# Precise tea leaf disease detection using UAV low-altitude remote sensing and optimized YOLO11 model

## TL;DR

This paper introduces a new lightweight model for detecting tea leaf diseases using drones, achieving higher accuracy and faster performance suitable for real-time monitoring.

## Contribution

The paper proposes FCHE-YOLO, an optimized YOLO11 model with novel modules for improved accuracy and reduced computational complexity for UAV deployment.

## Key findings

- FCHE-YOLO improves average accuracy (mAP) from 94.1% to 98.1% compared to YOLO11.
- The model's inference speed increases by 9.0%, reaching 47.5 FPS.
- FCHE-YOLO reduces computational complexity by 34.3% in FLOPs and 38.9% in parameter count.

## Abstract

Tea leaf diseases seriously affect its yield and quality, and consequently there is an urgent need for intelligent detection methods with high precision and edge deployment capabilities. To address low detection accuracy in complex backgrounds, overfitting due to limited data, and redundant parameters for existing methods, this paper proposes an improved lightweight detection model FCHE-YOLO based on the YOLO11, which aims to achieve rapid and accurate identification of tea leaf disease combining low altitude remote sensing with unmanned aerial vehicle (UAV). The model has made three key optimizations in the structure: Introduce the self-developed lightweight backbone module FC_C3K2, which significantly reduces computation and parameter count while enhancing the robustness of the model to complex scenarios; construct an efficient feature fusion structure HSFPN, optimizing multi-scale information integration and compressing model volume; design the detection head Efficient Head, integrating group convolution and lightweight attention mechanism to improve detection accuracy and suppress overfitting. The experimental results from the self built tea gardens show that the FCHE-YOLO improves the average accuracy (mAP) from 94.1% to 98.1% compared to the benchmark model YOLO11, with an improvement of 4.0 percentage points. Meanwhile, the inference speed of the model increases from 43.3 FPS to 47.5 FPS, with an increase of 9.0%, meeting the real-time detection requirements. More importantly, by network structure optimization, the model’s computational complexity is significantly reduced: The floating-point operations per second (FLOPs) decreases from 6.4 G to 4.2 G, with a decrease of 34.3%, and the parameter count decreases from 2.59 M to 1.46 M, with the compression rate reaching 38.9%, which makes the model more suitable for deployment on resource-constrained UAV edge devices. The final test show that the FCHE-YOLO significantly reduces the missed-detection rate, owns better detection accuracy and deployment practicality, and is suitable for real-time monitoring scenarios of tea leaf diseases with UAVs.