Crime is an act that violate the laws of a society, typically leading to prosecution and punishment by the state [1]. These acts range from minor infractions, such as petty theft and vandalism, to severe offenses like murder and terrorism. The effects of crime are profound and multifaceted, impacting individuals, communities, and society at large [2]. On an individual level, victims of crime can suffer physical harm, psychological trauma, and financial loss, which can lead to long-term emotional distress and a diminished quality of life [3]. Communities affected by high crime rates often experience a breakdown of social cohesion and trust, leading to fear, decreased property values, and economic decline. Businesses may be reluctant to invest in high-crime areas, exacerbating unemployment and poverty. Societally, crime strains public resources, including law enforcement, judicial systems, and correctional facilities. It also necessitates substantial government expenditure on policing, legal proceedings, and incarceration [4]. Moreover, pervasive crime can erode public confidence in the rule of law and governance, leading to broader social stability and development implications. Thus, understanding and addressing crime is essential for promoting safety, justice, and prosperity within any community [5].

A crime hotspot is a specific geographic area where the frequency of criminal activity is significantly higher compared to other areas. These hotspots often cluster crimes, such as theft, assault, or vandalism, making them focal points for law enforcement and community safety efforts [6]. Identifying and understanding crime hotspots is crucial for effective policing, as it allows for strategically deploying resources, targeted patrols, and proactive crime prevention measures. Hotspot detection is the process of identifying these high-crime areas using various analytical techniques and data sources. This process involves collecting and analyzing crime data, often including details such as the type of crime, location, time of occurrence, and other relevant factors. Advanced methods, such as statistical analysis, Geographic Information Systems (GIS), and machine learning algorithms like Kernel Density Estimation (KDE) are used to visualize and predict hotspots [7]. These tools help in mapping out the intensity and distribution of criminal activities, enabling law enforcement agencies to focus their efforts on areas that need the most attention, thereby enhancing public safety and reducing crime rates.

Traditional hotspot detection methods have primarily involved manual analysis, statistical approaches, and basic Geographic Information Systems (GIS). Manual analysis entails law enforcement personnel reviewing crime reports and records to identify high-crime areas. This method often involves plotting incidents on physical maps or simple digital tools and visually identifying clusters of criminal activity [8]. While this approach can provide insights, it is labor-intensive and time-consuming. Additionally, it is highly susceptible to human error and subjective bias, which can lead to inconsistent and inaccurate hotspot identification. The manual process’s limitations make it difficult to keep up with the dynamic and evolving nature of criminal activity. Statistical methods, such as point pattern analysis and spatial autocorrelation, have also been used to detect crime hotspots [9]. These techniques involve calculating the density and distribution of crime incidents within a given area. Point pattern analysis focuses on identifying statistically significant clusters of events, while spatial autocorrelation measures the degree to which crime events are spatially correlated. Although these methods offer a more systematic approach than manual analysis, they are often limited by their reliance on predefined statistical models and thresholds, which may not accurately capture the complexities of real-world crime patterns.

Basic GIS-based methods have advanced traditional hotspot detection by enabling the visualization of crime data on digital maps. Tools like heat maps and thematic maps allow for a more intuitive understanding of crime distribution [10]. However, these methods still have significant limitations. Basic GIS tools often lack the analytical depth required to identify nuanced patterns and trends. They may not effectively integrate multiple data sources or consider the influence of various socio-economic and environmental factors on crime. Furthermore, these methods usually provide static representations of crime data, failing to capture criminal activity’s dynamic and temporal aspects.

Addressing these challenges may require a better model and practical implementation strategies to ensure that machine learning models provide reliable and actionable insights for crime prevention and law enforcement [11,12]. The proposed CHART represents a significant advancement over traditional and existing machine learning methods for crime hotspot detection. By integrating comprehensive data preprocessing, robust feature engineering, and sophisticated algorithms like Adaptive Synthetic Sampling (ADASYN) and Kernel Density Estimation (KDE), CHART effectively addresses the limitations of previous approaches, such as inefficiencies, biases, and computational complexity. The use of a Random Forest-based model ensures high accuracy and robustness, mitigating overfitting and enhancing generalizability. CHART’s ability to deliver precise, timely, and actionable insights into crime patterns significantly outperforms benchmark methods, allowing law enforcement agencies to allocate resources and improve crime prevention and response strategically. This framework sets a new standard for dynamic spatial analysis and prediction, providing a powerful tool for enhancing public safety and community well-being.

The key contribution of the proposed research are as follows:

Utilization of advanced feature engineering techniques using domain-specific knowledge, such as time of day, location type, and historical crime frequency, to extract and transform relevant features for improved model performance.

The rest of the paper is organized as follows: Section 2 reviews the current literature on hotspot detection techniques, Section 3 outlines the core methodology of the proposed work, Section 4 presents the experimental evaluations and results. Section 5 illustrates the application of CHART and Section 6 discusses the conclusion and directions for future work.

