Evaluation of artificial intelligence algorithms with sensor data assimilation in estimating crop evapotranspiration and crop water stress index for irrigation water management

Authors

Abia Katimbo, University of Nebraska-LincolnFollow
Daran Rudnick, University of Nebraska-LincolnFollow
Jingwen Zhang, University of Illinois at Urban-Champaign
Yufeng Ge, University of Nebraska-LincolnFollow
Kendall C. DeJonge, United States Department of Agriculture
Trenton E. Franz, University of Nebraska-LincolnFollow
Yeyin Shi, University of Nebraska-LincolnFollow
Wei‑zhen Liang, University of Nebraska-LincolnFollow
Xin qiao, University of Nebraska - LincolnFollow
Derek M. Heeren, University of Nebraska-LincolnFollow
Isa Kabenge, Makerere University
Hope Njuki Nakabuye, University of Nebraska-LincolnFollow
Jiaming Duan, University of Nebraska-Lincoln

Document Type

Article

Date of this Version

1-7-2023

Citation

Smart Agricultural Technology 4 (2023) 100176. https://doi.org/10.1016/j.atech.2023.100176

Comments

Open access.

Abstract

Irrigation water management using automated irrigation decision support system (IDSS) as a smart irrigation scheduling tool can improve water use efficiency and crop production, especially under circumstances of limited water supply. The current study evaluated the performance of different artificial intelligence (AI) algorithms and their ensembles in forecasting Crop Evapotranspiration (ET_c) and Crop Water Stress Index (CWSI) against calculated single crop coefficient FAO56 ET_c and Jackson’s theoretical CWSI, respectively. Soil moisture, canopy temperatures (T_c) and Normalized Difference Vegetation Index (NDVI) were all measured from irrigated and non-irrigated maize plots in West Central Nebraska during 2020 and 2021 growing seasons. There were fifteen and twelve input combinations used for ET_c and CWSI predictions, respectively, having input variables such as weather and soil moisture as well as ancillary variables, including NDVI, reference evapotranspiration (ET_r), and cumulative growing degree days (CGDDs). While evaluating the models, four statistical performance indicators including coefficient of determination (r²), root mean square error (RMSE), mean absoluter error (MAE), and mean absolute percentage error (MAPE) were used. Furthermore, ranking scores were performed on statistical results to find the overall best model across all the input combinations. Based on total ranking scores, CatBoost (RMSE ranging between 0.06 – 0.09 unitless) was the best model in predicting CWSI, while Stacked Regression (RMSE ranging between 0.27 – 0.72 mm d^-1) was the best model for ET_c estimation. Future research will consider designing and evaluating an IDSS using identified best machine learning models to establish soil water and plant stress feedback for automated irrigation scheduling.