Abstract:

[Objective] Rice blast, caused by the fungus Magnaporthe oryzae, is among the most devastating diseases affecting rice production and global food security. Its occurrence and development are affected by rice varieties, meteorological conditions, geographical environments and field management, resulting in complex and unstable epidemic patterns. Traditional prevention and control methods rely on empirical experience and expert judgment, which are limited by scattered knowledge, inconsistent expression and insufficient intelligent decision-making ability. Knowledge graphs can organize agricultural knowledge effectively, but existing reasoning methods cannot handle complex semantic relationships and heterogeneous data well, and often suffer from insufficient feature extraction and oversmoothing problems. To overcome these limitations, a special knowledge graph is constructed for rice blast management and a Residual-AttentionGCN model is proposed. [Methods] Core agricultural entities closely related to rice blast management were focused on in the knowledge graph construction, including rice cultivars, planting regions, meteorological conditions, disease grading levels and field control measures. All entity extraction and normalization procedures were implemented based on the semantic attributes of diverse data sources. Rice cultivar data was mainly sourced from China Rice Data Center, while regional entities were defined based on four permanent experimental rice fields in South China. Meteorological entities were derived from five consecutive years of field observation records during rice growing seasons, including temperature and humidity fluctuation data. Disease severity grading strictly followed the 0–9 standard issued by the International Rice Research Institute (IRRI). Prevention measure entities were summarized and sorted out based on published agricultural research findings and practical guidance from plant protection professionals. The four experimental fields served not only as basic spatial entities but also as comprehensive data carriers, covering planted rice cultivars, daily field management, long-term meteorological records and historical rice blast incidence data. Multiple types of heterogeneous data were interconnected through predefined semantic rules, including cultivation matching, environmental adaptation, disease infection and prevention correlation relations. Such multi-dimensional associations eventually formed a hierarchical semantic network centered on the logical chain of "cultivar–region–climate–disease severity–prevention measure". The final constructed knowledge graph contained 11 521 instantiated nodes and 45 291 relational edges, presenting typical sparse and heterogeneous characteristics of practical agricultural graph data. Aiming at accurate relational reasoning, this research developed the Residual-AttentionGCN model based on classic graph convolutional network architecture. Residual feature propagation and adaptive attention aggregation mechanisms were innovatively combined to enhance the semantic learning ability for agricultural knowledge graphs. Residual connections were embedded to retain effective layer-wise feature information and alleviate gradient attenuation and oversmoothing problems in deep graph learning. Meanwhile, the attention mechanism dynamically optimized the weight distribution of adjacent nodes, enabling the model to focus on valid semantic associations and suppress interference from invalid noisy edges. Four mainstream similarity calculation methods, namely cosine similarity, Jaccard coefficient, Pearson correlation coefficient and Euclidean distance, were separately adopted to optimize edge weight assignment and characterize entity structural differences. Several classical graph neural network models, including BasicGCN, ResidualGCN, EnsembleGCN, GAT and AirGNN, were selected for comparative analysis. [Results and Discussions] Experimental results demonstrated that the proposed Residual-AttentionGCN model achieved superior convergence performance and relational prediction accuracy compared with all baseline models. Among all adopted edge weighting strategies, Euclidean distance delivered the best comprehensive performance. Under this optimal configuration, the model achieved a minimum average loss of 1.176, with optimal accuracy, recall and an F₁-Score of 0.933, 0.936 and 0.921, respectively. The average values of the three indicators across repeated experiments reached 0.796, 0.808 and 0.789. These results verified that the Euclidean distance-based weighting strategy was more applicable for measuring semantic differences between agricultural entities in this study's graph structure. Moreover, the integration of residual propagation and attention aggregation significantly enhanced the model's capability to capture multi-hop implicit semantic dependencies in rice blast knowledge graphs. On this basis, a prototype expert system was developed for intelligent rice blast prevention based on the Residual-AttentionGCN algorithm. The system supported core functions including disease risk assessment, implicit relation inference and targeted prevention measure recommendation. By inputting basic information such as rice cultivar, planting location and real-time climatic conditions, users could obtain quantitative disease risk evaluation and customized prevention strategies. The system maintained a rapid response time of 0.5–0.8 seconds under conventional computing environments, fully meeting the real-time interaction demands of field agricultural production with reliable practical applicability and engineering value. [Conclusions] In summary, the proposed Residual-AttentionGCN effectively mined latent semantic relations in agricultural knowledge graphs, providing a feasible technical reference for intelligent rice blast management, agricultural knowledge reasoning and refined smart farming applications.

Key words: rice blast, knowledge graph, graph neural networks, attention mechanism, residual learning