Abstract:
To achieve real-time, continuous, and batch prediction of dry matter content in pre-harvest kiwifruits, to the researchers verified the feasibility of using hyperspectral technology of dry matter content prediction for pre-harvest kiwifruits in field environment. The study conducted an experiment of hyperspectral data collection for pre-harvest kiwifruit on September 2021, in the main kiwifruit production areas of Xifeng and Xiuwen county, Guizhou Province. The hyperspectral data of pre-harvest kiwifruit samples were obtained with a hyperspectral camera, and the spectral bands were set from 500 to 900 nm with a spectral resolution of 2 nm, which had 193 spectral bands totally. The raw data were subjected to whiteboard correction, ROI (Region of Interest) cropping, and multiple scattering correction to obtain the spectral reflectance curves of the sample fruits. The characteristic bands were selected according to the characteristics of the spectral curves and the spectral absorption characteristics of water and chlorophyll, and the sample fruits were divided into a training set and a test set. The multispectral index was used to convert the spectral reflectance of the sample fruits in training set into index values. Through analyzing the correlation between index values and dry matter content, the optimal index was determined, and its fitting formula was used as the prediction model of dry matter content. The coefficient of determination, root mean square error, absolute error, relative error, and relative error mean were used as test indicators to calculate the error situation and verify the prediction effect of the model using the test set of sample fruits. The results showed that the spectral curves of all samples were similar, with high dry matter content of fruits having low spectral reflectance and low dry matter content of fruits having high spectral reflectance. The identified characteristic bands were 671.14, 745.08, 753.32, 805.14 and 886.93 nm, and five multispectral indices (I1-I5) were constructed according to the characteristic bands. Among the five multispectral indices, the determination coefficients of of I3, I4 and I5 reached above 0.8, and the highest reached 0.88 (I4). The prediction model of dry matter content for pre-harvest kiwifruit was established using this best-fitting multispectral index. In terms of the evaluation on the test set, the maximum absolute error of the model was 1.31% and the maximum relative error was 8.23%; the minimum absolute error was 0.03% and the minimum relative error was 0.23%. In terms of the overall prediction effect of sample fruits in the test set, the mean relative error of the prediction results was 3.13%, and the root mean square error was 0.65%. Compared with a previous study on the prediction of dry matter content of pre-harvest mango fruit using hyperspectral technology, which coefficient of determination was 0.64, and root mean square error was 1.08%. The proposed model also showed good accuracy. The experiment proved that it is feasible to use hyperspectral technology for dry matter content prediction on pre-harvest kiwifruit.