Weather data-centric prediction of maize non-stressed canopy temperature in semi-arid climates for irrigation management

Canopy temperature (T_c) measurements are increasingly being used to compute crop thermal indices for water stress estimation and improved irrigation management. Conventionally monitoring crop thermal response requires maintenance of a well-watered crop from which non-stressed canopy temperature (T_cns) is measured as a reference for thermal index computation. This study alternatively evaluated the performance of 36 weather data driven model combinations to predict peak time (12:00–17:00 h) T_cns in maize grown in semi-arid climates at the West Central Research, Extension, and Education Center (WCREEC) in North Platte, NE, and at the Limited Irrigation Research Farm (LIRF) in Greeley, CO. Data-driven models considered were multilinear regression (MLR), forward feed neural network (NN), recurrent neural network (RNN), multivariate adoptive regression splines (MARS), random forest (RF), and k-nearest neighbor (KNN). For each of these models, the following weather data combinations were tested: average air temperature (T_a), average relative humidity (RH), wind speed (U₂), and solar radiation (R_s) (combination 1); RH, U₂, R_s (combination 2), T_a, RH, R_s (combination 3); T_a, RH (combination 4); RH, R_s (combination 5); and T_a, R_s (combination 6). Ranking the performance of weather data × model combinations across both climate sites showed that MARS model with combination 1 was a better predictor of T_cns with R² of 0.866 and RMSE value of 0.966 °C at WCREEC and R² of 0.910 and RMSE value of 0.693 °C at LIRF. The performance of site specific (localized) and generalized model combinations was compared and indicated that cross site prediction of T_cns was primarily determined by weather data combinations, rather than model specificity.