Improving UAV-based soil moisture measurement using optimal feature selections and background information removal

【Background and Objective】Topsoil water content is a critical factor influencing crop growth and yield, yet traditional measurement methods are often limited in efficiency and scalability. UAV-based remote sensing provides a promising alternative for rapid, high-resolution in situ measurements. This paper evaluates the factors that affect the accuracy of UAV-based soil water content inversion and identifies the optimal combinations of data types, features, and modelling approaches for improving the accuracy of the UAV-based method.【Method】The experiment was conducted in a maize field during its early growth stage, characterized by substantial variation in canopy coverage. UAV imageries and ground-truth measurements were collected simultaneously. A threshold method was applied to remove the influence of soil background information and calculate vegetation coverage. Spectral and texture features were extracted, and vegetation coverage was integrated into different data combination patterns. Three regression methods: random forest regression, ridge regression and partial least squares regression, were used to construct the inversion model for estimating topsoil water content; comparison of their performance was analyzed under different scenarios.【Result】① The effect of background information removal on model accuracy varied with regression method and the data extracted from sensors. In particular, inversion accuracy improved after soil background information removal for RGB sensors but decreased for TIR sensors. ② The combination of visible and thermal infrared data significantly improved model accuracy, providing richer information and improving robustness. ③ Incorporating vegetation coverage improved accuracy of the predicted topsoil water content both with and without background information removal. For datasets without background information removal, the R2 of the methods using the RGB+TIR+FVC pattern increased by 0.01 compared to that of using the RGB+TIR pattern. After background information removal, their R2 increased by 0.11.【Conclusion】Our results show that different data combinations and inclusion of vegetation coverage had varying effects on the accuracy of UAV-based method for topsoil water content estimation. We screened optimal combinations and methods to increase the accuracy of the method for estimating topsoil water content in the early maize growing stage.