关键词: 水稻, 施肥, 行为识别, YOLOv8, 检测层, 注意力机制

Abstract: In order to automate the detection and recognition of manual fertilizer application during the rice cultivation process, this study focuses on the recognition of manual fertilizer application in the rice fields at the Digital Agriculture Center in Haishu District, Ningbo City, Zhejiang Province. The research utilizes the YOLOv8 algorithm for behavior recognition and detection of manual fertilizer application in the rice fields. Building upon the YOLOv8 algorithm, a target detection layer is introduced to retain shallow feature information and enhance the perception capability of the network model towards smallsized targets. Additionally, a global attention module is incorporated to improve the emphasis on global feature information by the network model. The research results demonstrate that the improved YOLOv8-GS recognition model achieves an mAP value of 98.4%, exhibiting a 2.4% improvement compared to the original YOLOv8 model. The detection time of each image is 1.7ms. Specifically for the test set featuring small-sized targets, the mAP value reaches 98.6%, indicating a 3.3% improvement from the previous version. These findings validate that the YOLOv8-GS model possesses advantages such as high precision, real-time performance, and multi-scale capabilities, especially for small-size targets with strong detection and recognition ability.

Key words: rice, fertilizer application, behavior recognition, YOLOv8, detection layer, attention mechanism