In this work, we provide a novel approach to tracking the progression of Bacterial Leaf Blight (BLB) disease in rice, which poses significant risks to rice production due to its rapid spread and extensive damage potential. Our proposed system addresses this challenge by pinpointing diseased leaf regions in images, facilitating better disease stage analysis. A custom data set of 6656 rice leaf images with symptoms of bacterial leaf blight caused by natural infection and artificial inoculation was collected. It is divided into five severity classes, from stage 1 to stage 5, based on the diseased leaf area. We initially evaluated five pre-trained deep CNN models (DenseNet-121, Xception, Inception-ResNet, MobileNet-v2, and NasNet-Mobile) for BLB identification without preprocessing. Subsequently, unsupervised segmentation methods (K-means, edge-based, contour-based, local binary pattern (LBP), and color saliency) were applied to preprocess the images, and the CNN models were re-evaluated. Our results demonstrated that MobileNet-v2, combined with the color-saliency segmentation method, achieved the highest test accuracy of 91.5%. This approach excelled in highlighting the color variations associated with different stages of severity of BLB. To enhance segmentation quality, a hard attention mask was generated, and regions of interest (ROIs) were extracted from segmented images. This extensive preprocessing step significantly improved the data, leading to substantial improvements in classification accuracy. Furthermore, our custom data set developed for this research will be published for further studies in this field, encouraging future advances in the management of agricultural diseases.