| dc.description.abstract | The early detection of paddy disease is essential for reducing the usage of chemical substances and pesticides, and preventing local and global transmission of diseases. An automated paddy disease diagnosis system makes this possible, and helps to improve crop production and the overall health of rice plants. For this study, we have explored the most prevalent paddy diseases and their visible symptoms. Additionally, the four segmentation models (UNet, VGG16 UNet, TransUNet and Deep Residual UNet) and four classification models (DenseNet, MobileNet, Vision transformer (ViT) and a custom ensemble model of DenseNet121 and Xception), have been reviewed with a comparative analysis highlighting their structural differences, advantages, and limitations. The significant research gaps in this domain have been identified and to address the lack of open-access paddy disease segmentation datasets, we have created a novel paddy disease segmentation dataset using image processing techniques. The applicability of the above-mentioned segmentation models has been evaluated for paddy diseases using this newly created dataset and we have identified that Deep Residual UNet is the most suitable model to be used in resource constraint applications—considering its quantitative and qualitative performance, model size and structural advantages and limitations. Furthermore, we have investigated the impact of training these models with a significantly higher number of augmented images—more than double the original dataset—and observed that while the quantitative performance increased with increased data, the qualitative performance degraded in a few cases. Moreover, due to the huge computational requirements and the data-hungry nature of ViTs, we assessed whether its performance could be achieved with traditional models or their ensembles and found it to be feasible. Additionally, we have explored the effect of augmentation intensity on the abovementioned classification models. | |
