Prediction of maize crop coefficient from UAV multisensor remote sensing using machine learning methods

In the upcoming irrigation management in agricultural production, accurate mapping of crop water consumption with a high spatial and temporal resolution at a farm scale is needed. In this study, we developed models for crop coefficients (Kc) estimation using unmanned aerial vehicle (UAV) remote sensing and machine learning (ML) techniques for irrigated maize in a semi-arid region in Northwest China. Kc values were calculated using a procedure given in FAO56 manual using field measurements. Multispectral vegetation indices (VIs), vegetation fraction (VF), thermal-based VIs, and texture information (TI) were derived from UAV-based multispectral, RGB, and thermal infrared imagery, respectively. These remotely sensed variables and their combinations were used to develop prediction models using six ML algorithms (linear regression-LR, polynomial regression-PR, exponential regression-ER, random forest regression-RFR, support vector regression-SVR, and deep neural network-DNN). Among these models, the RFR with the highest accuracy (R2 = 0.69, RMSE = 0.1019) was recommended to estimate maize Kc. The multispectral and thermal-based VIs and texture of the near-infrared band had greater contributions than RGB-based VF and TI in the Kc-RFR model under different irrigation treatments. Furthermore, the maize Kc-RFR prediction model had high accuracy in estimating cumulative evapotranspiration (R2 = 0.89, RMSE = 15.0 mm/stage) during different growth stages and daily soil water content (R2 = 0.85, RMSE = 0.0089 m3/m3) in the root zone. These results show that the integration of UAV remote sensing and ML provides a promising tool to help farmers make decisions using timely mapped crop water consumption, especially under water shortages or drought conditions.