Understanding the flowering process of litchi through machine learning predictive models

[Objective] China is the origin country of litchi (Litchi chinensis Sonn.) and the largest producer in the world. The low or unstable yield caused by unstable flowering is a prominent problem in litchi production, and the flowering time affects not only the maturity of fruit, but also the flowering rate and yield of litchi. The meteorological factors including air temperature, relative air humidity, rainfall, and wind level, and other factors including variety and tree age affect flower differentiation of litchi. However, there is a lack of systematic research on how the development stage of litchi flowers is affected by the meteorological factors. Accurately predicting the development of the inflorescence and the process of flowering duration, as well as correctly understanding the quantitative relationship between the flowering phenology and the meteorological factors, is very important for the high-yield and quality production of litchi. The machine learning algorithms can handle high-dimensional nonlinear data with complex interactions, outperform traditional statistical models in ecology, and have been effectively used for plant classification, phenology detection, crop growth detection, and yield prediction. The objective of this study was to develop regression models for litchi inflorescence development duration and flowering duration using machine learning algorithms including RF and STR and to analyze and assess the importance and relevance of selected features on the flowering duration according to RF algorithm in order to provide a theoretical basis for the prediction of litchi flowering period and realizing precise regulation. [Methods] Firstly, the litchi phenological period data were obtained from the National Litchi and Longan Industry Technology System (CARS), with a total of 2204 records. It covered 201 demonstration litchi orchards distributed in 53 cities and counties in Hainan Province, Guangdong Province, Guangxi Zhuang Autonomous Region, Fujian Province and Sichuan Province. The time span was 2009—2018, including 47 varieties such as Guiwei, Nuomici, Huaizhi, Feizixiao, etcs. The meteorological data were downloaded from the website “https://tianqi.911cha.com/” and recorded at a frequency of one hour, with meteorological factors including atmospheric temperature, atmospheric relative humidity, wind scale and rainfall. Feature engineering of the data, which involved removing irrelevant or redundant features and ensuring that there was no high correlation between the retained features, was used to improve the performance and generalization of the model. The six classical machine algorithms including Classified Regression Tree (CART), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Random Forest (RF), Stepwise Regression (STR) and Gradient Boosting Machine (GBM) were used for training. The algorithms (RF and STR) with the smallest Mean Absolute Error (MAE) and the highest residual error (RMSE) and the highest correlation coefficient (RP2) were selected for further parameter optimization and evaluation. A 5-fold cross-validation with 999 repetitions was performed on all trained machine learning models. The random seeds are set during resampling, parameter tuning and model training to ensure model reproducibility. The models were applied to be constructed in R-project (version 3.5.2) and the ‘caret’ package was applied to tune the machine learning algorithm parameters. [Results] The residual error of the model were 3.6-3.7 days, and the correlation coefficient were 0.97, so the models had high reliability; The model was further verified with blind test data set of two-year's phenological ecological characteristics, and the correlation coefficient was between 0.98-0.99. It was indicated that the series of prediction models could be applicable to accurately predict the development of inflorescence. Similarly, the residual error of the model predicted the shedding period were 1.2-2.6 days, and the correlation coefficient were 0.88-0.97, so the model had high reliability; The model was further verified with blind test data set of two year's phenological ecological characteristics, and the correlation coefficient was between 0.96-0.98, Indicating that the series of prediction models could be applicable to accurately predict the flowering duration. The daily accumulated temperature above 5 ℃, daily average temperature, wind level and rainfall were found to played an important role in the whole process of the florescence period of litchi. In addition, a daily accumulated temperature above 24 ℃ had great impact on the development of inflorescence, while daily a cumulated temperature above 18 ℃ had significant effect on the flowering duration. [Conclusion] The robustness and predictive fit of the regression model established in this study were high. After the verification of two year's data, the accuracy and stability of the prediction were ideal. These models would be important to judge and regulate the maturity period and market volume of litchi. And the characteristic features screened out were helpful to understand the complex influence of external meteorological factors on the flowering process.