Abstract: ：Heavy metal pollution in granular feed is a major, threat for the safety of animal-derived food and the balance of the ecological environment. This paper aims to solve the key problems in the near-infrared spectroscopy inversion analysis of the contents of heavy metals（copper and zinc）in granular feed, especially to improve the learning ability of deep learning models for the long-range correlations in spectral data. A parallel input-based dilated convolutional neural network（PIN-DCNN）model was proposed. This model processed data of different bands through a parallel input structure and introduces dilated convolutions to expand the receptive field, so as to enhance the ability to capture long-range features. The Savitzky-Golay first derivative smoothing method and Pareto Scaling were used to preprocess the spectral data to ensure the data quality. The preprocessed near-infrared spectroscopy dataset was used to train and test the model. The PIN-DCNN model was employed for the inversion prediction of the contents of copper and zinc. With the results that the Determination Coefficients（R²）reached 0.94 and 0.94 respectively, and the Root Mean Square Errors（RMSE）were 280 mg/kg and 410 mg/kg respectively. Compared with PLSR, SVR, BPNN, and 1D-CNN, PIN-DCNN shows significant advantages in all evaluation indexes. Through the strategies of parallel input and dilated convolution, the model in this paper effectively improves the feature extraction ability of spectral data, especially the learning ability for long-range correlations. The model performs remarkably well in the near-infrared spectroscopy inversion analysis of the heavy metal contents in granular feed, so it provides new ideas for the effective extraction of spectral features and the prediction and inversion of heavy metal elements.

Key words: granular feed, near-infrared spectroscopy inversion, parallel input architecture, dilated convolution