Estimating soil mineral nitrogen from data-sparse field experiments using crop model-guided machine learning approach
Sandy soils are susceptible to excessive nitrogen (N) leaching under intensive crop production which is linked with the soils low nutrient holding capacity and high-water infiltration rate. Estimating soil mineral nitrogen (SMN) at the daily time-step is crucial in providing fertilizer recommendations balancing plant nitrogen use efficiency (NUE) and N losses to the environment. Crop models [e.g., Decision Support System for Agrotechnology Transfer (DSSAT)] can simulate the trend of SMN in varied fertilizer rates and timing of application but are unable to replicate its magnitude due to the inability to capture high-water table conditions in a sub-irrigated soil. As an alternative to such physics-based model, time-series deep learning (DL) models based on a long short-term memory (LSTM) are promising in understanding nonlinearity among complex variables. Yet, purely data-driven DL models for crops are difficult to obtain due to the insufficient amount of data available and the excessive costs with producing more data. To address this challenge, a hybrid model (hybrid-LSTM) was developed by leveraging both the DSSAT and LSTM models to estimate daily SMN primarily using daily weather, applied fertilizer rates-timings, and the SMN sparse observations. This study used the observations from field trials conducted between 2010-2014 in Hastings, FL. The first step was to calibrate the DSSAT-SUBSTOR-Potato model to produce reliable SMN of the topsoil for treatments with varied N applied fertilizer rates split among the pre-planting, emergence, and tuber-initiation stages of the potato crop. Thereafter, the hybrid-LSTM model was trained on the calibrated DSSAT simulated SMN time-series and fine-tuned its predictions using the observed SMN to improve DSSAT simulated SMN. The hybrid-LSTM model was then tested on both calibrated and uncalibrated DSSAT SMN simulations where it outperformed the DSSAT model (range of improvement ranged [~]18-30% on comparing the normalized root mean squared error) in providing reliable estimates of SMN across most of the farms and years. This novel hybrid modeling approach could guide stakeholders and farmers to build sustainable N management with improved crop NUE and yield and help in minimizing environmental losses.