Forecasting Flowering and Maturity Times of Barley Using Six Machine Learning Algorithms

Abstract

Interior Alaska has a short growing season of 110 d. The knowledge of timings of crop flowering and maturity will provide the information for the agricultural decision making. In this study, six machine learning algorithms, namely Linear Discriminant Analysis (LDA), Support Vector Machines (SVMs), k-nearest neighbor (kNN), Naïve Bayes (NB), Recursive Partitioning and Regression Trees (RPART), and Random Forest (RF), were selected to forecast the timings of barley flowering and maturity based on the Alaska Crop Datasets and climate data from 1991 to 2016 in Fairbanks, Alaska. Among 32 models fit to forecast flowering time, two from LDA, 12 from SVMs, four from NB, three from RF outperformed models from other algorithms with the highest accuracy. Models from kNN performed worst to forecast flowering time. Among 32 models fit to forecast maturity time, two models from LDA outperformed the models from other algorithms. Models from kNN and RPART performed worst to forecast maturity time. Models from machine learning methods also provided a variable importance explanation. In this study, four out of six algorithms gave the same variable importance order. Sowing date was the most important variable to forecast flowering but less important variable to forecast maturity. The daily maximum temperature may be more important than daily minimum temperature to fit flowering models while daily minimum temperature may be more important than daily maximum temperature to fit maturity models. The results indicate that models from machine learning provide a promising technique in forecasting the timings of flowering and maturity of barley.