CNN deep learning performance in estimating nitrate uptake by maize and root zone losses under surface drip irrigation

• Published in: Journal of hydrology (Amsterdam) Vol. 625; p. 130148
• Main Authors: Azad, Nasrin, Behmanesh, Javad, Rezaverdinejad, Vahid, Khodaverdiloo, Habib, Thompson, Sally E., Mallants, Dirk, Ramos, Tiago B., He, Hailong
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.10.2023
• Subjects: Convolutional neural networks; Deep learning; HYDRUS-2D; Model emulation; Deep learning; HYDRUS-2D; Convolutional neural networks; Model emulation
• ISSN: 0022-1694; 1879-2707
• DOI: 10.1016/j.jhydrol.2023.130148

[Display omitted] •Machine-learning as an emulator can reduce the computational demands of optimization.•ETp, irrigation and fertilizer amount can accurately predict daily nitrate uptake.•CNN model has sufficient accuracy in estimation of nitrate uptake and root zone losses.•Root zone nitrate losses in sandy clay loam was more than sandy loam soil. Optimal fertigation regimes will minimize the leaching of agrochemicals while providing crops with sufficient nutrition for growth. A comprehensive objective function of optimization involves minimizing (i) difference between weekly plant nitrogen uptake and requirement and (ii) nitrate losses (nitrate leaching and accumulated nitrate in the root zone). Coupling a simulation model to an optimization algorithm for obtaining the objective function variables (N plant uptake and nitrate losses) typically requires long runtimes as the simulation model is repeatedly applied over iterations of the optimization algorithm. One attractive way to reduce the programming and computational demands associated with optimization is to construct a simpler, faster, but sufficiently accurate emulator of the simulation model based on machine-learning model. This study assessed the performance of convolutional neural networks (CNN) in modelling nitrate uptake by maize and estimation of nitrate losses in surface drip irrigation in three texturally different soils. To do this, the CNN was trained on output of the simulation model HYDRUS-2D. Various combinations of parameters influencing daily nitrate uptake (daily potential crop evapotranspiration, irrigation water and injected fertilizer amount) were evaluated as model inputs. Results indicated that daily nitrate uptake at any time t was predicted best with inputs of daily potential evapotranspiration at time t, irrigation and fertilizer amounts of the previous 1–7 days and the previous day’s nitrate uptake amount. Estimated nitrate losses for constant weekly fertigation (CWF) were about 32% in sandy loam and 38.5% in sandy clay loam soil. Variable weekly fertigation (VWF) reduced the nitrate losses to 24.5% in sandy loam and 31.5% in sandy clay loam soil, while losses in loam soil were intermediate to these. The study demonstrates that CNN estimation of nitrate uptake and root zone losses can be widely used in optimization of fertigation management.