Performance Evaluation of Hybrid Deep Learning Architectures for Plant Disease and Severity Classification

For sustainable agriculture and food security, it is crucial that diseases of crops are correctly identified along with the severity. With the increasing availability of annotated image datasets and computational resources, deep learning has become a promising solution to automate plant health surveillance. Summary This paper provides a through performance assessment of well-known hybrid deep learning-based architectures for plant disease and severity classification. The study covers a variety of models including CNN-based, CNN-LSTM hybrids, attention mechanisms and lightweight object detection architectures like YOLO and EfficientNet derivatives. We evaluate our methods using several benchmark datasets as well as field-acquired datasets of rice, cotton, tomato and sorghum. To evaluate these models for different disease types and severity stages, performance measures of accuracy and F1-score are utilized to compare between them. Experiments show that it often outperforms the pure CNN counterparts, especially for severity detection on multi-stage diseases. This paper also discusses the shortcomings of current methods and identifies promising research directions to further improve generalization, interpretability, and real-time application of models. The results are expected to help researchers and developers to choose the right architecture for precision agriculture applications.