文献类型： 外文期刊

作者： Sheng, Haiyuan ¹ ; Yao, Qing ² ; Luo, Ju ³ ; Liu, Yongjian ² ; Chen, Xiangfu ² ; Ye, Zhongru ¹ ; Zhao, Tiezhuang ⁴ ; Ling, Heping ⁵ ; Tang, Jian ³ ; Liu, Shuhua ³ ;

作者机构： 1.Zhejiang Sci Tech Univ, Sch Informat Sci & Technol, Hangzhou 310018, Peoples R China

2.Zhejiang Sci Tech Univ, Sch Comp Sci & Technol, Hangzhou 310018, Peoples R China

3.China Natl Rice Res Inst, State Key Lab Rice Biol, Hangzhou 311401, Peoples R China

4.Superhexa Century Technol Co Ltd, Beijing 100085, Peoples R China

5.Jinhe Nengyuan Technol Co Ltd, Changzhou 213200, Peoples R China

关键词： Rice planthopper; White flat plate; Intelligent survey; AR glasses; Deep learning; Object detection

期刊名称：COMPUTERS AND ELECTRONICS IN AGRICULTURE （ 影响因子：8.3； 五年影响因子：8.3 ）

ISSN： 0168-1699

年卷期： 2024 年 218 卷

页码：

收录情况： SCI

摘要： Rice planthoppers usually refer to brown planthoppers (Nilaparvata lugens), white-backed planthoppers (Sogatella furcifera), and small brown planthoppers (Laodelphax striatellus). Different species, wing types, and developmental stages of planthoppers often aggregate to suck rice sap or lay eggs at the stem base of rice plants, affecting rice quality and yield. The long-winged brown planthoppers and white-backed planthoppers can migrate longdistance with air currents. Real-time monitoring of population quantities of different species and developmental stages of rice planthoppers is a prerequisite for timely and effective prevention and control to ensure food security. At present, monitoring of planthoppers in fields still relies on surveyors going to paddy fields to collect planthoppers onto flat plates by patting rice plants, and then manually counting and recording the quantities of different species and developmental stages of rice planthoppers. This approach has several problems with low efficiency, low accuracy when dealing with a large number of rice planthoppers, and difficulty in tracing the survey results. To address the above-mentioned problems, we propose an innovative method that uses AR glasses with high-definition cameras to capture images of rice planthoppers on white plates and use deep learning methods to automatically identify and count different species and developmental stages of planthoppers. Due to the large number of non-target impurities, the tiny juvenile nymphs, and the imbalanced numbers of multiple planthopper classes in images, some deep-learning models have poor detection performance. To overcome these challenges, we developed a Cascade-RCNN-PH model. An adaptive positive sample matching algorithm was proposed to help small targets match the best anchors. A squeeze-and-excitation feature pyramid network (SEFPN) module was used to enable low-level feature maps to directly acquire comprehensive high-level semantic information, which improves the model's ability to distinguish features between targets and interference targets. Compared with the models of YOLO and Faster-RCNN, Cascade-RCNN-PH performed much better. The average recall rate of Cascade-RCNN-PH for detecting 10 different species and developmental stages of rice planthoppers was 83.43 %, and the average precision rate was 83.56 %. The combined method of AR glasses and AI models that we proposed for field surveys of rice planthoppers improves the survey efficiency of planthoppers, reduces manpower by half, significantly improves planthopper counting accuracy in high-density situations, and reduces the professional dependence of surveyors. This method can also be applied to the survey of other pests, such as aphids, whiteflies, etc.