Machine learning-based in-season nitrogen status diagnosis and side-dress nitrogen recommendation for corn

Reliable and efficient in-season nitrogen (N) status diagnosis and recommendation methods are crucially important for the success of crop precision N management (PNM). The accuracy of these methods has been found to be influenced by soil properties, weather conditions, and crop management practices. It is important to effectively incorporate these variables to improve in-season N management. Machine learning (ML) methods are promising due to their capability of processing different types of data and modeling both linear and non-linear relationships. The objectives of this study were to (1) determine the potential improvement of in-season prediction of corn N nutrition index (NNI) and grain yield by combining soil, weather and management data with active sensor data using random forest regression (RFR) as compared with Lasso linear regression (LR) using similar data and simple regression (SR) models only using crop sensor data; and (2) to develop a new in-season side-dress N fertilizer recommendation strategy at eighth to ninth leaf stage (V8-V9) of corn developement using the RFR model. Twelve site-year experiments examining corn N rates and planting densities were conducted in Northeast China. The GreenSeeker sensor data and corn NNI were collected at V8-V9 stage, and grain yield was determined at the harvest stage (R6). The soil information was obtained at planting and the weather data was measured throughout the growing season. The results indicated that corn NNI and grain yield were better predicted by combining soil, weather and management information with GreenSeeker sensor data using RFR model (R² = 0.86 and 0.79) and LR model (R² = 0.85 and 0.76) as compared with only using GreenSeeker sensor data (R² = 0.66 and 0.62–63) based on the test dataset. An innovative in-season side-dress N recommendation strategy was developed using the RFR grain yield prediction model to simulate corn grain yield responses to a series of side-dress N rates at V8-V9 stage. Based on these response curves, site-, and year-specific optimum side-dress N rates can be determined. The scenario analysis results indicated that this RFR model-based in-season N recommendation strategy could recommend side-dress N rates similar to those based on measured agronomic optimum N rate (AONR) or economic optimum N rate (EONR), with root mean square error (RMSE) of 17 kg ha⁻¹ and relative error (RE) of 14–15 %. It is concluded that combining soil, weather and management information with crop sensor data using RFR can significantly improve both in-season corn NNI and grain yield prediction and N management, compared with the approach based only on crop sensor data. More studies are needed to further improve and evaluate this approach under diverse on-farm conditions.