This section discusses the current literature that has been carried out in the domain of crime prediction. Dakalbab et al. [13] proposed a comparative analysis of different artificial intelligence based model to predict and prevent crime. The authors did a study where they reviewed 120 research papers about AI and crime prediction. They looked at the different types of crimes studied and the techniques used to predict them. They found that supervised learning was the most common approach used. They also looked at the strengths and weaknesses of the different techniques. They found that AI can be very effective in predicting crime, especially when used to identify crime hot spots. Hybrid models also showed promise. In the end, they suggested that more research should be done on hybrid models and that they plan to do further experiments to improve their own solution.

Xie et al. [14] discussed that spatial hotspot mapping is important in many areas like public health, public safety, transportation, and environmental science. It helps to identify areas with high rates of certain events like disease or crime, but traditional clustering techniques can give false results, which can be costly. To solve this problem, they developed a statistically robust clustering techniques, which use rigorous statistical methods to control false results. This article provides an optimized technique, including data modeling, region enumeration, maximization algorithms, and significance testing. The goal was to stimulate new ideas and approaches in computing research and help practitioners choose the best techniques for their needs. The work of Garcia-Zanabria et al. [15] discussed the challenges of understanding crime patterns in big cities. Crime is often spread out and hard to see, making it difficult and expensive to analyze. Their article introduced a new method called CriPAV, which helps experts analyze street-level crime patterns. CriPAV has two main parts: a way to find likely hotspots of crime based on probability, not just intensity, and a technique to identify similar hotspots by mapping them in a Cartesian space. CriPAV has been tested with real crime data in Sao Paulo and has been shown to help experts understand crime patterns and how they relate to the city.

Law enforcement authorities need to use data-driven strategies to prevent and detect crimes, as proposed by Al-Osaimi et al. [16]. However, their work limits the amount of data generated every day is increasing, which makes it difficult to process and store it. Their article also new Apriori algorithm to analyze crime by using various datasets. They designed a crime analysis tool for public safety and data mining that helps law enforcement officers to make better decisions. Wu et al. [17] proposed a place-based short-term crime prediction model that used patterns of past crimes to predict future crime incidents in specific locations. Their model was based on the concept human mobility that can contribute to limited crime generation. They used a large-scale human mobility dataset to evaluate the effects of human mobility features on short-term crime prediction. In addition to this, they also tested various neural network models on different cities with diverse demographics and types of crimes and found that adding human mobility flow features to historical crime data can improve prediction accuracy.

Cardone et al. [18] presented a fuzzy-based spatiotemporal hot spot intensity and propagation technique. Their work explained a new way to study “hot spots,” where a certain thing is happening a lot. The method involves using a computer program to find these hot spots and measure how strong they are. Their method was tested by looking at crime in the City of London over several years, and the results showed that crime has been decreasing in all parts of the city. Their method seems to be reliable and could be used in the future to study other things happening in different places. Appiah et al. [19] also discussed a model-based clustering of expectation maximization and K-means algorithms in crime hotspot analysis to fix crime in different areas. They used a mathematical method called Gaussian multivariate distributions to estimate potential crime hotspots. This involves finding the best way to group data points into clusters to identify areas where crimes are likely to occur. They used a large dataset of violent crimes and analyzed the data using a combination of K-means clustering and the expectation-maximization (E-M) algorithm. They found that this new method is efficient and fast and produced similar results to traditional methods.

Prathap et al. [20] discussed a geospatial crime analysis and forecasting with machine learning techniques paper. In their work, they discussed that people used social media to connect with others, share ideas and content, and for professional purposes. Researchers are able to analyze individual behavior and interactions on social media sites like Facebook and Twitter. Criminology is an area of study that uses data gathered from online social media to understand criminal activity better. Researchers can obtain valuable information about crime by analyzing user-generated content and spatiotemporal linkages. This research examines 68 crime keywords to categorize crime into subgroups based on geographical and temporal data. The proposed Naive Bayes-based classification algorithm is used to classify crimes, and the Mallet package is used to retrieve keywords from news feeds. Their study identifies crime hotspots using the K-means method and uses the KDE approach to address crime density. The study found that the suggested crime forecasting model is equivalent to the ARIMA model. The comparative overview of the proposed work with existing approaches is given in Table 1.

