PaperlayerView exact revision on arXiv

Source-linked AI summary

Classification of Urban Morphology with Deep Learning: Application on Urban Vitality

Wangyang Chen, Abraham Noah Wu, Filip Biljecki

arXiv:2105.09908v3cs.CVcs.CYcs.LG

TL;DR

The paper addresses the gap between numerical urban-morphology indices and visually perceived patterns. It introduces CRHD images and a ResNet-34 classifier, then uses the resulting features with LightGBM for urban-vitality prediction. The classifier reaches 0.875 overall accuracy, while road-network classification has a small positive association with vitality.

  • Problem

    Existing quantitative morphology methods measure density, proportion, and mixture but do not directly represent visually perceived morphological patterns.

  • Method

    The study generates Colored Road Hierarchy Diagrams, classifies them with ResNet-34, and evaluates probabilistic road-network features in LightGBM vitality models.

  • Results

    0.875 overall accuracy was achieved for road-network classification, and its effect on urban vitality prediction was small but positive.

  • Takeaways & Limitations

    The approach expands quantitative urban morphology with an automated visual classifier and supports scalable comparative studies across cities.

  • Takeaways & Limitations

    The labelled image set needs more worldwide samples, and the model captures overall patterns better than detailed information.

Abstract

from arXiv · show

There is a prevailing trend to study urban morphology quantitatively thanks to the growing accessibility to various forms of spatial big data, increasing computing power, and use cases benefiting from such information. The methods developed up to now measure urban morphology with numerical indices describing density, proportion, and mixture, but they do not directly represent morphological features from the human's visual and intuitive perspective. We take the first step to bridge the gap by proposing a deep learning-based technique to automatically classify road networks into four classes on a visual basis. The method is implemented by generating an image of the street network (Colored Road Hierarchy Diagram), which we introduce in this paper, and classifying it using a deep convolutional neural network (ResNet-34). The model achieves an overall classification accuracy of 0.875. Nine cities around the world are selected as the study areas with their road networks acquired from OpenStreetMap. Latent subgroups among the cities are uncovered through clustering on the percentage of each road network category. In the subsequent part of the paper, we focus on the usability of such classification: we apply our method in a case study of urban vitality prediction. An advanced tree-based regression model (LightGBM) is for the first time designated to establish the relationship between morphological indices and vitality indicators. The effect of road network classification is found to be small but positively associated with urban vitality. This work expands the toolkit of quantitative urban morphology study with new techniques, supporting further studies in the future.

1. Introduction

The paper addresses limits in quantitative urban morphology by combining deep learning with visual representations of road networks. It applies the approach to urban vitality prediction and uses advanced modeling to study morphology–vitality relationships.

  • Motivation: Urban morphology research commonly represents built environments with numerical indices, but these measures overlook patterns grasped visually.The paper positions visual perception as an underrepresented dimension of quantitative morphology.
  • Motivation: Deep learning is proposed as a way to extract features from morphological images and characterize urban form.The authors identify deep learning as underexplored for urban-form characterization.
  • Application: Urban vitality is examined as a case study because open spatial big data facilitate its measurement across urban areas.The paper adopts urban vitality as the application domain for evaluating the morphology method.
  • Research gap: OLS is widely used to relate urban morphology and vitality, but its spatial-independence and linearity assumptions are challenged.The introduction motivates using a more advanced model for this relationship.
  • Contributions: The study introduces supervised road-network classification, evaluates it for vitality prediction, adds multiple vitality dimensions, and adopts an advanced relationship model.The paper separates the general morphology method from the vitality case study.

2. Background and related work

Related work shows that urban morphology measures often capture quantities while missing visual patterns, and that vitality studies use diverse data sources. The paper also identifies nonlinear modeling as an underexplored complement to common OLS approaches.

  • Urban morphology: Urban morphology indices convey quantities but overlook visual patterns, while image-based deep learning has been less explored at mesoscale and macroscale.Existing image-based work has concentrated more heavily on street-view or microscale applications.
  • Urban vitality: Urban vitality has been measured using built-environment indicators such as POI density and road-junction density.These indicators represent the capacity of the built environment to support activity.
  • Urban vitality: People-based vitality measures commonly use location-based services, including social-media check-ins, reviews, and mobile-user density.The literature supplements built-environment measures with population and activity traces.
  • Urban vitality: Nighttime-light data provides another vitality measure linked in prior work to economic activity and nighttime urban conditions.It has been used to identify low-vitality areas and assess vitality at street-block scale.
  • Urban vitality: Broader vitality frameworks include economic, social, cultural, environmental, and innovation dimensions, though their data may lack granularity.Newer sources such as WorldPop and Airbnb are noted as ways to improve measurement detail.
  • Modeling relationships: OLS is common for morphology–vitality analysis, but spatial dependence and linearity assumptions motivate nonlinear models such as LightGBM.The paper identifies nonlinear modeling as an underexplored research direction.

3.1. Research Framework

The research framework combines traditional morphological indices with road-network classification, multiple vitality indicators, and LightGBM. It evaluates whether probabilistic road-network features improve vitality prediction across nine cities.

