Spatiotemporal Traffic Accident Prediction Using Deep Learning Models¶

A. Manninen — M.Sc. thesis, Aalto University, June 2024

The City of Helsinki has set a target of halving traffic deaths and serious injuries by 2030 against a 2020 baseline, which is the premise the thesis takes as given. Meeting a target like that requires knowing where and when to intervene, which is a forecasting problem before it is a policy one. This thesis asks whether a week's notice of elevated accident risk, at the resolution of a city block, can be extracted from data the city already collects.

The learning problem¶

Data points. One data point is a (grid cell, week) pair: a 1000 m × 1000 m subregion of Helsinki, paired with one week. A road-network mask restricts prediction to cells that actually contain road, which matters more than it sounds — without it, the model spends its capacity learning that parks and water have no collisions.

Features. Each prediction uses the preceding seven days of observations for that cell, drawn from several public sources and grouped into families the thesis ablates separately: traffic volume, weather (temperature, rainfall), calendar (month, quarter, Finnish holidays), land use (e.g. tree cover) and road-network structure (e.g. extent of paved road).

Labels. The label is the next week's accident count in that cell — a non-negative integer, making this regression rather than classification. The central difficulty is that the label is zero-inflated: accidents are rare, and at this spatial resolution the overwhelming majority of cell-weeks are zero. A model that always predicts zero scores well on average error while being useless, and much of the thesis is organised around that trap.

Hypothesis space. Three deep architectures are compared against a historical-average baseline: a long short-term memory network (LSTM); a convolutional neural network (CNN) paired with an LSTM, so that spatial and temporal structure are processed by different machinery; and a transformer.

Loss function. Mean squared error (MSE), minimised with RMSprop at a learning rate of 0.0001 and batch size 64, for up to 100 epochs with early stopping after five epochs without validation improvement. Dropout and early stopping both act as regularisation. Evaluation deliberately reports MSE, mean absolute error (MAE) and recall together. Under zero-inflation the error metrics and the detection metric disagree, which is the point.

What it found¶

CNN-LSTM was best on MSE (0.01546) and recall (0.2764); plain LSTM was best on MAE (0.02602). No single model dominates, and the thesis makes the sensible recommendation to pick by which error you care about: CNN-LSTM when you want to catch accidents, LSTM when you want calibrated magnitudes.

One honest anomaly is reported rather than smoothed over: validation loss came out lower than training loss for every model. The thesis works through the candidate explanations rather than leaving it as a curiosity: more accidents fall in the training period than the validation period, and dropout is active during training but not evaluation. That discussion is a good model of what to do when a diagnostic looks wrong.

Watch the defence¶

Full text¶

Read the thesis at Aaltodoc (open access)