Neural network model predictions for phosphorus management strategies on tile-drained organic soils

Abstract

Abstract The organic soils of Holland Marsh, Ontario are used for intensive vegetable production, which demands high-phosphorus (P) fertilizer applications. Such high-fertilizer applications on these tile-drained lands lead to eutrophication in surrounding water bodies. This study investigated the application of neural network (NN) models for deriving P management strategies. Seven NN models were assessed using the following two approaches: a time series with 1-year training and 1-year testing of the models and a randomization analysis where a random 80% of data was used for model training and the remainder for model testing. The feed-forward model using the randomization and the long-short-term memory model using time-series outperformed all other models. Two strategies for P management were evaluated: a direct approach that predicts P loads using new fertilizer rates or controlled drainage discharge rates, and a particle swarm optimization (PSO) that used a percent reduction of actual P loads to predict an optimal water table management strategy. Overall, the direct approach identified a water table level of 30 cm from the soil surface during the spring and 80 cm during the summer period as optimal to reduce P loads. The PSO analysis showed that a reduction of P loads by 20% in the spring and up to 40% in the summer through water table control would not compromise crop production.