Weedex : Deep learning enabled autonomous robot for detection and removal of paddy weeds

This research addresses the challenge of detecting and removing paddy weeds through the integration of computer vision, deep learning, and robotics. A dataset comprising the five weed species, Fimbristylis miliacea, Panicum repens, Cyperus rotundus, Commelina diffusa, and Limnocharis flava, was collected using a digital camera. The study focused on developing a weed removal robot, named Weedex, with two core components: hardware design for mechanical weed removal and software model development for automated weed detection. Once weeds are detected by the software model, Weedex should be able to autonomously locate and mechanically remove the identified weed plants. To enhance model robustness and address class imbalance, data augmentation techniques such as random flips and color/brightness adjustments were applied. Object detection was performed using the algorithms YOLOv8-s, YOLOv8-x, and YOLOv8-seg. Deep feature extraction from detected objects utilized multiple pretrained architectures, including EfficientNet-B0, EfficientNet-B5, ResNet50, MobileNetV2, and a custom convolutional neural network. Model performance was evaluated using precision, recall, F1-score, and accuracy. Results demonstrate that EfficientNet-B5 consistently outperformed alternative feature extractors, while the combination of YOLOv8-seg with EfficientNet-B5 embeddings achieved the best performance (≈ 99% accuracy, F1 ≈ 0.99) on the test data. This improvement was attributed to pixel-level weed localization, reduced background interference in feature extraction, and robust similarity-based classification. This research provides a practical solution for automated weed management, potentially reducing herbicide use and labor requirements, and promoting sustainable and precision agriculture practices.