The weed detection tool has undergone substantial updates for enhanced functionality. It employs classical computer vision techniques and open-vocabulary models for weed detection in diverse plantations. The workflow starts with the X-Decoder model identifying "soil" or "ground" areas, crucial for targeting weeds and excluding irrelevant sections like bushes or the sky. GroundingDINO is then used for open-vocabulary detection, enabling the identification and classification of weeds without pre-defined vocabulary. Finally, SAM-HQ segments bounding boxes to precisely define weed areas.

The tool's adaptability is demonstrated in various contexts, where it assists a robot developed for Rumex removal. Validation with a public dataset of over 4,000 weed images, representative of rapeseed crops, shows impressive results: an F1 score of 0.96, precision of 0.955, and recall of 0.969, indicating the approach's effectiveness using both zero-shot open-vocabulary models and specifically trained models. This application marks a significant advancement in weed detection and management, offering a flexible and precise solution for modern agricultural challenges.