Machine learning-based canola yield prediction for site-specific nitrogen recommendations

An effective crop yield prediction is critical for making precise nitrogen (N) management decisions. A multi-site-year study was conducted across eastern Canada with the objectives to (1) construct machine-learning random forest regression (RFR) models for canola yield prediction; and (2) develop site-specific N recommendations using the RFR modelling approach. Soil characteristics, weather, plant growth and spectral index data collected from 2013 to 2015 were used to train the RFR models and the 2016 data for model validation in terms of making N decisions. Results showed that canola yields responded positively to N application rates in 16 out of the 18 site-year environments, with sometimes higher yields and an average of 8% higher N use efficiency for the split-N strategy than for the same amount of N applied only at preplant. Heat stress and precipitation distribution were identified as of critical importance in total yield variation. The RFR model by combining weather, soil and plant growth features with the spectral indices displayed highly improved prediction reliability, up to 85%, with 53–57% lower root mean square errors, compared with the model based only on leaf chlorophyll or normalized difference vegetation index. Based on the RFR algorithm, an average economic optimum N rate of 150 kg N ha −1 was recommended for most canola production scenarios in the test year. This study demonstrated that the machine learning-based RFR modelling approach can be used to implement optimal nutrient management strategies for sustainable crop production, which is sensitive and better adapted to environment-induced abiotic stresses.