Accurate Sugarcane Detection and Row Fitting Using SugarRow-YOLO and Clustering-Based Spline Methods for Autonomous Agricultural Operations

Sugarcane is mostly planted in rows, and the accurate identification of crop rows is important for the autonomous navigation of agricultural machines. Especially in the elongation period of sugarcane, accurate row identification helps in weed control and the removal of ineffective tillers in the field. However, sugarcane leaves and stalks intertwine and overlap at this stage. They can form a complex occlusion structure, which poses a greater challenge to target detection. To address this challenge, this paper proposes an improved target detection method, SugarRow-YOLO, based on the YOLOv11n model. The method aims to achieve accurate sugarcane identification and provide basic support for subsequent sugarcane row detection. This model introduces the WTConv convolutional modules to expand the sensory field and improve computational efficiency, adopts the iRMB inverted residual block attention mechanism to enhance the modeling capability of crop spatial structure, and uses the UIOU loss function to effectively mitigate the misdetection and omission problem in the region of dense and overlapping targets. The experimental results show that SugarRow-YOLO performs well in the sugarcane target detection task, with a precision of 83%, recall of 87.8%, and mAP₅₀ and mAP_50-95 of 90.2% and 69.2%. In addition to addressing the problem of large variability in row spacing and plant spacing of sugarcane, this paper introduces the DBSCAN clustering algorithm and combines it with a smooth spline curve to fit the crop rows in order to realize the accurate extraction of crop rows. This method achieved 96.6% in the task, with high precision in sugarcane target detection and demonstrates excellent accuracy in sugarcane row fitting, offering robust technical support for the automation and intelligent advancement of agricultural operations.