Abstract
Determining the quantitative content of chlorophylls in plant leaves by their reflection spectra is an important task both in monitoring the state of natural and industrial phytocenoses, and in laboratory studies of normal and pathological processes during plant growth. The use of machine learning methods for these purposes is promising, since these methods allow inferring the relationships between input and output variables (prediction model), and in order to improve the quality of the prediction, a researcher may modify predictors and selects a set of method parameters. Here, we present the results of the implementation and evaluation of the random forest algorithm for predicting the total concentration of chlorophylls a and b from the reflection spectra of plant leaves in the visible and infrared wavelengths. We used the reflection spectra for 276 leaf samples from 39 plant species obtained from open sources. 181 samples were from the sycamore maple (Acer pseudoplatanus L.). The reflection spectrum represented wavelengths from 400 to 2500 nm with a step of 1 nm. The training set consisted of the 85 % of A. pseudoplatanus L. samples, and the performance was evaluated on the remaining 15 % samples of this species (validation sample). Six models based on the random forest algorithm with different predictors were evaluated. The selection of control parameters was performed by cross-checking on five partitions. For the first model, the intensity of the reflection spectra without any transformation was used. Based on the analysis of this model, the optimal ranges of wavelengths for the remaining five models were selected. The best results were obtained by models that used a two-point estimation of the derivative of the reflection spectrum in the visible wavelength range as input data. We compared one of these models (the two-point estimation of the derivative of the reflection spectrum in the range of 400-800 nm with a step of 1 nm) with the model by other authors (which is based on the functional dependence between two unknown parameters selected by the least squares method and two reflection coefficients, the choice of which is described in the article). The comparison of the results of predictions of the model based on the random forest algorithm with the model of other authors was carried out both on the validation sample of maple and on the sample from other plant species. In the first case, the predictions of the method based on a random forest had a lower estimate of the standard deviation. In the second case, the predictions of this method had a large error for small values of chlorophyll, while the third-party method had acceptable predictions. The article provides the analysis of the results, as well as recommendations for using this machine learning method to assess the quantitative content of chlorophylls in leaves.