Predicting flower induction of litchi (Litchi chinensis Sonn.) with machine learning techniques

Abstract

The flower induction is a critical physiological change during which vegetative buds transit to floral buds. The duration of flower induction (DFI) for litchi plays determinative role for the success and the quality of flowering. It is hard to be reliably predicted because multiple factors including ages of trees, variety and dynamically environmental and climate variables have important impacts. Here we predicted the DFI for four litchi varieties using random forest (RF) implicit and stepwise regression (STR) explicit machine learning models. These models were trained and validated from the data consisting phenological phases from the mature of the last autumn shoots to the flower shedding, and the corresponding meteorological factors from 2009 to 2020. The DFI predictive models consider timescales from 1 h up to 10 days, and the determination coefficients from the 5-fold cross validation achieves 0.96 to 0.99. The high accuracy was maintained in the blind test, with the determination coefficients of 0.97–0.98 for the data in 2019 and 0.78–0.88 for 2020. The reliability is still sufficient to guide the cultivation of litchi although the dramatic climate change weaken the inherent multiple correlations for DFI. From these robust predictive models, we found ten important features that can be used to determine the DFI with consistent positive or negative contributions. They are the minimum cooling rate, the maturity time of the last autumn shoot, the maximum heating rate, the mean of minimum temperature, the mean of maximum temperature, the amount of heat accumulation at hot day with the minimum daily temperature above 22 °C, the maximum daily temperature below 26 °C, the minimum daily temperature below 6 °C, age, and the atmospheric relative humidity.