M2NN: Multi-view multi-task graph neural network using congestion heatmap imagery for predicting traffic incidents across heterogeneous areas

• Multi-view input flexibly combines traffic sensor data with congestion heatmap imagery. • Multi-task learning jointly predicts link-level and area-level incident probabilities. • Deep learning architecture with GNN, CNN, and RNN captures complex spatiotemporal traffic patterns. • Consistency loss enforces alignment between link-level and area-wide predictions. • Framework supports customisable sub-area shapes and scalable network-wide forecasting. In Intelligent Transportation Systems (ITS), accurate and reliable traffic incident forecasting is essential for proactive traffic management. Traditional incident prediction methods face limitations: they either focus on specific road segments, missing the broader traffic network context, or conduct a global network-wide analysis, decreasing prediction specificity. These models also primarily rely on numeric data, such as traffic measures and network attributes. However, incorporating emerging data such as traffic map imagery can provide richer context to help models to better capture incident patterns and better generalise. This study addresses these challenges by introducing a Multi-view Multi-task Neural Network (M2NN) model, a deep graph neural network (GNN) that flexibly incorporates ‘multi-view’ inputs by combining traffic data with congestion heatmap images and adopts a ‘multi-task’ learning approach by leveraging both link-level incident prediction and subarea-level prediction as model’s target outputs. The evaluation results based on real-world data show that the M2NN model outperforms other benchmark models in link-level incident prediction tasks, which are highly challenging. The ablation analysis reveals that both the multi-view and multi-task learning components significantly enhance the prediction performance, demonstrating the benefit of the proposed architecture.