An Adaptive Spatial Resolution Method Based on the ST-ResNet Model for Hourly Property Crime Prediction

Hong Zhang, Jie Zhang, Zengli Wang, Hao YinISPRS Int. J. Geo Inf 2021 · DOI

TL;DR

This paper introduces an adaptive spatial resolution method to optimize crime prediction models. By leveraging the ST-ResNet deep learning architecture and evaluating performance across various grid sizes, they demonstrate that an intermediate spatial resolution (2.4 km) yields the best hourly property crime prediction accuracy, crucial for effective police deployment.

What Problem Does This Paper Address?

Predictive policing aims to efficiently allocate police resources and deter crime by forecasting where and when crimes are likely to occur. While many methods exist, a critical challenge is determining the optimal spatial resolution for these predictions. Crime is not uniformly distributed, and the scale at which data is aggregated can significantly impact model accuracy due to the modifiable areal unit problem. Traditional methods often struggle with the dynamic nature of crime and the need for precise, short-term forecasts. This paper addresses the need for a method that can automatically identify the most effective spatial resolution for hourly property crime prediction, balancing the trade-offs between capturing localized crime patterns and maintaining manageable data aggregation for practical police deployment.

Methods

The core of this research is an adaptive spatial resolution method built upon the ST-ResNet deep learning model. ST-ResNet is chosen for its ability to capture both spatial structures (using Convolutional Neural Networks - CNNs) and temporal dependencies (using Recurrent Neural Networks - RNNs) in crime data.

Here's a breakdown of their approach:

  • ST-ResNet Model: This model predicts crime risk by considering historical crime patterns (trend, period, and closeness) and external factors like weather.
  • Adaptive Spatial Resolution: Instead of fixing a grid size, the researchers create a "data pyramid" where crime data is aggregated into grids of varying resolutions (from 1.2 km to 9.6 km).
  • Hierarchical Pyramid Structure: This approach allows for efficient processing by generating different resolutions from a base resolution, similar to how web mapping services display maps at various zoom levels.
  • Prediction Accuracy Index (PAI): To evaluate the performance at each spatial resolution, they use the PAI, a metric designed to assess the hit rate of predicted hotspots relative to the study area size. A higher PAI indicates better prediction accuracy.
  • Iterative Evaluation: The ST-ResNet model is trained and evaluated at each spatial resolution. The resolution that yields the highest PAI is selected as the optimal one.
  • External Factor: Daily highest temperature was identified as a significant external factor influencing property crime and was incorporated into the model.

Datasets

  • Property Crime Data: Historical property crime data (burglary, theft, robbery) from January 2012 to October 2013 in Suzhou, China. This dataset includes case number, crime type, occurrence time, and geographical coordinates. Over 130,000 cases were used.
  • Weather Data: Daily average, highest, and lowest temperature data for the same period (January 2012 to October 2013), collected from the China Ministry of Agriculture Planning and Design Institute.

Key Insights and Ideas

  • The "Sweet Spot" for Spatial Resolution: The most significant finding is that neither the finest nor the coarsest spatial resolution is optimal for hourly crime prediction. An intermediate resolution (2.4 km in this study) strikes the best balance, leading to higher prediction accuracy (PAI).
  • Trade-off Between Granularity and Practicality: While finer resolutions (e.g., 1.2 km) might capture more localized crime events, they can lead to unstable counts and make it difficult for police to deploy resources effectively. Coarser resolutions can mask important localized hotspots.
  • Temporal Patterns Matter: The study confirms the importance of temporal patterns (closeness, period, trend) in crime, with "closeness" (crimes occurring closer in time) being particularly relevant for hourly predictions.
  • Temperature as a Predictor: Daily highest temperature was found to be a significant external factor, explaining a notable portion of the variance in property crime.
  • Adaptive Method's Value: The proposed adaptive method automates the selection of the optimal spatial resolution, saving researchers and practitioners the effort of manual experimentation and ensuring a more robust prediction model.

Key Results

  • The adaptive spatial resolution method, using ST-ResNet, identified 2.4 km as the optimal spatial resolution for hourly property crime prediction in Suzhou.
  • At this optimal resolution, the model achieved a PAI of 8.8 on the test dataset.
  • The RMSE at the optimal resolution was 7.81.
  • The study found that the highest temperature explained 30.9% of the variance in property crime.
  • The near-repeat analysis indicated a significant pattern within the first week and within 300m, which weakened with time and distance.

Limitations and Future Work

  • Limited External Factors: The study primarily focused on temperature as an external factor. The authors acknowledge that many other factors (e.g., socioeconomic, environmental) can affect criminal behavior and suggest incorporating more factors in future work to improve prediction accuracy.
  • Generalizability: While the method is presented as adaptive, its performance might vary across different cities or crime types. Further validation in diverse contexts would be beneficial.
  • Data Privacy: The authors note that data is available on request due to privacy restrictions, which is a common limitation in crime data research.

Future work could involve exploring a wider range of external factors, applying the adaptive method to different crime types and geographical areas, and investigating more advanced deep learning architectures for enhanced spatio-temporal crime prediction.

Summary updated

BibTeX
@article{zhang2021adaptive,
  author = {Hong Zhang and Jie Zhang and Zengli Wang and Hao Yin},
  title = {{An Adaptive Spatial Resolution Method Based on the ST-ResNet Model for Hourly Property Crime Prediction}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2021},
  doi = {10.3390/ijgi10050314},
}
Figure