Objectives Remote sensing method, based on spectral index, is widely used for the real-time monitoring of crop growth, however, the soil background at the early stage and the loss of sensitivity at the later stage of crop growth limit its accuracy. With the rapid development of artificial intelligence, using machine learning algorithms has become a widely accepted method to remove the defect. In this study, we compared the accuracy of potato chlorophyll content estimation by partial least squares and random forest algorithm methods.

Methods Field experiments with different nitrogen levels were conducted in the main potato producing areas of Inner Mongolia from 2019 to 2020. Spectral data and chlorophyll values were obtained at the key growth period of potatoes. Spectral index band optimization was conducted using the experimental data to find the optimal spectral index sensitive to chlorophyll. Total reflection spectral band and optimized spectral index were used as input variables to estimate potato chlorophyll content in random forest and partial least squares model. The experimental data were divided into 75% and 25% as modeling set and verification set, respectively, to compare the accuracy of the models and evaluate the models at the same time.

Results The optimized bands were mainly concentrated in the purple and green light ranges, Opt-NDVI based on the central band of 408 nm and 552 nm had the best optimization effect. The correlation between the spectral index and potato chlorophyll content was significantly improved through optimization, but still influenced by the growth period. The correlation was higher in the post-flowering period than in the pre-flowering period. Random forest and partial least squares were used for modeling with the total reflection spectral band and optimized spectral index, respectively. Compared with the optimized spectral index, both the machine learning models significantly improved the accuracies of potato chlorophyll estimation by using the optimized spectral index. Under the same variables, the stochastic forest model showed was better in modeling ability, the estimated values showed better linear relationship with the measured values, and the influence of the growth period was neglected.

Conclusions The prediction ability of spectral index is greatly affected by the growth period. The optimization of spectral index as the input variables improves the computational efficiency and prediction accuracy of random forest algorithm and partial least square method. The random forest algorithm based on optimized spectral indices is more accurate than the partial least square for estimation of potato chlorophyll, with better linear relationship between the predicted and measured values and negligible growth period impaction.