Multi-scale Patch-Based Extraction via Dual Attention Mechanism for Weed Classification

Abstract

Accurate and efficient weed classification plays a pivotal role in precision agriculture by enabling targeted herbicide application and reducing environmental impact. This study proposes a novel deep learning-based framework for robust weed species classification using the DeepWeeds dataset. The methodology incorporates multi-scale patch-based feature extraction to capture spatially diverse patterns in complex backgrounds, followed by a hybrid CNN for deep feature encoding. To further enhance model discriminability, a dual attention mechanism—combining spatial and channel-wise attention—is employed to emphasize informative regions and feature maps. Additionally, handcrafted features are integrated with deep representations to enrich semantic information. A feature selection and dimensionality reduction stage is applied using Uniform Manifold Approximation and Projection (UMAP) and the fused features are classified using a RandomForest classifier.  Experimental evaluation on the DeepWeeds dataset demonstrates the superiority of the proposed approach, achieving a classification accuracy of 99.59%, with precision, recall and F1 scores exceeding 99% across most weed categories. Comparative analysis shows substantial improvement over traditional CNNs and baseline transfer learning models. This work highlights the effectiveness of combining deep features with expert knowledge through attention mechanisms and selective feature fusion. Future work includes extending the model for real-time inference on embedded edge devices, incorporating temporal weed progression using UAV imagery, and generalizing the framework for multi-crop and multi-season weed classification.