Smart Agro-Weed Eradication with AI Driven Weed Detection and Removal Using Vision Transformers and Neuromorphic Computing

Efficient weed management is essential for improving the productivity and sustainability of crops cultivation. The swift rise of herbicide-resistant weeds has highlighted the necessity for novel strategies to tackle the difficulties related to accurate weed identification. Conventional methods of weed eradication, like manual labor or pesticide application, often demand considerable effort, entail substantial costs, and may adversely affect the environment. Conventional machine learning methods necessitate substantial labeled datasets and encounter difficulties with real-time processing. This research introduces an AI-driven method for weed detection and eradication, employing the Crop and Weed Detection Data with Bounding Boxes dataset to train a Vision Transformer (ViT)-based model for accurate classification. In contrast to conventional CNN s, ViT effectively captures long-range dependencies in images, enhancing feature extraction for intricate weed-crop discrimination. An Active Learning (AL) architecture is implemented to reduce manual labeling efforts by choosing only doubtful samples for human annotation. This diminishes the labeling burden while enhancing model generalization to novel weed species. A neuromorphic computing-based robotic system is utilized for real-time weed eradication, utilizing low-power spiking neural networks (SNNs) for expedited decision-making in the field. Proposed model achieves 99.69% of accuracy, 98.29% of precision, 98.04% of recall, 98% of F1-Score and 1.6J of Energy consumption.