Convolutional neural networks fusing spectral shape features with attentional mechanisms for accurate prediction of soluble solids content in apples

Abstract

The soluble solids content (SSC) of apples is a key factor for evaluating their flavor and texture. However, convolutional neural networks (CNNs) still encounter challenges in effectively extracting relevant features for accurate SSC prediction. This study integrated spectral shape feature (SSF) and the convolutional block attention mechanism (CBAM), to enhance CNN performance in predicting apple SSC. The optimal CNN parameters were determined to be a batch size of 20, an ‘adam’ optimizer with an exponentially decaying learning rate, and the ‘relu’ activation function. Comparative analysis revealed that the CNN model fusing SSF and CBAM (SSF-CBAM-CNN) outperformed models such as partial least squares regression (PLSR) and backpropagation neural networks (BPNN), with an increase in the determination coefficient (R²) by 11% and 8%, respectively. These findings demonstrate that integrating SSF with spectral features significantly enhances model accuracy, establishing SSF-CBAM-CNN as a reliable and high-performance solution for precise SSC detection in apples.