Crime is a critical social issue that threatens public safety, and accurately predicting and reducing crime in hotspot areas is a key objective. Crime patterns are inherently complex, with substantial spatial and temporal variability, which makes accurate prediction challenging. This study addresses these intricacies by leveraging advanced deep learning models, including Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), to enhance predictive accuracy in crime hotspots. Specifically, we integrate ST-ResNet for spatio-temporal analysis with Long Short-Term Memory (LSTM) networks to model temporal dependencies, overcoming ST-ResNet's limitation in capturing sequential patterns in crime data. To further improve predictive accuracy, we introduce a feature: the daily Euclidean distance of each crime to the nearest park as a secondary input channel for ST-ResNet, supplemented by historical crime data. Additionally, weather conditions and temporal variables are incorporated into the model's fully connected layer as additional inputs. Our method achieves a mean hit rate of over 88% at finer spatial resolutions (500 m), outperforming existing models, which tend to perform better only at coarser resolutions (1000 m). By capturing finer spatial details, our approach proves more effective in identifying theft hotspots across Chicago, offering valuable insights for developing targeted crime prevention strategies.
Keywords: Hotspot mapping; LSTM networks; Real-time crime prediction; ST-ResNet; Spatio-temporal analysis.
© 2025. The Author(s).