Blooming Level And Density Estimation Of Apple Flowers Based On Deep Learning

Apple tree thinning is beneficial for improving apple yield and fruit quality.Choosing a reasonable timing and intensity of flower thinning helps reduce fruit trees’load level,improve their disease resistance,reduce yield fluctuations,and improve the quality of flower buds.The estimation of apple flower blooming level and the accurate judgment of flower thinning timing influence the decision of orchard managers to make production planning.At the same time,the high density and distribution of apple flowers affect the intensity of thinning.The production management of apple orchard flower thinning relies on manual experience,which is labor-intensive and lacks objective estimation.The use of computer vision to estimate orchard blooming levels and flower density distributions has received widespread attention.Traditional apple flower detection methods based on adjusting colors and thresholds are limited by weather,lighting conditions,and shooting locations,making the method inappropriate for orchards in complex natural environments.Deep learning-based image detection algorithms have been shown to have higher accuracy and robustness in agricultural detection.In this study,a deep learning-based apple flower blooming level estimation algorithm and an apple flower density detection method were proposed respectively,which completed the estimation of apple flower blooming level and the determination of peak blooming date to provide decision support for flower thinning timing decision and obtained an apple flower density distribution map to provide objective flower density data for variable flower thinning intensity thinning machinery.The main research contents and results of this paper are as follows:（1）An apple flower data collection device was built and completed to collect apple flower data at different blooming levels and in different natural environments,in multiple scenarios,to produce apple flower blooming level data set and apple flower density data set.（2）In this study,an apple flower blooming level estimation method is proposed for global or block-level blooming level estimation in orchards.The method consists of a deep learning-based apple flower detector,a blooming level estimator,and a peak blooming date finder to achieve the estimation of six blooming levels in apple orchards.The YOLOv5s model is used as the apple flower detector and the model is improved by adding a co-attentive mechanism and a small target detection layer and replacing the model neck with a bidirectional feature pyramid network（Bi-FPN）structure to improve the performance of detecting apple flowers at different blooming levels.The robustness of the apple flower detector under different light conditions and the generalization across years of data were tested using apple flower data from 2021-2022.The average precision of the apple flower detector reached 77.5%.The statistical results showed that the blooming level estimator followed the trend of orchard blooming level;the peak bloom date finder determined the peak bloom time and provided information on the timing decision of flower thinning.（3）Detecting apple flower density and extracting spatial variation information of flowering density in orchards provide important references of flower thinning intensity for variable flower thinning intensity thinning machinery.This study proposes a U-Net model modified by adding ASPP module to implement pixel-level apple flower segmentation for detecting apple flower density.The model achieved an F₁-score of 88.3%and an Io U value of79.1%for apple flower segmentation.The inference results of the model were fused with the depth information to remove the background pixels to obtain the apple flower density map.The density map successfully reflected the differences in flower density distribution among fruit trees and between canopies of the same fruit tree.