Abstract:

 Based on a dataset of 177 samples of fruit set rate and meteorological factors obtained from experiments in Fujian from 2020 to 2024, a correlation analysis was used to screen for characteristic meteorological factors that significantly affect fruit set rate. Three machine learning algorithms, that of multiple stepwise regression (MSR), random forest (RF) and gradient boosting decision tree (GBDT), were employed to develop meteorological models for predicting the fruit set rate of yellow passion fruit. The study aimed to provide a scientific reference for optimizing cultivation management and enhancing the production of yellow passion fruit in Fujian. The results showed that: (1) the MSR model showed that only four out of 40 characteristic meteorological factors significantly indicated the fruit set rate of yellow passion fruit at the 0.05 significance level. Specifically, the higher average temperatures from 10d before flowering date to the flowering date, shorter sunshine duration during the 11−20d before flowering date, greater precipitation on the flowering date and longer sunshine duration on the flowering date could lead to a decreased fruit set rate of yellow passion fruit. (2) Analysis of the top eight most important meteorological factors among 40 characteristic meteorological factors for the RF and GBDT models revealed that the common determinant meteorological factors influencing the fruit set rate of yellow passion fruit were the maximum temperature from 10d before flowering date to the flowering date, the average temperature from 10d before flowering date to the flowering date, the precipitation on the flowering date, relative humidity on the 20d before flowering date, precipitation during the 11−20d before flowering date. (3) The root mean square error (RMSE) values of the MSR, RF and GBDT models on the test set were 16.7, 10.7 and 11.7 percentage points, respectively, while the determination coefficients (R²) were 0.66, 0.86 and 0.83, indicating favorable simulation performance across all three models. The MSR model demonstrated strong interpretability, although its prediction accuracy was significantly lower than that of the GBDT and RF models. The RF model achieved the best performance on the test set, exhibiting superior generalization capability. In contrast, the GBDT model showed the best fit on the training set (RMSE=6.8 percentage points, R²=0.95), yet its performance on the test set was slightly inferior to that of the RF model.

Key words: Yellow passion fruit, Fruit set rate, Machine learning, Meteorological model