基于YOLO-PLM的撒播豌豆苗田间杂草检测与区域密度估计多任务模型

Multi-task Model for Weed Detection and Regional Density Estimation in Broadcast Sown Pea Seedling Fields Based on YOLO-PLM

摘要:

针对撒播豌豆苗田间复杂背景下杂草目标尺度小、分布密集、遮挡严重，且现有检测方法难以为变量除草提供区域分布信息的问题，本研究提出一种基于改进YOLO 11n的杂草检测与区域密度估计多任务模型YOLO-PLM。在YOLO 11n基础上，引入区域密度估计分支，使模型最终在完成杂草目标检测的同时输出区域密度热力图，为分区喷施和变量除草提供区域尺度依据;增加P2小目标检测层，以增强模型对小尺度杂草目标和局部聚集模式的感知能力;采用轻量化自适应特征提取模块（Lightweight adaptive extraction，LAE）替换部分下采样卷积，并引入MBConv（Mobile inverted bottleneck convolution）模块对检测头进行轻量化重构，以降低模型参数量与浮点运算量。试验结果表明，YOLO-PLM精确率、召回率、mAP@0.5和mAP@0.5:0.95分别达97.4%、92.9%、96.9%和83.4%，较YOLO 11n-dens分别提高1.9、4.0、2.0、5.3个百分点;区域密度估计MAE和RMSE分别为0.762、1.066株/区;模型参数量和浮点运算量分别为1.95×10^6和7.7×10^9。与YOLO v8n-dens、YOLO v9t-dens、YOLO v10n-dens、YOLO 11n-dens和YOLO 12n-dens模型相比，YOLO-PLM在检测精度、区域密度估计能力和轻量化方面均表现出较优的综合性能。研究结果表明，本文所提模型在撒播豌豆苗田间复杂背景下具有较好的杂草识别与区域分布表征能力，可为田间精准变量除草作业提供技术支撑。

Abstract:

Aiming to address the challenges of small-scale, densely distributed, and heavily occluded weed targets in the complex background of broadcast sown pea seedling fields, as well as the difficulty of existing methods in providing regional distribution information for variable-rate weeding, YOLO-PLM, a multi-task model for weed detection and regional density estimation was proposed based on an improved YOLO 11n framework. A regional density estimation branch was introduced to enable the model to simultaneously detect weeds and generate regional density heatmaps, thereby providing regional scale information for zone specific spraying and variable rate weeding. In addition, a P2 small-object detection layer was incorporated to enhance the perception of small weed targets and local clustering patterns. To reduce computational cost, part of the downsampling convolutions was replaced with a lightweight adaptive extraction (LAE) module, and the detection head was lightweight reconstructed by using the mobile inverted bottleneck convolution module (Detect_MBConv). Experimental results showed that YOLO-PLM achieved a precision of 97.4%, recall of 92.9%, mAP@0.5 of 96.9%, and mAP@0.5:0.95 of 83.4%, which was improved by 1.9, 4.0, 2.0, and 5.3 percentage points, respectively, over that of YOLO 11n-dens. The proposed model also achieved an MAE of 0.762 plants per region and an RMSE of 1.066 plants per region for regional density estimation, with 1.95×10^6 parameters and 7.7×10^9 FLOPs. Compared with YOLO v8n-dens, YOLO v9t-dens, YOLO v10n-dens, YOLO 11n-dens, and YOLO 12n-dens, the proposed model showed superior overall performance in detection accuracy, regional density estimation, and lightweightness. The results indicated that YOLO-PLM can provide effective technical support for precision variable-rate weeding in broadcast sown pea seedling fields.