Research on Cotton Crop Coefficient Estimation and Inversion Model Based on UAV Remote Sensing and Soil Moisture Monitoring

Objective The crop coefficient is one of the fundamental parameters for calculating crop water requirements, and rapid acquisition of the crop coefficient is an important step in estimating evapotranspiration in farmland.

Methods To explore the feasibility of coupling UAV remote sensing data with ground-based soil moisture monitoring data for estimating the cotton crop coefficient and determining an appropriate irrigation schedule, this study took cotton at the Changji Huaxing Farm Experimental Base as the research object. Based on soil, meteorological, and UAV multi-spectral remote sensing data, the dual crop coefficient method, modified by meteorological data and cotton crop coverage, was used to estimate the cotton crop coefficient under different drip irrigation treatments. Two methods, namely Least Angle Regression (LAR) and Random Forest (RF) regression, were employed to analyze the relationships between the crop coefficient (Kc) and four vegetation indices derived from UAV remote sensing data: Normalized Difference Vegetation Index (NDVI), Simple Ratio (SR), Soil-Adjusted Vegetation Index (SAVI), Enhanced Vegetation Index (EVI), and soil water content (SW). Combined with model accuracy evaluation metrics, the optimal crop coefficient estimation inversion model was constructed, and a spatial distribution map of the cotton crop coefficient was ultimately generated.

Results Under different irrigation treatments, Kc showed high correlations with NDVI, SR, SAVI, EVI, and SW, indicating that establishing a cotton crop coefficient estimation model based on these indices is feasible. Under different irrigation treatments, the coefficients of determination (R²) of the optimal LAR estimation models were 0.90, 0.90, and 0.85; the root mean square errors (RMSE) were 0.16, 0.11, and 0.13; and the normalized root mean square errors (nRMSE) were 11%, 11%, and 13%, respectively. The coefficients of determination (R²) of the optimal RF estimation models were 0.98, 0.97, and 0.98; the RMSE were 0.05, 0.07, and 0.05; and the nRMSE were 5%, 6%, and 5%, respectively. Therefore, the accuracy of the optimal RF estimation model was slightly higher than that of the optimal LAR estimation model.

Conclusion It is feasible to construct a cotton crop coefficient estimation model and determine an appropriate irrigation schedule using UAV remote sensing data and ground-based soil moisture monitoring data.