Prediction model of rice protein content based on hyperspectral image and deep feature

In order to fully explore useful spectral and image information of hyperspectral images, a method of extracting the deep feature of hyperspectral images based on stacked auto-encoder (SAE) was investigated in this study, then the extracted deep feature was used to establish support vector regression (SVR) model to realize non-destructive detection of protein content in rice. Firstly, 420 rice samples (70 in each group) were placed in 6 groups at different storage time (0, 24, 48, 72, 96 and 120 h) under the condition of high temperature (45℃) and high humidity (95% relative humidity). Secondly, the hyperspectral images (400-1 000 nm) of 6 group rice samples with different protein content were collected by the hyperspectral image acquisition system. After hyperspectral images collection, the Kjeldahl method for determination of nitrogen was used to detect the protein content of rice samples. According to the chemical detection results, the protein content of the rice samples decreased from 7.78 to 7.62 g/(100 g) with the increase of storage time. Thirdly, in order to separate samples from background, the threshold segmentation method was used to obtain the sample mask in ENVI software, then the mask was applied to the hyperspectral image containing only rice samples. The rice sample area was chosen as region of interest (ROI), and the average spectral data and image information of ROI was extracted respectively. Finally, SAE was used to extract deep feature of spectral data, image information and fusion data. 1) For spectral information, savitzky-golay (SG) was used to pre-process the obtained spectral data, and SAE was used to extract the deep feature, then SVR was used to establish the prediction model, and grid search method was used to optimize the kernel parameter g and penalty factor c in SVR. The results showed that the optimal scale of SAE was 478-400-290-70, and RC 2 (determinant coefficient of calibration set), RP 2 (determinant coefficient of prediction set), RMSEC (root mean square error of calibration set), RMSEP (root mean square error of prediction set) using deep feature extracted by SAE were 0.976 2, 0.939 2, 0.068 6 g/(100 g), 0.115 3 g/(100 g), respectively. 2) For image information, the RGB images in each band were extracted first, the size was 100 pixels×100 pixels. There was much redundant information in the original images, so we unified them to the gray images with 28 pixels×28 pixels, then flattened and converted them to one-dimensional column vector, after deep feature extraction by SAE on the vector, the prediction results of the model built with SVR showed that the optimal scale of SAE was 784-700-480-30, and RC 2, RP 2, RMSEC, RMSEP modeled using deep feature extracted by SAE were 0.915 4, 0.821 0, 0.051 0 g/(100 g), 0.111 8 g/(100 g), respectively. 3) For the fusion data of spectral data and image information, the initial dimension was 1 262, the fusion data combined all the redundant information of spectral and image, so dimension reduction becomes critical. After feature extraction by SAE, the dimension was reduced to the low level and the efficiency of network training was improved. The optimal scale of SAE was 1262-550-450-30, and RC 2, RP 2, RMSEC, RMSEP of the model build with deep feature extracted by SAE were 0.971 0, 0.964 4, 0.077 2 g/(100 g), 0.085 1 g/(100 g), respectively. Compared with using spectral data or image information alone, the prediction effect of the model build with the fusion data was improved obviously. To summarize, the method in the paper fully fused the spectral data and image information of hyperspectral image, and then deep feature extracted by SAE improved the prediction accuracy of the model established by SVR effectively, which provides a theoretical basis for non-destructive detection of protein content in rice. © 2019, Editorial Department of the Transactions of the Chinese Society of Agricultural Engineering. All right reserved.