Evaluation of Cotton Water Stress Estimation Using Multispectral Satellite Images Based on M5 Tree Model

Water stress occurs as a result of the imbalance between soil water in the root zone and plant water use, which necessitates determining the water stress index of the plant. Surface soil moisture is directly related to plant water content. Availability of satellite data has led to temporal and spatial resolution of field data and offers new opportunities for monitoring crop conditions. In this research, accurate and continuous monitoring of soil moisture content, as a representative of soil moisture stress, was done with field measurements of soil moisture, and comparison with multispectral data of Landsat 9 and Sentinel 2 satellite images. The relationship between plant indices, as an independent variable, and soil surface moisture, as a dependent variable, was studied using linear multivariate regression and M5 tree regression methods. Considering the non-linearity of the relationship between soil moisture and spectral reflectance, linear multivariate regression did not show satisfactory results with coefficient of determination (R2) of 0.46 and 0.34 for Landsat 9 and Sentinel 2 satellites, respectively, as well as the root mean square error (RMSE) equal to 0.043 and 0.052. However, M5 tree regression showed more acceptable results, such that by establishing 16 and 20 regression relationships for Landsat 9 and Sentinel 2 satellites, the soil moisture was estimated withR2 of 0.70 and 0.67 and RMSE of 0.033 and 0.038, respectively. The results showed that the estimation of soil moisture with methods based on machine learning, such as the M5 model, increases the accuracy of calculations. In the M5 decision tree regression, a high number of variables does not necessarily lead to an increase in the accuracy of soil moisture estimation, and a relationship with the highest accuracy was found in the low number of variables. Therefore, the relationship obtained at the field level can be used to evaluate soil water stress and determine irrigation time in agricultural lands on a large scale, without measuring soil data.