Abstract: Rice is one of the most crucial crops worldwide in modern agriculture. The current diseases (such as rice blasts) have seriously threatened the superior quality and yield of the rice in recent years. However, manual disease monitoring cannot fully meet the large-scale rice cultivation in the extensive fields. Furthermore, the mechanical equipment for disease detection can frequently cause unnecessary harm to the rice plants. As a result, it is highly required for the intelligent and large-scale detection of diseases to reduce the labor costs for the high precision of monitoring. Deep learning models (such as GoogLeNet) can be expected to accurately identify the rice blast. Nevertheless, an optimal balance between detection accuracy and processing speed is very necessary for GoogLeNet in the practical applications of rice blast detection. This study aims to classify and detect the rice blast diseases at the seedling stage using an improved GoogLeNet model. Particularly, the rice was vulnerable to diseases during the tillering stage. Specifically, the research targets were also collected from the rice images in the critical period. The disease features were determined to significantly influence the subsequent rice growth using refined models. Several key steps were also involved: Initially, a comprehensive and diverse dataset of rice blast images was obtained after field research, expert consultations, and advanced techniques of image capture, followed by image processing. Subsequently, data augmentation (including rotation and brightness adjustments) was employed to obtain the final dataset with 2,000 rice blast images. An attention mechanism was then integrated into the GoogLeNet model. The distinct features of rice blasts were focused on after optimization. An ablation test was conducted to assess the effectiveness of the improved model. Comparative experiments were also performed to illustrate its advantages. The results indicated that this modification significantly improved the detection accuracy. Specifically, the attention mechanism module (GoogLeNet+DSCAM) was incorporated to increase the recall by 9.29 percentage points, compared with the original model. The attention mechanism has effectively enhanced the detection of critical information. Furthermore, the improved GoogLeNet model also surpassed the original model, in terms of all evaluation metrics. The better performance was achieved to improve by 15.33, 15.80, and 15.61 percentage points in the precision, recall, and F1 score, respectively, Thereby the refined network architecture substantially enhanced the detection performance. The superiority of the improved GoogLeNet was observed in the tasks of rice blast classification. Several widely-recognized classification models (including AlexNet, ResNet, VGG, and the original GoogLeNet) were selected as the benchmarks for comparison. The rice blast datasets and an independent test set were utilized to fully train and then evaluate these models. Meanwhile, the comparative analysis showed that there were distinct advantages and practical efficacy of the improved GoogLeNet model. Such classification also demonstrated some challenges. The enhanced GoogLeNet model exhibited exceptional performance overall evaluation metrics, indicating a marked superiority over AlexNet, ResNet, VGG, and the original GoogLeNet. Specifically, the notable improvements were also achieved by 16.11, 17.20, 25.95, and 15.33 percentage points in the precision, respectively; There were the advantages of 17.91, 20.24, 29.67, and 15.80 percentage points, respectively, in the recall; There were an even more pronounced increases of 17.07, 18.67, 27.94, and 15.61 percentage points, respectively, in the F1 score. These significant performances can provide high efficacy to enhance the accuracy and the efficient speed of the detection. This finding can also offer valuable insights into preventing and controlling rice diseases.