Abstract:

The inversion of the fine spatial distribution of soil water content in the seedling stage of winter wheat based on multispectral remote sensing of UAV can be used as a reference to plan irrigation in agricultural and improve irrigation efficiency. This study took the soil water content in the top layer (5cm) of winter wheat seedlings in Zhengzhou and Xinxiang as the inversion object, based on multispectral data from drones, selected the optimal spectral features, compared and validated the simulation results of random forest (RF) and gradient boosting (GB) machine learning models, and performed grid inversion of soil water content in the experimental area based on the optimal model. The results showed that the GB and RF models had a better inversion effects on the topsoil water content in Zhengzhou and Xinxiang during the wheat seedling stage, with R² and nRMSE ranging from 0.926 to 0.983 and 5.6% to 14.4%, respectively. The modeling accuracy of GB and RF based on the aggregated data from both sites was good, with R² and nRMSE of 0.902, 0.787 and 6.9%, 10.2%, respectively. The simulation results of the GB model were better than the RF model. The spatial accuracy of soil water inversion during the winter wheat seedling stage was 2cm, which better revealed the spatial heterogeneity of soil water in farmland. Both models performed well for different underlying surfaces and weather conditions and had high model generalization. The results can provide theoretical and technical support for accurate inversion of the water content of farmland soil using multi−spectral remote sensing from UAVs, which can benefit the development of precision agriculture and smart agriculture. 

Key words: Soil water content, UAV, Multispectral, Machine learning