A hybrid atrous CNN-SVM model for rice leaf disease classification

Abstract

Rice leaf diseases can seriously hurt crop yields, so spotting them early and accurately is crucial for keeping things under control. In this work, we put forward a hybrid deep learning approach called Atrous CNN-SVM, which takes a pre-trained VGG19 setup and boosts it with atrous (or dilated) convolutional layers to pull out features across different scales, then pairs it with a Support Vector Machine for reliable classification. We put the model through its paces using a practical dataset gathered from Dong Thap in Vietnam, and double-checked it on an outside dataset. The experiments revealed that Atrous CNN-SVM topped out at an accuracy of 86.49% when trained on 90% of the local data, doing better than plain CNN-SVM, MobileNet, and classic models relying on hand-engineered features. Overall, this points to the real value of blending atrous convolutions with SVM for automatic rice disease detection, and it drives home how vital it is to weave in location-specific data for precision farming.