COMPARATIVE ANALYSIS OF MACHINE LEARNING AND DEEP LEARNING MODELS FOR PLANT DISEASE DETECTION

Abstract

Diseases in plants remain a serious hazard to the world’s food supply, farming results and environmental protection, especially in farming-reliant areas. Diagnosing plant diseases using old methods mostly involves looking at plants and this can be slow, subjective and easily incorrect. Because more people are being born, there is more demand for food which makes crop health and lower losses to diseases extremely important. Using pesticides as pest controls endangers health and damage to the environment and the growing resistance of pathogens has weakened their effectiveness. So, fast and accurate methods for detecting plant diseases are urgently needed. In modern developments in AI, ML and DL, it is now possible to use computers to identify plant diseases through digital image processing. SVM, RF, DT and GB are promising at classifying plants from their leaf images. These simple models depend on color, texture and shape brought out by HOG and LBP algorithms. However, although these methods are easy to use and explain, they need domain specialists to create features by hand and often fail with complex visual patterns. CNNs and similar models deliver results regardless of visual differences, since they can discover needed patterns right from the unprocessed images. Examples of these architectures such as VGGNet, ResNet, AlexNet and EfficientNet, have proven better at detecting and naming many plant diseases on datasets like Plant Village. With these models, it’s possible to reuse networks already trained without having much data in your domain. The use of flipping, rotation and brightness adjustment makes models work better and helps them avoid overlearning. In this study different models of DL & ML approaches are tested and assessed for infected plant detection using the Plant Village data in this research. The Objective was to discover the model that best and generally recognized plant leaf diseases through image data. Test accuracy was highest with 99% for the Random Forest classifier and was followed by the DT with 96% and GB with 95%. While the training accuracy for the SVM was high at 96%, its test accuracy fell to just 85%. As part of DL, popular training models were transferred and used in the field of deep learning. ResNet50 performed the best, reaching 96.8% test accuracy, while VGG16 had 95.2% and AlexNet came in at 94.1%. They worked well even with new, unseen data, as a result of using extra data and fine-tuning. Generally, the findings indicated that Ensemble Classifiers and CNN-based CNNs offer the best accuracy when using ML and DL, respectively. Using both DL and ML together will play a major role in the future of agriculture. As a result, they help to automatically recognize and identify plant diseases, aiding the practice of precision farming. Thanks to new developments in data gathering, understanding models and running them on phones, these technologies will greatly help with real time monitoring and crop health care.