In conclusion, the extensive review of current literature underscores the significant strides made in crime prediction and hotspot detection using artificial intelligence and machine learning techniques. Various methodologies, including supervised learning, hybrid models, statistically robust clustering techniques, and geospatial analysis, have demonstrated substantial efficacy in identifying and predicting crime hotspots. The introduction of advanced algorithms, such as those leveraging human mobility data and fuzzy-based spatiotemporal techniques, highlights the innovative approaches employed to enhance crime prediction models’ accuracy and reliability. Building upon these advancements, the proposed CHART framework offers a comprehensive and superior approach to intelligent crime hotspot detection and real-time tracking. By utilizing a robust methodology that includes comprehensive data collection from the ICT police database, sophisticated data preprocessing, domain-specific feature engineering, and applying Kernel Density Estimation for heat map generation, the CHART framework effectively visualizes crime density and identifies hotspots. Incorporating a Random Forest-based model for hotspot detection further enhances the predictive accuracy and reliability of the framework.

This section discusses the core methodology of CHART, which is majorly composed of data collection, preprocessing, feature extraction, and prediction. The proposed methodology in Fig. 1 combines various data sources, including police reports, public databases, and social media, to create a comprehensive crime prediction model, CHART. It starts with a robust text preprocessing pipeline that cleans and prepares data by removing URLs, converting text to lowercase, removing numbers, joining text tokens, and stripping punctuation. This clean data is then subjected to feature extraction, where domain knowledge is applied to extract relevant crime-related attributes like time, location, and type. These features are ranked and labeled to prepare for machine learning analysis. The model uses Kernel Density Estimation (KDE) to generate a heat map, visually representing crime hotspots as smooth data point distributions. Finally, a Random Forest-based approach is employed for crime detection, utilizing decision trees that aggregate predictions through majority voting or averaging to identify potential crime hotspots effectively. This comprehensive approach aims to enhance real-time crime prediction and hotspot detection, providing law enforcement with actionable insights for targeted interventions.

This section presents the study’s results on crime analysis and hotspot prediction using map mark points in a safe city context.

Accuracy, precision, recall and F1 are prescribed measure that has been used to evaluate the effectiveness of classification models, as given in Eqs. (14)–(16). These measures provide insights into different aspects of the model’s performance. (14)Accuracy=TP+TNTP+TN+FP+FN (15)precision=TPTP+FP (16)recall=TPTP+FN

This section presents the experimental results of the proposed model on a crime dataset on different datasets. The dataset is divided into 70% training and 30% testing. The training set is used to fit the model and learn the underlying crime patterns, while the test set is reserved for evaluating the model’s predictive performance on unseen data. This split ratio was chosen based on common machine learning practices and provides a balanced approach to train the model effectively while preserving enough data to assess its generalization capabilities. The details of hyperparameters are shown in Table 4.

The evaluation of the proposed approach was based on three key metrics: accuracy, precision, and recall. The results, as shown in Fig. 3, demonstrate the effectiveness of the model in analyzing and predicting crime patterns in Islamabad.

The confusion matrix results for datasets CD-I, CD-II, and CD-III are illustrated in Fig. 4, showing the model’s classification performance across the three datasets. Each confusion matrix highlights the number of true positives, false positives, true negatives, and false negatives, providing a detailed view of how well the model distinguishes between different classes in each dataset.

In the absence of direct ground truth data, we evaluated the effectiveness of the proposed method using proxy ground truth derived from historical crime patterns recorded in the Islamabad Police dataset. This dataset spans multiple years and includes detailed crime attributes such as location, time, and crime type. To assess the model’s accuracy, we compared the predicted crime hotspots with historical crime data, achieving a 92.8% match between the model’s predictions and actual crime concentrations from the Islamabad Police dataset. Additionally, domain experts from local law enforcement evaluated the predictions, confirming the relevance of the identified hotspots with an expert agreement rate of 90.2%. We further validated the method using unsupervised evaluation metrics like the silhouette score, which resulted in a score of 0.92, indicating well-clustered and distinct crime hotspots. These combined evaluations demonstrate the model’s reliability and effectiveness in predicting crime patterns in the absence of explicit ground truth.

The performance of CHART has been compared with several well-known machine learning algorithms, including Naive Bayes, Support Vector Machine (SVM), Logistic Regression, Linear Regression, and K-Nearest Neighbors (KNN). Naive Bayes is a probabilistic classifier based on Bayes’ theorem, assuming independence between features. Support Vector Machine (SVM) is a classifier that identifies the hyperplane that best separates the data into different classes. Logistic Regression is a statistical model that uses a logistic function to model a binary dependent variable. Linear Regression uses a linear equation to estimate the relationship between a dependent variable and one or more independent variables. K-Nearest Neighbors (KNN) is a non-parametric algorithm that classifies a data point based on the classifications of its nearest neighbors. The performance of these algorithms was assessed based on three key metrics: Accuracy, Precision, and Recall. The results, summarized in Table 5, demonstrate the effectiveness of the proposed approach compared to these traditional classifiers.

