Deep Learning-Based Evaluation of Models for Rice and Sorghum Weed Classification Using the MFWD Dataset

Precision agriculture aims at optimizing resource utilization and increasing crop yield while minimizing environ-mental impact. Correct weed classification is the foundation of accurate weed management, avoiding the misuse of herbicides, and optimizing sustainability. Traditional approaches, such as mechanical eradication and broad-spectrum herbicide application, are wasteful and labor-intensive. Deep learning techniques specially Convolutional Neural Networks or CNNs which offers a computer-based approach by learning discriminative features from images. The study here aims at investigating the imple-mentation of deep learning models for the purpose of effective weed species classification which is done using an extensive data of 30 species among which 5 balanced species is selected for evaluation. Based on proper examination and trails deep learning architectures, MobileNet v2, ResNet 50, InceptionV3, EfficientNet- V2, Swin 1 are used for evaluating their effectiveness in classifying the weed species. The pre-trained models were further fine-tuned using the transfer learning technique with the primary goal of enhancing the feature extraction that addresses the potential problems of under- and over-fitting. The results obtained showed that among the models used, Swin 1 had given the highest accuracy value of 97.15, demonstrating that the model's superior capability in generalizing compared to other tested architectures.