Abstract:The primary factor in crop growth and one of the fundamental indicators used to monitor the wetness of fields is soil moisture. The relationship between the size of spectral information window of sampling points and soil moisture was mainly studied to solve the problem of soil moisture inversion error caused by spatial heterogeneity. UAV remote sensing technology was utilized to acquire multispectral orthophoto images during the corn filling and wheat seedling stages, under both sprinkler irrigation and border irrigation. Initially, the sliding window method was employed to extract 34 groups of spectral characteristic variables, capturing the average spectral information across various spatial scales. Subsequently, the optimal window size of spectral information at the sampling points was determined by using three machine learning models: extreme gradient Boost (XGBoost), support vector machine regression (SVR), and partial least squares regression (PLSR). Next, the feature variables extracted the 34 groups of spectral features were screened by using the Pearson correlation coefficient feature variable screening method (R) in conjunction with the XGBoost and SVR machine learning models. Subsequently, the feature variables that demonstrated sensitivity to soil water were selected. Lastly, the estimation of soil moisture was conducted. The results indicated that the optimal spectral information window for sampling points under sprinkler irrigation was smaller compared with that under border irrigation. Specifically, the optimal window size for sprinkler irrigation was ranged from 11×11 to 21×21, while for border irrigation, it was ranged from 15×15 to 29×29. The eigenvariable screening method, employing the Pearson correlation coefficient in combination with machine learning models, can significantly enhance the accuracy of soil moisture inversion. Among the five machine learning models (R_XGBoost, R_SVR, XGBoost, SVR, PLSR), the R_XGBoost model exhibited the highest accuracy in estimating soil moisture. The R_XGBoost model achieved R2 values of 0.80 and 0.83, and RMSE values of 1.27% and 0.98% under spray irrigation and border irrigation, respectively. Additionally, the R2 values were 0.76 and 0.79, and the RMSE values were 1.68% and 0.85%, respectively. The accuracy of the soil water inversion models was higher under border irrigation compared with that of sprinkler irrigation. The research result can serve as a valuable reference for information mining and soil moisture monitoring through the analysis of UAV multi-spectral images.