The results highlight the superior performance of the proposed approach in comparison to the other classifiers. Specifically, the proposed approach achieved an accuracy of 95.65%, significantly higher than the other classifiers. The closest contender, SVM, achieved an accuracy of 88.47%, demonstrating the robustness of the proposed method. With a precision of 93.87%, the proposed approach outperformed all other classifiers, with SVM and Logistic Regression following at 86.53% and 84.74%, respectively. The recall of the proposed approach was 94.56%, indicating its effectiveness in identifying true positive cases, and notably higher than the recall values for SVM (87.29%) and Logistic Regression (85.10%). The comparative analysis indicates that the proposed approach not only surpasses traditional machine learning algorithms in terms of accuracy but also excels in precision and recall. This high performance can be attributed to the model’s ability to effectively learn from the crime data, capturing the intricate patterns and distributions associated with various crime types and locations within Islamabad. By achieving higher precision, the proposed approach ensures that the majority of the identified crime hotspots are indeed areas with high crime rates, reducing false positives. Similarly, the high recall value ensures that most high-crime areas are correctly identified, minimizing false negatives. Overall, the results validate the effectiveness of the proposed approach as a reliable tool for crime prediction and analysis, outperforming conventional classifiers and demonstrating its potential for broader applications in urban planning and law enforcement.

The ablation study conducted highlights the contributions of each key component of the CHART framework by evaluating performance metrics such as accuracy, precision, and recall as shown in Table 6. The full CHART model, which integrates Kernel Density Estimation (KDE), Random Forest, ADASYN for data balancing, and advanced feature engineering, achieves the highest overall performance with 95.65% accuracy, 93.87% precision, and 94.56% recall. When ADASYN is removed from the model (KDE + Random Forest without ADASYN), the performance declines, with accuracy dropping to 91.12%, demonstrating the importance of addressing class imbalance in crime data. Models relying solely on KDE or Random Forest perform worse, achieving 88.47% and 90.12% accuracy, respectively, highlighting the effectiveness of combining these approaches. Further, using KDE with unbalanced data results in the lowest performance (84.76% accuracy), underscoring the need for data balancing. For comparison, spatio-temporal KDE, as proposed by Hu et al., shows moderate performance with 86.78% accuracy, but it does not match the robustness of the CHART framework. This study illustrates the necessity of combining multiple advanced techniques to achieve superior performance in crime hotspot detection.

The web-based application has been developed to provide an interactive platform for analyzing and visualizing crime data. The dashboard, displaying data from police calls has been shown in Fig. 5, allows users to filter by Police Circle, Police Station, and Criminals for focused analysis. It categorizes crimes into types such as Assault, Burglary, Kidnapping, and Theft. The dashboard shows a total of 45.29 K police calls recorded between 01 January 2024 and 30 May 2024. Users can filter data by Police Circle, Police Station, and Criminals, allowing for a more focused analysis. This also helps in understanding temporal patterns and spatial distribution, aiding in effective police deployment during high-risk times. This application is valuable for law enforcement and city planners, enabling data-driven decisions to enhance public safety and crime prevention. The application also provides real-time feedback and insights, helping to strategize patrol routes, allocate resources efficiently, and design safer urban layouts.

The hotspot identification interface has been shown in Figs. 6 and 7 provides a clear and detailed visualization of crime hotspots and trends using Safe City mark points. By pinpointing these high-risk areas, law enforcement agencies can deploy police resources more effectively, ensuring a more strategic and efficient approach to crime prevention. The spatial distribution highlights police stations with the highest and lowest crime rates, facilitating targeted interventions in areas with higher crime concentrations.

This allows for a more focused allocation of resources to the areas that need them the most, enhancing the overall safety of the community. Additionally, the dashboard provides insights into the most common types of crimes and their hotspots, which assists law enforcement agencies in prioritizing specific crime prevention strategies.

This research proposes an advanced machine learning framework, CHART (Crime Hotspot Analysis and Real-time Tracking), designed to address the challenges in crime prediction and prevention. The methodology begins with comprehensive data collection from the ICT police database, encompassing detailed attributes such as crime type, location, time, and demographic information. Key steps include data preprocessing, domain-specific feature engineering, heat map generation using Kernel Density Estimation (KDE), and hotspot detection with a Random Forest model to predict crime likelihood in various areas. Experimental evaluation demonstrated CHART’s superior performance over benchmark methods, significantly improving crime detection accuracy, achieving 95.24% for CD-I, 96.12% for CD-II, and 94.68% for CD-III. The integration of sophisticated preprocessing techniques, balanced data representation, and advanced feature engineering ensures the model’s reliability and practicality for real-world crime analysis. Visualization of crime hotspots allows law enforcement agencies to strategize effectively, focusing resources on high-risk areas to enhance overall crime prevention and response efforts. Future work will explore incorporating real-time data streams, enhancing model adaptability to emerging crime patterns, and integrating additional contextual factors such as socio-economic indicators to further improve prediction accuracy and utility.