A machine learning approach for estimating forage maize yield and quality in NW Spain

Crop models simulate crop growth and development according to different climatic, soil and crop management conditions. The CSM-CERES-Maize model (DSSAT) was adapted to simulate forage maize yields by calibrating the genetic parameters of six cultivars: SE1-200, SE2-300 and SE3-400 in three sites and three years in Asturias, and XU1-220, XU2-300 and XU3-400 in four sites and three years in Galicia. Calibration using the CSM-CERES-Maize model, together with the use of historical meteorological data (2000-2022) from the study sites, enabled simulation of forage maize yield (whole plant dry matter yield) and quality (whole plant net energy for lactation yield and whole plant crude protein yield) for six cultivars during the 23-year period. LightGBM models (a machine learning technique) were used with the simulated forage maize yield, quality data, historical weather, soil, and management data to capture non-linear relationships in the data and to identify the most influential variables for crop yield and quality predictions. The results of the model evaluation yielded an accuracy of 94.7%, (R2 score = 0.86) for forage maize yield, an accuracy of 94.0% (R2 score = 0.84) for the net energy for lactation yield and an accuracy of 93.0% (R2 score = 0.85) for the crude protein yield. Variable importance plots revealed Growing Season and Radiation from sowing to harvest to be the top two most influential predictor variables. In Asturias and Galicia, the cultivars with the longest cycle (cultivars cycle 400) are those with the highest values for the variables studied in the 23 years of historical meteorological data (average of three sites in Asturias and four sites in Galicia with three sowing dates in each site). The models will be available to make predictions for forage maize yield and quality by non-specialist users, using the geographical location of the crop field, cultivar type, sowing and harvest date and probable values of weather variables during the growing season as input data.