Machine learning methods for efficient and automated in situ monitoring of peach flowering phenology

Abstract

Accurate knowledge of peach flowering phenology is essential for scheduling precise irrigation and managing artificial pollination for breeding. However, in situ monitoring of peach flowering phenology is laborious and challenging. This study divided peach flowering phenology into eight stages covering inflorescence emergence and flowering using BBCH scales. Web-connected recording cameras and weather stations automatically-collected flower images and meteorological data from three peach cultivars. Daily variables, including area proportion of specific color ranges in the images, temperature, illuminance and heat/chilling requirement data were applied for training and testing four machine learning models: random forest (RF), support vector machine (SVM), naïve Bayes (NB), and k-Nearest Neighbors (KNN). Grid tuning and 10-fold cross-validations were conducted on these models to determine the optimal method for identifying peach flowering phenological stages. The RF model obtained highest optimal F1 score of 98.82 % (harmonic average of model accuracy and recall) among the models on the testing set. The area proportion of specific color ranges in the real-time images are essential for model performance. Heat requirement data (biological growth time and growing degree days) can improve the overall accuracy. The classification accuracy at different BBCH phenological stages of the three cultivars is acceptable for real-time monitoring of peach flowering phenology. These are applicable for peach breeding, heat stress management, and irrigation scheduling. The RF model also shows acceptable accuracy with flower images and public meteorological data input, which reveals the extendibility for phenology monitoring of other trees or crops.