Signal, Image and Video Processing Обработка сигналов

YOLOv8-WD: an improved YOLOv8-based multi-class weed detection algorithm

Qie, Yage, Fang, Chao, Huang, Jinghua, Wu, Donghao, Jiang, Jian
Springer Nature · 2026-11-13 · DOI 10.1007/s11760-026-05865-7

Аннотация

Weeds are one of the major threats confronted by agricultural production. Some researchers have proposed using intelligent machines to identify and remove weeds. However, in the complex farmland environment, there are problems such as a vast diversity of weed species, variable lighting conditions, and mutual occlusion of leaves, which render it impossible to detect weeds efficiently and accurately. Therefore, this paper presents a multi-class weed detection method based on the improved YOLOv8 algorithm. This method is capable of efficiently detecting weeds in environments with different weather conditions, locations, and weed densities. To enhance the sensitivity of the model to the local features of weeds, we integrate the Large Separable Kernel Attention (LSKA) mechanism with the Spatial Pyramid Pooling-Fast (SPPF) layer to aggregate feature information from different positions, thereby achieving better feature extraction. Meanwhile, we redesign the detection head by utilizing the Diverse Branch Block (DBB) reparameterization module, aiming to capture more features during the training phase.Finally, the dynamic upsampler DySample is employed to replace the traditional nearest neighbor interpolation method. It captures more features during the training stage, reduces the computational amount, and improves the overall performance. The experimental results show that the mAP0.5 of the improved algorithm reaches 94.5%, which is 3.8% higher than that of the original YOLOv8 model. This validates the effectiveness of the improved algorithm and demonstrates its applicability to the weed identification system in smart agriculture.

Читать у издателя →