Estimation of Chlorophyll Content in Potato Leaves Based on Machine Learning

To improve the accuracy of the photo chlorophyll content estimation model, the remote sensing images of different growth stages of potatoes under control and drought treatments were obtained using a multi-spectral camera on a UAV platform. Thirteen vegetation indices were selected as input variables of the chlorophyll content inversion model, and the estimation model of potato chlorophyll content was constructed by using multiple linear regression (MLR), support vector regression (SVR), random forest regression (RFR) and decision tree regression (DTR). Correlation analysis between vegetation index and chlorophyll content showed that at the tuber formation stage of the control treatment, the absolute values of correlation coefficients between CIre, GNDVI, NDVIre, NDWI, GRVI, LCI and chlorophyll content were above 0.5, and their were significant (p<0.05) or highly significant (p<0.01) correlations. In other growth stages of potato, the absolute values of correlation coefficients between 13 vegetation indexes and chlorophyll content were all above 0.5, which was a highly significant correlation (p<0.001). In addition, the accuracy of MLR, SVR, RFR and DTR models were compared. The results showed that the SVR model has the best prediction effects in the tuber formation stage, tuber expansion stage and starch accumulation stage of the control treatment. The control treatment's R-2 and RMSE were 0.89 and 2.11 in the tuber formation stage, 0.59 and 4.03 in the tuber expansion stage, and 0.80 and 3.18 in the starch accumulation stage. The RFR model produces the best prediction effects in the tuber formation, tuber expansion, and starch accumulation stages of the drought treatment. The outcomes of R-2 and RMSE on drought treatment were 0.90 and 1.57 in the tuber formation stage, 0.87 and 2.16 in the tuber expansion stage, and 0.63 and 3.01 in the starch accumulation stage. This study presents a new approach for monitoring the chlorophyll content of potatoes, and a corresponding estimating model can be selected based on the specific potato growth stage and different experimental treatments in future.