Abstract Predicting the dry matter yield (DMY) of the first harvest of perennial ryegrass ( Lolium perenne L.) following overwintering helps assess the risk of winter damage. This study compared machine learning (ML) models for the prediction of DMYs based on well‐organized examined data for cultivar registration trials across 13 locations in Hokkaido, Japan, over the period from 1983 to 2022, including the DMYs of first harvests, daily weather data, seedling and examined years, number of harvests in the previous year, date of last harvest in the autumn of the previous year, date of first harvest and experimental designs (drilled‐row plots or sward plots, and whether mixed with other species). The yield estimation models were developed using 17 ML methods with default hyperparameter settings. Among the ML models, successful methods included tree‐based algorithms. The accuracy scores ( R 2 ) were 0.80–0.91 for training data (5‐fold cross‐validation), and 0.79–0.92 for test data. The permutation feature importance in the best model (light gradient boosting machine) indicated that the daily minimum temperature and snow depth in winter could have a significant impact on the DMYs of the first harvests following overwintering. Additionally, the partial dependence plots revealed that frequent cutting increased the risk of overwintering damage and that it identified a specific autumn harvest period that negatively affected yield and persistence after overwintering. In conclusion, the ML model, incorporating weather and pasture management data, demonstrated high predictive performance ( R 2 = 0.92) for forecasting overwintered yield loss of perennial ryegrass in frozen soil areas of Hokkaido.
Tanaka et al. (Tue,) studied this question.