 99.66%, respectively, outperforming traditional machine learning approaches. Compared with the Stacked1DCNN model trained using spectral data alone, these metrics improved by 2.81%, 2.78%, 3.20%, 3.01%, and 1.11%. This study offers theoretical and technical insights into the development of automated sorting and non-destructive detection systems for seed potatoes.