Geospatial-based data and knowledge driven approaches for burglary crime susceptibility mapping in urban areas

Abstract

The Damansara-Penchala region in Malaysia, is well-known for its high frequency of burglary crime and monetary loss based on the 2011-2016 geospatial burglary data provided by the Polis Diraja Malaysia (PDRM). As such, in order to have a better understanding of the components which influenced the burglary crime incidences in this area, this research aims at developing a geospatial-based burglary crime susceptibility mapping in this urban area. The spatial indicator maps was developed from the burglary data, census data and building footprint data. The initial phase of research focused on the development of the spatial indicators that influence the susceptibility of building towards the burglary crime. The indicators that formed the variable of susceptibility were first enlisted from the literature review. They were later narrowed down to the 18 indicators that were marked as important via the interview sessions with police officers and burglars. The burglary susceptibility mapping was done based on data-driven and knowledge-driven approaches. The data-driven burglary susceptibility maps were developed using bivariate statistics approach of Information Value Modelling (IVM), machine learning approach of Support Vector Machine (SVM) and Artificial Neural Network (ANN). Meanwhile, the knowledge-driven burglary susceptibility maps were developed using Relative Vulnerability Index (RVI) based on the input from experts. In order to obtain the best results, different parameter settings and indicators manipulation were established in the susceptibility modelling process. Both susceptibility modelling approaches were compared and validated with the same independent validation dataset using several accuracy assessment approaches of Area Under Curve - Receiver Operator Characteristic (AUC-ROC curve) and correlation matrix of True Positive and True Negative. The matrix is used to calculate the sensitivity, specificity and accuracy of the models. The performance of ANN and SVM were found to be close to one another with a sensitivity of 91.74% and 88.46%, respectively. However, in terms of specificity, SVM had a higher percentage than ANN at 57.59% and 40.46% respectively. In addition, the error term in classifying high frequency burglary building was also included as part of the measurements in order to decide on the best method. By comparing both classification results with the validation data, it was found that the ANN method has successfully classified buildings with high frequency of burglary cases to the high susceptibility class with no error at all, thus, proving it to be the best method. Meanwhile, the output from IVM had a very moderate percentage of sensitivity and specificity at 54.56% and 46.42% respectively. On the contrary, the knowledge-driven susceptibility map had a high percentage of sensitivity (86.51%) but a very low percentage of specificity (16.4%) which making it the least accurate model as it was not able to classify the high susceptible area correctly as compared to other modelling approaches. In conclusion, the results have indicated that the 18 indicators used in this research could be employed to successfully map the burglary susceptibility in the study area. Furthermore, it was also found that residential areas within the vicinity of Brickfields, Bangsar Baru, Hartamas and Bukit Pantai are consistent to be classified as high susceptible areas, meanwhile areas of Jalan Duta and Taman Tunku are both identified as the least susceptible areas across the modelling methods.