Enhancing Agricultural Health: A Deep Learning Model for Plant Disease Detection

In this work, we have presented a Multi-Fusion Deep Learning Framework for plant disease diagnosis by addressing the challenge of accurate detection using multiple types of sources. The current research combines several image-based input data from the Plant Village dataset and environmental and field data (e.g., weather and soil conditions). The Hybrid Feature Extraction method combines pre-trained models with handcrafted features that incorporate Histogram of Oriented Gradients. Additionally, the multi-branch CNN architecture is further developed with a spatial attention mechanism on diseaseaffected regions. Furthermore, the ensemble of models is based on the use of weighted averaging and stacking to enhance the quality of prediction. The model offers a superior application of transfer learning to adjust to new patterns of illnesses based on the type of disease and incremental learning is added to make the model innately adjusted over time. Furthermore, explainable AI methods, like Grad-CAM and SHAP, are applied to the model to explain a rational choice. The framework is evaluated on Accuracy, Precision, Recall, and F1 score as the evaluation matrices along with real-world validations via field trials. This approach offers a holistic solution to the efficient and scalable implementation of plant disease detection in Precision Agriculture.