Hybrid ST-ResNet and LSTM approach for precise crime hotspot prediction

Area of Science:

• Computer Science
• Criminology
• Urban Planning

Background:

• Crime prediction is challenging due to complex spatial-temporal crime patterns.
• Existing models struggle with fine-grained spatio-temporal crime variability.
• Accurate crime hotspot identification is crucial for public safety and targeted prevention.

Purpose of the Study:

• To improve the accuracy of crime prediction in urban hotspots.
• To leverage deep learning models for enhanced spatio-temporal crime analysis.
• To identify theft hotspots with greater precision for effective crime prevention.

Main Methods:

• Integrated ST-ResNet with Long Short-Term Memory (LSTM) networks for spatio-temporal analysis.
• Incorporated daily Euclidean distance to nearest park as a novel input feature.
• Utilized historical crime data, weather conditions, and temporal variables as model inputs.

Main Results:

• Achieved a mean hit rate exceeding 88% at a fine spatial resolution of 500m.
• Outperformed existing models that perform better at coarser resolutions (1000m).
• Demonstrated superior effectiveness in identifying theft hotspots in Chicago.

Conclusions:

• The proposed deep learning approach enhances spatio-temporal crime prediction accuracy.
• Incorporating proximity to parks and weather data improves hotspot identification.
• The method provides valuable insights for developing targeted crime prevention strategies.