基于改进YOLOv8m-pose的作物株高检测方法

A Lightweight Crop Plant Height Detection Method Based on Improved YOLOv8m-pose

关键词:  无人机  株高测量  轻量化网络  小目标检测  注意力机制  表型信息

Key words:  Unmanned aerial vehicle  Plant height measurement  Lightweight network  Small object detection  Attention mechanism  Phenotypic information

摘要:

【目的】针对无人机三维重构测高受限于算力及地面高程反演误差，以及现有轻量化网络在微小关键点定位时易丢失特征且受同质化背景干扰等问题，提出一种高精度、轻量化的作物株高检测方法。【方法】构建涵盖稻麦完整生育期的无人机与E字尺联合数据集，设计基于关键点像素映射的株高线性缩放算法。基于改进YOLOv8m-pose提出轻量化网络Height-YOLO：设计高通感知模块HPModule，通过自适应高通滤波与通道-空间联合注意力机制强化微观尖点特征提取；引入颜色对比注意力模块CCA，利用多尺度膨胀卷积与显式差分建模提升同质化背景下的目标边界辨识度。【结果】对比实验表明，Height-YOLO的关键点定位精度mAP(P)_(50-95)达0.9879，平均绝对误差（MAE）为3.5265，均方根误差（RMSE）为4.4389，决定系数R^2达0.9130。较次优通用模型，〖mAP(P)〗_(50-95)提升0.29个百分点，MAE与RMSE分别显著降低8.4%和7.9%，R^2提升1.6%。消融实验验证了所提模块的协同有效性；跨场景应用测试验证了模型对参照物的宏观泛化能力，并明确了跨物种形态学差异的局限。【结论】Height-YOLO模型有效克服了复杂农田环境下微小目标的特征衰减与定位漂移，在保持轻量化的同时显著提升了株高测量精度，为无人机边缘端实时表型监测提供了可行方案。

Abstract:

[Purpose] In the domain of precision agriculture, high-throughput phenotyping is essential for crop breeding and field management. Traditional unmanned aerial vehicle (UAV) three-dimensional reconstruction methods for measuring crop plant height are fundamentally constrained by high computational power requirements for point cloud generation and significant ground elevation inversion errors caused by dense crop canopies. Concurrently, when existing lightweight deep learning networks are applied to two-dimensional key point localization tasks, they frequently encounter severe bottlenecks. Specifically, they are prone to high-frequency feature loss during continuous downsampling and suffer from localization drift caused by severe homogeneous background interference. To address these critical limitations, this study proposes a high-precision, robust, and lightweight crop plant height detection method specifically optimized for edge computing applications in complex field environments. [Method] First, a comprehensive joint dataset comprising UAV-captured high-resolution RGB imagery and E-shaped physical reference scales was meticulously constructed, covering the complete growth stages of field-grown rice and wheat crops under diverse illumination conditions. Based on this dataset, a linear scaling algorithm for plant height calculation was designed utilizing a direct spatial pixel mapping strategy derived from precisely detected keypoints. Subsequently, a novel lightweight convolutional neural network, designated as Height-YOLO, was proposed based on the structural enhancement of the YOLOv8m-pose baseline architecture. To tackle the issue of microscopic feature attenuation, a High-Pass Perception Module (HPModule) was innovatively designed, which synergistically enhances the extraction capability of microscopic apex features through adaptive high-pass filtering combined with a channel-spatial joint attention mechanism. Furthermore, a Color Contrast Attention (CCA) module was introduced into the deep feature layers. The CCA module effectively utilizes local multi-scale dilated convolutions and explicit differential modeling to significantly improve the discriminability of target boundaries against highly homogeneous agricultural backgrounds. [Result] Extensive comparative experiments demonstrate that the proposed Height-YOLO achieves exceptional state-of-the-art performance. The core key point localization precision mAP(P)_(50-95) reaches 0.9879, the mean absolute error (MAE) is minimized to 3.5265, the root mean square error (RMSE) is recorded at 4.4389, and the coefficient of determination R^2 achieves 0.9130. Compared with the suboptimal general-purpose baseline model, Height-YOLO achieves a notable improvement of 0.29 percentage points in 〖mAP(P)〗_(50-95), while simultaneously reducing the MAE and RMSE by a significant margin of 8.4% and 7.9%, respectively, alongside a 1.6% increase in the R^2 metric. Comprehensive ablation studies rigorously verified the synergistic effectiveness of the proposed functional modules. Moreover, cross-scenario zero-shot application tests in field cotton cultivation environments validated the model''s robust macroscopic generalization capability in identifying standard reference objects, while objectively identifying the inherent localization bottlenecks caused by cross-species morphological phenotypic differences between grass and broadleaf crops. [Conclusion] The proposed Height-YOLO model successfully overcomes the persistent challenges of microscopic feature attenuation and severe localization drift for small agricultural objects. By striking an optimal balance between maintaining a lightweight architectural profile and significantly enhancing the overall measurement accuracy, this research provides a highly feasible, efficient, and reliable technical solution for real-time crop phenotypic monitoring deployed directly on UAV edge computing devices.