How Semi‑Supervised Deep Learning Detects Road Closures in Real‑Time

1. Business Background

Dynamic events such as road closures, construction, and accidents affect traffic capacity and user travel. Closures represent a severe reduction in capacity, forcing vehicles to turn around and detour, which significantly impacts user experience. Gaode leverages user trajectory data to discover these events.

The detection problem is split into two parts: addition (new closures) and dissipation (closure removal). Addition deals with rare events across the whole network, while dissipation handles online events that are mostly closures.

2. Solution

Gaode’s pipeline consists of three layers: data, discovery, and verification. A semi‑supervised deep‑learning solution is applied at each layer, supporting both offline and real‑time mining.

Basic data: static road network attributes and dynamic data (user trajectories, routing, deviation, heatmap).

Recall module: identifies suspected closures using traffic drop, increased U‑turns, deviation spikes, and heatmap cuts.

Feature extraction: models traffic, routing, deviation, and heatmap on topological and temporal dimensions, producing a 39‑dimensional feature vector.

LSTM+ResNet prediction: combines temporal LSTM with spatial ResNet (LSTMResNet) for online inference.

Layered output: high‑confidence results are auto‑deployed; medium/low confidence results receive manual assistance.

3. Modeling Methods

3.1 Road‑Network Modeling

The road network is a directed graph where each edge is a link. Modeling includes spatial partitioning (upstream, current, downstream links), business data modeling on these links (producing a 39‑dimensional vector), and temporal modeling using multi‑day sequences.

3.2 Algorithm Modeling

Experiments progressed from classic time‑series models (LSTM, GRU) to Temporal Convolutional Networks (TCN), which outperformed them. ResNet was also evaluated; although slightly weaker than TCN, it performed comparably to GRU and inspired the combined LSTM+ResNet architecture.

The final LSTMResNet model uses a 28‑day input sequence (each day a 39‑dimensional vector). LSTM outputs feed into a 7‑block ResNet, whose output connects to a fully‑connected layer producing a two‑node confidence score for binary classification. To mitigate over‑fitting, Batch Normalization and dropout are applied.

4. Business Deployment

4.1 Semi‑Supervised Boost

The semi‑supervised approach first trains on a small set of high‑quality labeled samples, then predicts on unlabeled online data. High‑confidence predictions are added as pseudo‑labels for a second training round, yielding the final model.

Evaluation shows the semi‑supervised model (V2) improves top‑N accuracy by 10 percentage points over the baseline (V1), confirming its effectiveness for high‑confidence closure detection.

4.2 Business Data Validation

Model outputs are examined against four key features—traffic, routing, deviation, and heatmap—to ensure consistency with business expectations. Validation on Beijing data confirms that confidence scores align with these features, and higher confidence correlates with higher closure likelihood.

Confidence matches business logic across all four features.

Confidence reliably indicates presence of an event.

Higher confidence implies greater closure probability.

5. Conclusion

The article introduces dynamic and closure events, proposes a semi‑supervised deep‑learning solution that integrates road‑network modeling, LSTM, and ResNet, and validates the model against real‑world business data. The approach enhances route planning and user experience by accurately detecting road closures.