An evaluation method for northern corn leaf Bligh disease in northern china based on reprocessing multi-source habitat information

• Integrates multi-source datasets for comprehensive Northern Corn Leaf Blight (NCLB) assessment. • Establishes a county-level habitat suitability model for NCLB based on epidemiological features. • Improves disease severity classification accuracy using an uncertainty-guided correction module. • Predicts spatial risk distribution patterns across maize-growing regions in northern China. With changes in global climate and farming systems, the outbreak of Northern Corn Leaf Bligh (NCLB) disease in Northern China becomes increasingly more frequent, posing a serious threat to corn production. However, the risk evaluation of NCLB is still facing a range of problems, such as insufficient remote sensing data diversity, low model accuracy, and intra-class heterogeneity. Aiming at these issues, a county-level habitat suitability evaluation model for NCLB (Stacking_MUGLSC) was established based on multi-source data and a Multi-Class Uncertainty-Guided Local Synergistic Correction (MUGLSC) postprocessing module. Then, a multidimensional feature space was constructed by fusing remote sensing temporal characteristics, dynamic meteorological factors, variety resistance information, disease occurrence data, and epidemiological patterns. Subsequently, Pearson correlation analysis and the Relief-F algorithm were utilized to screen for ecological variables including the temperature variation rate from August to September, the vapor pressure variation rate from July to August, and variety disease resistance. Under the Stacking ensemble framework, the classification performance of the proposed model was significantly improved by employing Logistic Regression (LR) as a meta-model, combining the advantages of multiple Machine Learning (ML) algorithms. The experimental results show that the Stacking_MUGLSC model outperformed other ML models in terms of AUC (0.97), OA (0.85), and F1 Score (0.88). SHAP and LIME interpretability analyses further demonstrated that the selected multi-source data are closely correlated with the distribution of NCLB. Overall, the model’s prediction results were highly consistent with the actual disease distribution, confirming that the fusion of multi-source data can effectively characterize the spatiotemporal occurrence pattern of NCLB. The evaluation framework established in this paper not only provides a scalable technical foundation for crop disease early warning, but also reveals a multi-source data interaction mechanism for optimizing farmland management strategies. Our findings are of important practical value for ensuring food security over the world.