Deep learning-driven intelligent weed mapping system: Optimizing site-specific weed management

In response to the issue of weed control in agricultural practices and the demand for Site-Specific Weed Management (SSWM), this study proposes a low-altitude orthophoto weed map generation system for agricultural fields based on deep learning and image processing techniques, which aims to provide accurate and timely weed maps for early-stage weed control. By developing an end-to-end Intelligent Weed Mapping System (IWMS) and implementing the proposed AgroYOLO model, rapid and effective monitoring of weeds and crops in agricultural fields is enabled. Due to the significant differences in the number of weeds and crops in the field leading to an imbalanced sample distribution, we designed an Adaptive Copy-Paste Algorithm (ACPA) to address the issue of sample imbalance in multi-object detection. In addition, this study investigates the effects of different ground sampling distances (GSD) on the detection performance. Experimental results show that the designed system achieves an average accuracy of 94.3 % on a custom dataset, with an average processing time of 85ms per square meter of the ground mapping area, thereby improving the effectiveness of weed detection and control and providing support for sustainable agricultural field management.