Multi-crop plant counting and geolocation using a YOLO-Powered GUI System

• Developed a multi-crop counting method using an AI model with up to 99% accuracy. • Improved detection accuracy by segmenting orthomosaic images into smaller tiles. • Converted bounding box coordinates into GPS locations for precise field mapping. • Advanced precision farming by minimizing manual counting and enhancing decision-making. Crop counting has traditionally relied on manual field assessments or complex machine learning algorithms, which often struggle to identify small objects or underestimate the total count of objects present in the field. This study proposes a multi-crop graphical user interface integrated with an object detection model to accurately detect and count objects in the field while converting bounding box coordinates into global positioning system (GPS) coordinates. We utilized tiles extracted from an orthomosaic. The system consists of four core modules: The first module splits the orthomosaic into smaller tiles with adjustable overlap and sizes. The second module uses an object detection model to process each tile and saves the detections in .txt files. The third module transforms bounding box coordinates into real-world GPS locations using tile metadata and position within the original orthomosaic. The fourth module eliminates duplicate detections by applying buffer zones around neighboring detections, merging overlapping instances, and assigning a single centroid for each cluster. Three crops (tobacco, strawberry, and watermelon) were evaluated using two tile overlap configurations (non-overlap and 20% overlap between tile edges) and three tile sizes (640, 1280, and 2048 pixels). A YOLOv11x model was trained for each crop, and the methodology's accuracy was assessed by comparing the total objects identified by the models against the actual number of objects in the field. The results demonstrated that the proposed methodology accurately counted objects in the field with an accuracy ranging from 91% to 99%. The 640-pixel overlapped tile approach resulted in the highest occurrence of multi-detections for the same object. In contrast, the most accurate count was obtained using the non-overlapping approach with 2048-pixel tiles. This study demonstrates the feasibility of applying this method for different crop systems, reducing the reliance on manual counting and improving decision-making for precision farming practices.