Abstract:
Rice plant diseases has been a growing concern in recent years. Different kinds of rice plant diseases cause significant
damage to the rice plants which results in reduced production of rice yield. Early detection of rice plant disease is crucial for crop
protection system. However, the conventional disease detection techniques used by farmers are not highly accurate to identify the
rice plant diseases in timely manner. Recent developments in Convolutional Neural Networks (CNN) has greatly improved the image
classification accuracy and hence they are particularly useful in various plant disease detection. In this study, we proposed an
optimized CNN architecture based on depthwise separable convolutions to identify various rice plant diseases. After collecting 12
types of disease affected rice plant images along with healthy rice plants from different rice fields of Bangladesh the images were
preprocessed and augmented by which a dataset of 16770 images has been constructed. Along with the proposed CNN model,
different lightweight state-of-the-art CNN architectures have been used on the dataset and results have been analyzed. The
experimental analysis indicates that MobileNet v2 architecture provided the best validation accuracy of 98.7% among the all state-ofthe-art CNN architectures. The proposed lightweight CNN architecture outperformed all the other state-of-the-art CNN architectures
with a testing accuracy of 96.3%. Considering a small parameter size, it is evident that the proposed Convolutional Neural Network
model performed significantly well in detecting the rice plant diseases accurately.