Non-destructive prediction of soluble solids content in kiwifruit using hyperspectral imaging and a multi-scale channel-attention network

Soluble solids content is an important indicator of sweetness, ripeness, and edible quality in kiwifruit. Conventional refractometry requires destructive sample preparation and is unsuitable for rapid batch inspection, whereas many existing spectral modeling methods rely mainly on mean spectra and make limited use of the spatial information contained in hyperspectral images. Therefore, this study developed a hyperspectral imaging-based deep learning framework for the nondestructive prediction of soluble solids content in “Xuxiang” kiwifruit. A total of 400 mature fruits were collected, and near-infrared hyperspectral images were acquired over the spectral range of 900–1700 nm. Following image calibration, region-of-interest extraction, and principal component analysis, the reduced hyperspectral images were used as model inputs. A lightweight multi-scale channel-attention two-dimensional convolutional neural network was developed to extract spatial–spectral features. The influence of retaining different numbers of principal components was evaluated, and the proposed network was compared with partial least squares regression, support vector regression, ResNet50, and VGGNet. The results showed that the number of retained principal components affected prediction performance, with 15 components providing a suitable balance between information preservation and redundancy reduction. Under this setting, the proposed network achieved a testing coefficient of determination of 0.9714 ± 0.0054, a root mean square error of 0.3305 ± 0.0682 Brix, and a residual predictive deviation of 5.9695 ± 0.5221. These findings demonstrate the potential of hyperspectral imaging combined with multi-scale channel-attention learning for rapid and nondestructive kiwifruit quality evaluation.