IFSA-Inception-CBAM: An Early Detection Model for Rice Blast Disease Based on Integrated Feature Selection and a Deep Convolutional Neural Network

Rice blast disease is one of the most contagious and destructive diseases affecting rice, posing a serious threat to global rice production and the agricultural economy. To enable accurate early detection under field conditions, this study proposes an integrated feature sorting algorithm (IFSA). The algorithm integrates five spectral feature selection methods—partial least squares, successive projections algorithm (SPA), principal component analysis loading (PCA-Loading), genetic algorithm (GA), and random forest (RF)—and employs the Borda count method for comprehensive feature ranking and selection. Field experiments were conducted in Haicheng, Anshan, Liaoning Province, China, using the rice cultivar Yanfeng 47. A total of 4893 hyperspectral samples were collected under natural field conditions. The results demonstrate that IFSA effectively identifies key spectral wavelengths for the early diagnosis of rice blast disease, achieving significantly higher detection accuracy than conventional single-method dimensionality reduction approaches. Based on the IFSA-selected wavelengths, an early detection model (Inception-CBAM) was further developed by integrating a multi-channel convolutional neural network with a convolutional block attention module, thereby enhancing the extraction and recognition of early disease-related features. Compared with six baseline models (InceptionV4, ResNet, BiGRU, RF, support vector machine, and extreme learning machine), Inception-CBAM achieved an overall accuracy of 95.44 ± 0.50% and a Kappa coefficient of 93.92 ± 0.67% for early rice blast disease detection, outperforming all competing methods. This study confirms the effectiveness of IFSA for hyperspectral feature selection and demonstrates that the proposed Inception-CBAM model provides strong capability for early disease detection. Nevertheless, the data were collected from a single cultivar and a single region; therefore, the model’s generalization performance across broader environments requires further improvement. Future work will extend the evaluation to multi-cultivar and multi-region scenarios to facilitate practical deployment for real-time field diagnosis.