CMRNet: An Automatic Rapeseed Counting and Localization Method Based on the CNN-Mamba Hybrid Model

文献类型： 外文期刊

第一作者： Li, Jie

作者： Li, Jie;Yang, Chenbo;Zhu, Chengyong;Qin, Tao;Tu, Jingmin;Wang, Binhui;Yao, Jian;Qiao, Jiangwei

作者机构：

关键词： Autonomous aerial vehicles; Crops; Vegetable oils; Flowering plants; Feature extraction; Training; Deep learning; Accuracy; Location awareness; Resistance; Counting; localization; lodging; rapeseed; state space model

期刊名称：IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING （ 影响因子：5.3； 五年影响因子：5.6 ）

ISSN： 1939-1404

年卷期： 2025 年 18 卷

页码：

收录情况： SCI

摘要： Lodging a major agricultural issue, significantly compromises the yield, stability, and quality of oilseed crops, particularly rapeseed (Brassica napus L.). Real-time monitoring and accurate assessment of lodging are critical for precise yield estimation and the development of lodging-resistant varieties. However, traditional methods for quantifying lodging rates, which rely on manual measurements of lodged plant proportions, are often labor-intensive and prone to inaccuracies, limiting their utility in large-scale breeding programs. This article provides an indirect method for lodging assessment by simplifying the lodging issue to the enumeration of upright plants. First, we use a deep learning model for plant counting from Unmanned aerial vehicle (UAV) imagery in plot level. A novel CMRNet model is developed for upright plants counting and localization, leveraging a hybrid CNN-Mamba backbone network. The model synergizes local feature extraction via CNN with the global modeling strengths of the Mamba state space model, yielding semantically rich features while significantly enhancing computational efficiency and inference speed. Then, we created a new Upright Rapeseed Center Point (URCP) dataset using high-altitude UAV remote sensing orthoimages, encompassing rapeseed fields at various maturity stages and lodging degrees. Training and validation of CMRNet on the URCP dataset yielded exceptional performance metrics, with mean absolute error (MAE) of 5.70, relative root mean square error (rrMSE) of 8.08, and coefficient of determination (R-2) of 0.9220. These results significantly outperformed existing TasselNetV2, RapeNet, and RPNet models. The number of parameters in our model is only 7.94 M, which is lower than SOTA counting networks. In addition, we also verified the robustness on different rape materials in two years, 2023 and 2025, and the R-2 were all above 0.8, indicating that the model should cope with different field conditions.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序