From Pixels to Insights: Harnessing Deep Learning for Accurate Plant Pathology Diagnosis

The growing threat to the world's food supply from plant diseases shows that we really need better and more accurate ways to find these diseases. This research is about using advanced deep learning technology to solve this problem. The study looks at how well different models work for figuring out plant diseases. Some of the models it looks at are AlexNet, DenseNet169, Inception_v3, ResNet34, Squeezenet1_1, and VGG13. The information used includes 38 different plant diseases from the PlantVillage dataset extracted from Kaggle and also a background class from the DAGS dataset. The results show that there are small differences in performance between different types of designs and ways of training. For example, AlexNet is very accurate when it is retrained deeply, reaching 99.24% On the other hand, DenseNet169 performs very well with 99. 72% accuracy when it undergoes deep retraining. In addition to measuring accuracy, the research also includes visualization experiments to help us better understand how the models behave. The occlusion experiment shows heat maps that visually represent how different parts of an image affect the decisions made by the computer program. The saliency map experiment uses math to figure out which pixels are most important. This helps us understand how the model makes decisions. Images for certain plant diseases like Early Blight, Late Blight, and Septoria Leaf Spot are shown using AlexNet and VGG13. In summary, this study compares different deep learning models for spotting plant diseases and also gives important information about how they work. The results help choose the best models and training plans, which improves the creation of strong and trustworthy automated systems for diagnosing plant diseases. This research is an important step in bringing advanced disease detection to farming using the latest technology.