Estimating crop evapotranspiration using drone imagery, ground canopy temperature, and machine learning techniques

Abstract

Efficient irrigation management relies on accurately estimating crop evapotranspiration (ETc), yet conventional methods often face limitations, such as cost, spatial coverage, data requirements, and the need for local calibration. This study had two main objectives: 1) to quantify daily ETc of sweet corn (SC) and green beans (GB) using crop water stress index (CWSI) calculated from canopy temperatures (Tc) and crop coefficient (Kc) estimated vegetation indices and 2) to evaluate the potential of machine learning (ML) models in estimating daily ETc and Tc. Irrigation experiments were conducted during the winter seasons of 2020–2021 and 2021–2022 at the University of Florida's Tropical Research and Education Center (TREC), Florida, USA. Networks of above-canopy infrared thermocouples (IRTs) and soil moisture (SM) sensors were used to collect Tc and SM. Multispectral images were also collected using an unmanned aerial vehicle (UAV)-based RedEdge-MX sensor. Sub-hourly changes in SM during dry periods were aggregated to estimate the daily measured ETc of SC and GB. Time series of CWSI were generated from Tc, while Kc was estimated using eleven vegetation indices (VIs) generated from drone imagery. Moreover, four ML models, i.e., CatBoost (CB), Random Forest (RF), k-Nearest Neighbors (kNN), and Extreme Gradient Boosting (XGB), were evaluated for simulating ETc. Six models, i.e., CB, kNN, RF, XGB, Light Gradient Boosting Machine (LGB), and Deep Learning (DL), were also evaluated for simulating the Tc of SC and GB. The results showed that the CWSI approach was acceptable in estimating ETc with an average MAE of 0.90 mm day^-1 for SC and 0.62 mm day^-1 for GB. Four out of eleven vegetation indices (VIs) demonstrated superior performance in estimating daily ETc, including the Soil Adjusted Vegetation Index (SAVI), Normalized Green-Red Difference Index (NGRDI), Red Edge Normalized Difference Vegetation Index (RENDVI), and NIR-RE normalized difference vegetation index (NIRRENDVI). The ML models captured ETc and Tc with better accuracy. Averaged root mean square error (RMSE) for ETc across the four models was 0.86 mm day^-1 for SC and 0.89 mm day^-1 for GB. The average RMSE of the ML models for simulating Tc was ±1.1 °C for SC and ±1.5 ^°C for GB. Overall, CWSI, spectral reflectance-based Kc, and ML models proved to be useful tools for estimating ETc at finer spatial and temporal scales with reasonable accuracy.