Semi-supervised Deep Learning Solution for Detecting Road Closure Events

The article presents a semi‑supervised deep learning framework designed by Gaode (Amap) to discover and verify road‑closure events, which are a type of dynamic traffic incident that severely impacts user navigation.

Business Background : Dynamic events such as closures, construction, and accidents change road capacity. Closures cause traffic flow to drop to near zero and force users to reroute, making timely detection essential.

Solution Architecture : The processing pipeline is divided into three layers – Data Layer, Discovery Layer, and Verification Layer. The overall solution follows a hierarchical, semi‑supervised design that can operate both offline and online.

Modeling Methods :

Road‑network modeling: spatial modeling (upstream, current, downstream links), business‑data modeling (features from planning, flow, deviation, heatmap) resulting in a 39‑dimensional feature vector, and temporal modeling (time‑series of these vectors).

Algorithm modeling: exploration of classic time‑series models (LSTM, GRU), state‑of‑the‑art TCN, and convolutional models (ResNet). The final model combines LSTM and ResNet (LSTMResNet) to leverage both temporal and spatial representations.

The LSTMResNet network receives a 28‑day sequence of 39‑dimensional vectors, processes them through an LSTM layer, feeds the output to a 7‑block ResNet, and ends with a fully‑connected layer that outputs a binary confidence score.

Regularization techniques such as Batch Normalization and dropout are applied to mitigate over‑fitting.

Business Deployment :

High‑confidence predictions are required to match manual labeling accuracy; semi‑supervised learning is used to boost confidence accuracy.

Risk mitigation includes explainability analysis to ensure model outputs align with business expectations.

Semi‑supervised Training : A small set of high‑quality labeled samples trains an initial model, which then labels high‑confidence online samples for a second‑round training, improving top‑N accuracy by 10 percentage points.

Validation : Model outputs are examined against traffic flow, planning, deviation, and heatmap features. The confidence scores correlate with the likelihood of a closure, and the model’s behavior conforms to domain knowledge.

Conclusion : The semi‑supervised deep learning approach effectively detects road‑closure events, improves user routing, and demonstrates the integration of big‑data analytics with AI techniques in map services.