Early Apple Yield Prediction Based on Flowering Stage Image Thinning Simulation Characteristics.

Area of Science:

• Horticulture and Agricultural Science
• Computer Vision and Machine Learning
• Precision Agriculture

Background:

• Current fruit tree yield prediction relies on late-stage data or complex environmental factors, hindering early estimation.
• Flowering period images present challenges due to dense, overlapping flower structures, limiting traditional feature extraction.
• Existing methods often overlook the spatial clustering of flowers and its correlation with yield.

Purpose of the Study:

• To develop an early apple yield prediction model utilizing flowering stage image thinning simulation characteristics.
• To address limitations in early yield prediction by analyzing complex flower spatial distributions.
• To explore the relationship between flower clustering patterns and final fruit yield.

Main Methods:

• Collected blossom images and yield data from 100 apple trees, processing images to extract flower masks.
• Enhanced DBSCAN clustering with KDTree and adaptive multi-scale mechanisms (AMS-DBSCAN) for efficient flower identification.
• Designed density and spatial uniformity thinning simulation strategies to create interpretable phenotypic features, fused with statistical features, and optimized using Lasso regression.

Main Results:

• The XGBoost model achieved high predictive performance (R² = 0.856) with the hybrid feature set.
• Feature optimization with Lasso further improved XGBoost's predictive accuracy (R² = 0.900).
• The proposed method demonstrated reliable early yield estimation capabilities.

Conclusions:

• The novel approach based on flowering stage image thinning simulation enables accurate early apple yield prediction.
• This method offers a valuable tool for precision management and timely decision-making in apple cultivation.
• The study highlights the importance of spatial clustering and thinning simulation in understanding yield determinants.