Evaluation of peanut kernel oil content using 3D-WPCA-CNN based on hyperspectral imaging

The rapid, non-destructive, and precise prediction of oil content in peanut seeds is of great significance for accelerating the breeding process and meeting the demands of the food industry. This study proposes a novel 3D weighted pooling and channel attention convolutional neural network (3D-WPCA-CNN) model to enhance the accuracy of oil content prediction. The model integrates two key components: a weighted average pooling layer and a channel attention mechanism, both designed to optimize feature extraction and information weighting in high-dimensional data. The weighted average pooling layer is developed to enhance the extraction of critical features while reducing interference from irrelevant or noisy information. Unlike conventional pooling methods, such as max pooling or average pooling, which treat all spatial features equally, this layer introduces learnable weight matrices that dynamically adjust based on feature importance. During the training process, these weights are continuously optimized, allowing the model to highlight the most informative spatial and spectral features while suppressing less relevant ones. The pooling operation is implemented using weighted 3D convolution, where each feature channel is assigned an independent weight matrix, ensuring that different spectral channels are processed distinctly. Additionally, the model adopts grouped convolution to maintain independence among channels and improve computational efficiency. The learnable weight parameters are initialized randomly and progressively fine-tuned through backpropagation, enabling the network to adaptively focus on key spectral and spatial information. The channel attention mechanism further enhances the model's feature selection capability by dynamically adjusting the importance of different channels in the hyperspectral image. Traditional convolutional neural networks (CNNs) treat all channels as equally important, which may not be optimal for hyperspectral data, where certain spectral bands carry more relevant information for predicting oil content. To address this, the proposed mechanism first applies global average pooling across the spatial dimensions to extract a channel-wise feature descriptor, summarizing the overall contribution of each channel. This descriptor is then processed through a lightweight fully connected network, which learns to assign importance weights to each channel. The network includes a hidden layer with a ReLU activation function, followed by an output layer with a sigmoid activation function, ensuring that the learned weights are normalized. The attention weights are then applied to reweight the input feature channels, emphasizing high-contribution channels while suppressing redundant or irrelevant ones. This process enables the model to refine its feature representation dynamically, improving its predictive accuracy. To validate the superiority of the proposed approach, the model was compared against a convolutional neural network (CNN) based on mean spectral features. The proposed 3D-CNN model demonstrated higher predictive accuracy (R² = 0.809 2, RMSE = 1.713 1%) compared to the model based on average spectra (R²= 0.698 3, RMSE = 2.183 7%), 3D Average pooling convolutional neural network without Channel Attention（R²=0.721 7, RMSE=2.069 3%）and the 3D Weighted pooling convolutional neural network without Channel Attention (R² = 0.738 4, RMSE = 2.005 9%). Furthermore, the RPD of the 3D CNN model was 2.338 9, indicating its reliable performance in predicting oil content. This study provides a novel high-dimensional modeling approach for hyperspectral image-based crop seed composition prediction, offering valuable practical implications for the precise assessment of quality traits in peanuts and other agricultural crops.