Rice plant disease classification using dilated convolutional neural network with global average pooling

The Indian economy is thought to be primarily dependent on agriculture. In plants with various climatic circumstances, illness is highly prevalent and natural. As a result, the quality of crop deteriorates. Getting the best quality and quantity of harvest is farmers' most challenging task due to recent changes in weather cycles. Crop diseases must be identified and prevented as soon as possible to improve productivity. Deep learning is an artificial intelligence branch. It has been actively discussed in academic and industrial circles in recent days, because of the advantages of autonomous learning and feature extraction. The use of learning-based algorithms in plant leaf disease recognition can help to strengthen the objectivity of plant leaf disease feature extraction, minimize the limitations of intentionally selecting disease spot features, and speed up the study. In this paper, we examine existing approaches to detecting plant leaf disease using deep learning and high-end imaging methods, as well as their challenges. We anticipate that our research will be useful to researchers interested in plant disease identification. The traditional CNN has the issue of using up too much computational power. To address this issue, this research developed a DCNN (Dilated Convolutional Neural Network) model with Global Average Pooling (GAP), which will be constructed by changing regular CNN convolution kernels with dilated convolution kernels and the fully connected layer in traditional CNN replaced by Global Average Pooling. The dilated convolution gives the advantages of less computational cost and reduced memory usage then GAP avoids overfitting. These two new concepts are implemented with CNN and the results of this method is compared with other learning and hybrid learning methods using performance metric such as precision, recall, f1-score and accuracy. The classification includes four classes such as bacterial blight, blast, brown spot and turgo. The performance metrics shows that, in the same experimental setup, the DCNN model with GAP improves training accuracy by 5.49 percent on average, compared to the classic CNN model and the results are compared with other learning and hybrid learning methods.