TL;DR
This paper introduces a novel multi-scale recursive method for identifying urban functional areas by integrating taxi trajectory and Points of Interest (POI) data. By combining clustering with a Random Forest model (CA-RFM) and employing a hierarchical road network division, the approach achieves significantly improved accuracy and efficiency in urban functional zone mapping compared to single-scale or single-data-source methods.
What Problem Does This Paper Address?
Accurately identifying and mapping urban functional areas (UFAs) is crucial for urban planning, resource allocation, and understanding city dynamics. However, existing methods often rely on single data sources (like land-use maps or remote sensing) which can be outdated, lack detailed semantic information, or struggle with spatial resolution. Using only one scale of analysis can lead to over- or under-fitting the complex urban structure. This paper aims to overcome these limitations by proposing a method that leverages multi-source data and a multi-scale approach to achieve more comprehensive and accurate UFA identification.
Methods
The core of the proposed method is the CA-RFM model, which integrates a clustering algorithm (K-medoids) with a Random Forest classifier.
- Time Series Generation: Taxi trajectory data is transformed into time-series features representing passenger inflow and outflow patterns across different traffic analysis zones (TAZs) for both weekdays and weekends.
- K-Medoids Clustering (with DTW): Dynamic Time Warping (DTW) is used to calculate similarity matrices for the time-series data. K-medoids clustering is then applied to these matrices to group similar regions and extract significant feature regions for training.
- CA-RFM Model: This model combines the clustering results (as training samples) with POI data to train a Random Forest classifier for initial UFA identification. This approach aims to reduce errors from manual sample selection and leverage both dynamic (taxi data) and static (POI data) information.
- Multi-Scale Recursive Identification: The city is hierarchically divided using road networks of different grades (highways, main roads, ordinary streets, service roads). The CA-RFM model is applied at each scale. A cross-validation step, comparing CA-RFM results with POI quantitative identification (using frequency density and category ratio), determines if a block unit needs further subdivision. This recursive process allows for fine-grained identification from large-scale to small-scale units.
Datasets
- Taxi Trajectory Data: Collected for Shenzhen, China, covering one week (December 7-13, 2015), comprising over 419 million records. This data captures dynamic movement patterns.
- Points of Interest (POI) Data: Collected for Shenzhen in 2018 from the Gaode Map Open Platform. POIs were reclassified into five primary categories: public management/service, commercial service, residential, industrial, and green space/square.
- Road Network Data: Obtained from OpenStreetMap, used to create hierarchical divisions of the study area.
Key Insights and Ideas
- Synergy of Dynamic and Static Data: Combining taxi trajectory data (dynamic) with POI data (static) provides a richer understanding of urban functions than either source alone. Taxi data captures movement patterns, while POIs offer semantic context.
- Hierarchical Road Networks for Segmentation: Using multi-level road networks as a basis for spatial segmentation is more aligned with urban planning and can better capture the hierarchical nature of urban structure than simple grid-based methods.
- Recursive Validation for Accuracy: The iterative cross-validation between the CA-RFM model and POI quantitative identification at each scale is crucial for refining the identification and ensuring consistency, especially in complex or mixed-function areas.
- DTW for Time-Series Similarity: DTW is effective in handling temporal variations in taxi data, allowing for meaningful comparisons of activity patterns across different regions.
Key Results
- Improved Accuracy: The proposed multi-scale recursive method achieved an overall accuracy (OA) of 0.874 and a Kappa coefficient of 0.853. This is significantly higher than single-scale methods or methods using only one data type. For instance, the multi-scale CA-RFM model (D) achieved 0.757 OA, while the single-scale POI identification (A) achieved 0.672 OA.
- Faster Delineation: The method is reported to be faster in delineation compared to traditional methods.
- Reduced Training Sample Error: The combination of clustering and Random Forest (CA-RFM) effectively reduces the error associated with manual selection of training samples.
- Effective Functional Area Identification: The study successfully identified various functional areas, including mixed residential and commercial zones, mature commercial areas, and industrial/public service zones, with good accuracy.
Limitations and Future Work
- Taxi Data Limitations: Taxi data represents only one mode of transportation and may not fully capture the travel patterns of all residents. The data also has inherent positioning errors.
- Data Coverage and Constraints: Regional transportation, economic, and infrastructure constraints can lead to incomplete taxi trajectory data, potentially biasing identification results.
- Boundary Mismatch: Discrepancies between the actual boundaries of delineated study units and functional areas can lead to inaccuracies.
- Future Work Suggestions:
- Integrate more diverse multi-source urban big data (e.g., cell phone check-in data, microblog data) for a more comprehensive understanding of urban morphology.
- Analyze the spatial structure and correlations of identified functional zones to assess urban space utilization and suggest optimization strategies.
- Explore richer data and more precise divisions for improved recognition accuracy.
Summary updated