Architecture-based machine learning models for peach yield prediction before bloom

Abstract

Early yield prediction is essential for optimising fruit crop management. This study evaluates three machine learning models—random forest (RF), extreme gradient boosting (XGBoost), and support vector machines (SVM)—to predict peach (‘N48-52’) yields across four tree levels using non-destructive pre-bloom architectural measurements. Data on structural dimensions, fruit count, and weight were collected in 2019, 2021, and 2022. Eighty percent of the dataset was used for model training and 20% for validation. The results showed that SVM performed best for the third level (R2 = 0.91), while RF achieved the highest accuracy for the first, second, and fourth levels (R2 = 0.79, 0.91, and 0.93, respectively). Additional data collected in 2023 were used to further validate model stability and accuracy. The models maintained strong predictive performance with R2 values of 0.9054, 0.7684, 0.8768, and 0.7964. A comparison between estimated and actual production showed statistically similar results at a significance level of α = 0.05, confirming the reliability of the proposed models for early yield prediction in peach orchards. For fruit crop management to be optimised, early yield prediction is essential. In order to predict peach yields (‘N48-52’) across four tree levels using non-destructive, pre-bloom architectural measurements, this study assesses three machine learning (ML) models: random forest (RF), extreme gradient boosting (XGBoost), and support vector machines (SVM). Structural dimensions, fruit count, and weight were among the data gathered in 2019, 2021, and 2022. In order to train the models, 80% of the dataset was used, and the remaining 20% was used for validation. According to the results, SVM performed best for the 3rd level (coefficient of determination (R2) = 0.91), while RF was the most accurate for the 1st, 2nd, and 4th levels (R2 = 0.79, 0.91, and 0.93, respectively). To further enhance the accuracy of the proposed models, additional data points were randomly collected from different trees in 2023. These data included measurements of the complete path to a given level along with its fruit production, allowing for verification of the precision and stability of the proposed models. In 2023, the models maintained an accuracy of R2 = 0.9054, 0.7684, 0.8768, and 0.7964, respectively, for RF (1st level; 2nd level; 4th level) and SVM (3rd level). A comparison between the estimated production from the trees and the actual production showed a statistically similar result (accepted statistical error for the analysis of variance statistical test (α = 0.05).