Click on “Download PDF” for the PDF version or on the title for the HTML version.

If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Highlights

A lightweight feature extraction backbone network based on CSPDarknet53, integrated with a lightweight C2f_RVB block, reduces model parameters and computational load without sacrificing the ability to extract texture features of maize seedlings.

Abstract. This paper optimizes the YOLOv8 algorithm and proposes a maize seedling detection model (EMBS-YOLO) based on an efficient multi-branch and multi-scale feature fusion pyramid network. This model can accurately count maize seedlings, evaluate sowing quality, and perform tasks such as gap detection and supplemental planting, thus achieving rapid detection of maize seedling numbers. The model employs a feature extraction backbone network based on CSPDarknet53, integrated with the lightweight C2f-RVB convolutional module, which significantly reduces the number of parameters and computational complexity of the model, while maintaining the target feature extraction capabilities. Furthermore, an efficient multi-branch and multi-scale feature fusion pyramid network (EMBSFPN) is introduced. Compared to the traditional PANet, EMBSFPN utilizes the Channel to Spatial Attention Aggregation Module (CSAAM), which applies weighting to the feature maps extracted by the backbone network in both channel and spatial dimensions, allowing the model to focus more on key features while minimizing interference from irrelevant information. EMBSFPN also increases the depth of the feature fusion network, employing different fusion methods at different network depths, combined with cross-layer connections to achieve deeper fusion between features of different resolutions. Finally, we adopt the Focaler-IoU loss to replace the original CIoU loss, which enhancing the model's detection ability for challenging samples, such as those with blurred boundaries, occlusion, and overlapping objects. Experimental results show that the EMBS-YOLO model improves the spatial features of maize seedlings while reducing network complexity. The model achieves an AP_50-95 of 69.25%, an AP₅₀ of 97.97% and a recall rate of 95.2%. In addition, the number of parameters used is reduced by approximately 33.0% and GFLOPs are reduced by 15.8% compared to YOLOv8s. The EMBS-YOLO model can accurately count maize seedlings and estimate planting density while maintaining low model complexity, providing technical support for managing the maize seedling stage.