# Accurate detection and density estimation of peach tree inflorescences using an improved YOLOv11 model

## TL;DR

This paper introduces an improved YOLOv11 model for accurately detecting peach tree inflorescences in complex orchard environments.

## Contribution

The novel MDI-YOLOv11 model enhances detection accuracy and adaptability for peach inflorescences using a new architecture and feature fusion techniques.

## Key findings

- MDI-YOLOv11 achieved an AP50 of 0.919 and AR50 of 0.964 for detecting peach inflorescences.
- The model outperformed YOLOv11s with a 0.033 increase in AP50 and AR50.
- The model generated accurate inflorescence density maps for orchard management.

## Abstract

Flower thinning plays a vital role in peach production, which significantly affects fruit yield and quality. Obtaining precise information about inflorescences is the key to scientific thinning and refined orchard management. However, the accurate detection of peach inflorescence still faces great challenges due to the complex and changeable light conditions, dense occlusion between flowers and significant scale differences in the actual orchard environment. In order to solve these problems, an enhanced YOLOv11s peach inflorescence detection model, termed MDI-YOLOv11, is proposed in this study to achieve accurate and stable recognition of flowers and buds. Considering the characteristics of small target and frequent occlusion in peach inflorescences, a collaborative design of the neck feature fusion structure and the backbone feature attention mechanism is adopted. Specifically, the RFCAConv module is added to the backbone network to increase sensitivity to salient regions, while a P2 layer for small target detection is embedded within the neck network and integrated with the RepGFPN structure to enhance multi-scale feature fusion, thereby improving detection accuracy and adaptability in complex orchard environments. The model’s performance was systematically assessed on a self-built dataset comprising 1,008 images. The dataset labeled 41,962 target instances after sample balancing, including 22,803 flower targets and 19,159 bud targets, covering typical orchard scenes with varying illumination, color characteristics, and high density occlusion. The five-fold cross-validation experiment demonstrated that MDI-YOLOv11 achieved an AP50 of 0.919 and an AR50 of 0.964 for peach tree inflorescences detection, along with a detection time of 13.46 ms per image. 10.97 million parameters, and a model size of 21.51MB, all of which meet practical application requirements. Compared with the YOLOv11s model, the MDI-YOLOv11 model achieved a 0.033 increase in both AP50 and AR50, and the detection performance and model complexity are better than YOLOv11m. Based on the detection results of MDI-YOLOv11, this study generated row-by-row inflorescence density distribution maps that intuitively displayed the spatial density distribution of peach inflorescences. The results indicate that the proposed method enables efficient and accurate detection of peach flowers and the generation of inflorescence density maps, which is expected to provide effective support for refined orchards management.