Variable weighted convolutional neural network for the nitrogen content quantization of Masson pine seedling leaves with near-infrared spectroscopy.

Spectroscopy is a powerful non-destructive quantization tool. In this paper, the technology is used to predict the nitrogen content of Masson pine seedling leaves. Masson pine is widely planted in China, and its nitrogen content is an important index for evaluating the vigour of seedings. To establish a better prediction model, an improved 1D convolutional neural network architecture, named the variable weighted convolutional neural network (VWCNN), is proposed in this research. The new model can automatically force the network attention onto the important spectrum wavelengths and is able to improve the generalization ability of the basic 1D-CNN model. For 219 fresh Masson pine seedling leaves, it shows better results in the prediction accuracy and robustness compared to those derived from the traditional shallow prediction model and other CNN-based models. VWCNN can achieve a 0.984 R2 and 0.038 RMSE value in training dataset and 0.925 R2 value and 0.075 RMSE value in the test dataset. The proposed model was also tested on a public corn kernel dataset. For the dataset output, moisture, oil, protein, and starch, the new model also achieves state-of-the-art prediction results.