Coordinate descent based agricultural model calibration and optimized input management

Authors: BHAR, ANUPAM - Iowa State University; KUMAR, RATNESH - Iowa State University; QI, ZHIMING - McGill University - Canada; Malone, Robert - Rob

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 3/14/2020
Publication Date: 3/31/2020
Citation: Bhar, A., Kumar, R., Qi, Z., Malone, R.W. 2020. Coordinate descent based agriculture model calibration and optimized input management. Computers and Electronics in Agriculture. 172:105353. https://doi.org/10.1016/j.compag.2020.105353.
DOI: https://doi.org/10.1016/j.compag.2020.105353

Interpretive Summary: Agricultural models that accurately simulate field conditions are critical for developing optimal fertilizer and irrigation management practices under the vast array of expected conditions (e.g., climate, soil, management combinations). Problems include determining the best set of: 1) model input variables to optimize model performance compared to field data and 2) management practices to optimize crop production and producer profit. A few methods have been previously developed to automatically optimize agricultural models; however, the Coordinate Descent (CD) algorithm that is popular in Machine learning communities has not been used for this purpose. Here, we use CD to optimize the Root Zone Water Quality Model (RZWQM) and then use the optimized model in conjunction with several algorithms (e.g., Differential-Evolution, Sequential-Least Square) to optimize fertilizer and irrigation management. The CD optimization resulted in more accurate model predictions of crop production, evapotranspiration, and soil water content than previous optimization attempts at this field site in northeastern Colorado. The automated methods to optimize fertilizer and irrigation management increased both RZWQM simulated crop yield and producer profit compared to field management. This research provides a new technique to optimize agricultural system models and will help model developers and model users more efficiently estimate model input variables. The aspect of the study that investigated automated methods to optimize fertilizer and irrigation management may help agricultural scientists and the agriculture industry more effectively and efficiently design sustainable systems.

Technical Abstract: A well-calibrated agricultural system model with many parameters is critical for optimized management. Here, the Root Zone Water Quality Model is calibrated automatically using the Coordinate Descent (CD) algorithm against measured data from a fully irrigated corn field in terms of yield, plant height, leaf area index, evapotranspiration, and soil water content. Fifty-six soil hydraulic and three crop parameters were calibrated. Average Model Efficiency (ME) using the manual and automated calibration was 0.71. The CD method reported an average R2 of 0.79 vs 0.77 for the manual calibration. The CD calibrated model was validated against data from an adjacent deficit irrigated field. The average R2 measure was found to be 0.77 (against 0.74 for manual) and the average ME was 0.64 (against 0.61 for manual) for validation. Once calibrated, fertilization and irrigation are optimized so that farm profit is maximized. Profit is defined as (corn price)*(yield) – (fertilizer N cost)*(amount of N) – (irrigation cost)*(amount of irrigation). The optimization variables are amount of fertilizer and irrigation water. Three global optimization methods, namely, Differential-Evolution, Basin-Hopping, and Particle-Swarm and two local optimization methods, namely, Sequential-Least-Square and Constrained-optimization-by-linear-approximation were tried. All methods increased the yield by 7% and profit by 10% as compared to what was actually applied in the field.