Detection of soluble solids content in mold-infected Hami melons based on hyperspectral image fusion spectral-texture using 3DCNN-TexNet-Att

Postharvest decay caused by pathogenic fungi seriously affects the quality and economic value of Hami melons, yet early infections lack visible symptoms, making timely detection challenging. Soluble solids content (SSC), a key quality indicator, may vary under fungal infection. This study proposes a multidimensional feature fusion method integrating hyperspectral data, texture features, and deep learning models to predict SSC in Hami melons and to evaluate the impact of Aspergillus infection on SSC. During the period of pathogen infection, the SSC of Hami melons shows a continuous downward trend. The damage caused by pathogenic bacteria promotes the degradation of sugar and organic acids, thereby affecting the quality of the fruit. The key spectral bands associated with SSC were identified through two-dimensional correlation spectroscopy (2D-COS) analysis of hyperspectral data. Concurrently, texture features were extracted via image processing, selecting those that demonstrated a pearson correlation coefficient > 0.2 with SSC. Subsequently, a three-dimensional convolutional neural network (3DCNN) was employed to synergistically analyze the fused spectral-textural features. Experimental results demonstrated that the lightweight 3DCNN-MobileNetV2 model exhibited superior capability in handling high-dimensional and complex features compared to traditional machine learning models (PLSR and SVR). Furthermore, the multidimensional data fusion strategy improved SSC prediction accuracy by 5.74 % over single-spectral models. The 3DCNN-TexNet-Att model, enhanced with a self-attention mechanism, demonstrated optimal SSC prediction performance, achieving an R² of 0.911, RMSE of 0.541, and RPD of 3.345. Research shows that this method effectively detects SSC in mold-infected Hami melon and provides an effective tool for quality monitoring during postharvest transportation and storage.