A unified framework with U-Net + + and CNN-RNN-BiGRU architectures for automated weed detection in precision agriculture using AI

Abstract The automated weed identification is necessary to enhance crop yield and the sustainable precision agriculture. The use of manual labour and chemical herbicides in traditional practices makes them very expensive, harmful to the environment, and causes more herbicide resistance. To overcome, this paper presents an AI-driven model AgroWeedX-Ensemble, which combines enhanced preprocessing, state-of-the-art segmentation, multi-scale feature extraction, feature selection, and efficient weed detection. An enhanced WeedNet-Adaptive Pre-Processing Optimizer (WN-APO) improves the quality of data by dynamically changing bilateral filtering, augmentation, and normalization to accommodate lighting changes and shadows. The proposed Attention-ASPP Enhanced Hybrid Dilated Network (AA-HDN) is based on the U-Net + + , attention, ASPP, and hybrid dilated convolutions to precisely differentiate between overlapping weed-crop areas. Multi-Scale Residual Spatial Feature Extractor (MS-RSFE) uses HOG, Gabor filters, and ResNet-50 with Feature Pyramid Networks to identify a variety of morphologies of weeds. The Parrot-Wheel Feature Selector (PWFS) is used to reduce feature redundancy by combining Parrot Optimizer and Binary Waterwheel Plant Optimization. Lastly, WeedAttnX-Net is a CNN-based model that adds RNN, Bi-GRU, attention layers, and the Bi-GRU-Attention to lower the false positives and enhance the reliability of the detection. The experimental findings show that AgroWeedX-Ensemble is highly accurate (0.99), precise (0.98), and robust in different field conditions with a lower false positives (0.02) and false negatives (0.009). The suggested system provides a scalable and useful solution to automated weed detection in precision agriculture.