Understanding the flower rate of litchi (Litchi chinensis Sonn.) through machine learning predictive models

Flower rate is an important indicator dominating the harvest litchi fruit of high quality, which has complex interactions among inherent characteristics of individual trees, climate, soil, fertilizer and responsive managements. Quantitative prediction of the flower rate utilizing machine learning algorithms is feasible, so to determine important features and to implement appropriate management responses. In this study, based on a comprehensive collection of data in climate conditions, cultivar types, tree ages, phenological phases and historical flowering patterns coupled with the flower rate of litchi spanning the years from 2009 to 2020 in the southern region of China, a set of robust predictive models were constructed. Binary classification models delivered an AUC in the range from 0.73 to 0.90, and regression models achieved the coefficient of determination (R 2 ) from 0.64 to 0.73 in the blind tests on annual data. SHAP analysis revealed that the date of floral white millet emergence, the accumulation of chilling hours below 9, 10 and 13 °C in the process of flower induction and their annual average, ages for trees, and the mean atmospheric relative humidity per hour in the process of flower induction are the most important features. Through the exploration of their positive or negative contributions to the floral induction, stimulate the timely emergence of “white millet” through irrigation, pruning, and foliar fertilization, practical guidance can be obtained and validated through expertized rationality. This study provided solid guidance to improve the flower rate of litchi that can promote the efficient responsive management to harvest litchi fruit of high quality.