Deep Learning Approaches for Weed Species Classification: Efficient Deployment on Edge Device

Abstract

This study focuses on developing and evaluating deep learning models for weed species classification using the DeepWeeds dataset. Ten deep learning architectures were trained using two approaches: freezing convolutional layers and training the entire architecture. Transfer learning with ImageNet-initialized weights was employed to enhance training efficiency. Both multiclass and multilabel classification techniques were implemented, with appropriate dense layers and activation functions tailored for each type. Models such as MobileNet, EfficientNetB0, and DenseNet121 demonstrated high classification accuracy, with EfficientNetB0 achieving the highest multiclass accuracy of 99.7%. This best-performing model was further assessed for resource efficiency and deployed on an edge device for real-time application. The findings highlight the application of deep learning methods to address agricultural challenges, specifically weed species classification, and their potential for real-world implementation.