  • Analytical framework: The framework has three parts: measuring urban morphology, measuring urban vitality, and modeling their relationship.The morphology measurement combines traditional indices with the proposed road-network classification.
  • Road-network classification: The proposed classification generates Colored Road Hierarchy Diagrams, trains a ResNet-34 model, and exports category maps and probabilities.The probabilities can augment the feature space of quantitative morphology studies.
  • Urban vitality measurement: Urban vitality is assessed through built-environment, human-activity, nighttime, population, and tourism indicators.The indicators use POI density, tweet density, nighttime-light brightness, population size, and Airbnb-listing density.
  • Relationship modeling: LightGBM models morphology–vitality relationships by comparing a traditional-index baseline with an augmented model containing probabilistic road-network features.The framework also examines relationships with separate vitality aspects.
  • Study design: The study covers nine cities on multiple continents, divided into approximately 1km × 1km grids.The grids align with WorldPop’s 1km resolution grids.

3.2. Measurement of Urban morphology

The morphology measurement combines Colored Road Hierarchy Diagrams with a ResNet-34 classifier and traditional street, building, and block indices. Road networks are represented through four visual categories and probabilistic features.

  • Colored Road Hierarchy Diagrams: Colored Road Hierarchy Diagrams represent road hierarchy through both line thickness and color.Color also enables thresholding to tone down selected road grades.
  • Road-network categories: The road-network taxonomy contains radial, organic, gridiron, and no-pattern categories.The no-pattern category is included to avoid forcing unpatterned networks into one of the three patterned classes.
  • Road-network categories: Representative CRHDs are shown at 2km × 2km scale using OpenStreetMap road-network data.The figure presents quintessential diagrams for the different categories.
  • Classification model: A ResNet-34 model classifies CRHDs and outputs four softmax probabilities as an embedded feature vector.The probabilities represent membership in the four road-network categories.
  • Classification model: Training uses manually labelled representative samples from the study areas and 108 cities worldwide, with 20% and 80% contributions respectively.This split balances precision in the study areas with robustness to unfamiliar areas.
  • Grid classification: Prediction uses 1km-radius CRHDs concentric with grids, assigning each input to the category with the highest returned probability.The doubled CRHD extent incorporates contextual road networks from adjacent grids.
  • Morphological indices: Traditional indices cover road, intersection, building, block, and land-use characteristics, while classifier probabilities supplement them.The indices are generated automatically with the Morphoindex generator.

3.3. Measurement of urban vitality

The study measures urban vitality across multiple dimensions, combining built-environment, human-activity, nighttime, population, and tourism indicators.

  • Urban vitality includes built environment vibrancy, human activity density, nighttime light brightness, population density, and tourism vibrancy.These dimensions are measured using POI density, tweet density, nighttime light brightness, population size, and Airbnb listing density.

3.4. LightGBM

LightGBM is presented as a tree-based gradient boosting model for learning relationships between morphological features and vitality outcomes.

  • LightGBM extends gradient boosting decision trees with Exclusive Feature Bundling and Gradient-based One-Side Sampling.These techniques accelerate training and improve accuracy.
  • Morphological indices and urban vitality indicators are organized as inputs for quantitative analysis.The study inventories morphological indices and vitality indicators with their data sources.
  • The model estimates feature importance from how often each feature is selected for a tree-node split.

4. Results

The classification model accurately distinguishes road-network patterns, maps their distribution across cities, and supports subgroup analysis and urban-vitality prediction. Its augmentation improves prediction metrics, although patterned road categories contribute less than development-intensity features at grid scale.

  • 4.1. Road network classification: The ResNet-34 model classifies 521 CRHD images into four road-network categories using training, validation, and test splits.Minor roads were truncated from CRHDs to reduce noise.
  • 4.1. Road network classification: 100% accuracy was achieved for no pattern and gridiron networks, compared with 0.8 for organic and 0.7 for radial networks.Organic and radial patterns were harder to classify because their major-road structures vary more across samples.
  • 4.1. Road network classification: The classified maps identify gridiron-dominated US cities, radial concentrations in Paris and London, and organic dominance in Singapore and Shanghai.
  • 4.1. Road network classification: The nine cities form four subgroups: gridiron-dominated, gridiron-organic-mixed, radial-organic-mixed, and organic-dominated.The clustering is based on the proportions of road-network categories.
  • 4.2. Effectiveness of road network classification in predicting urban vitality: The augmented model improves all reported error metrics over the baseline, but the positive effect of road-network classification is not significant.The augmented model combines traditional morphological indices with road-network category probabilities.
  • 4.3. Differences among city subgroups: Morphology-vitality relationships vary substantially among the identified city subgroups.
  • 4.4. Relationship between urban morphology and vitality at the grid scale: Development-intensity features, including building and intersection density, rank highly across vitality dimensions, while most patterned road-network probabilities matter less.The conclusion identifies development intensity as more influential than comparatively patterned road-network categories at grid scale.

5. Discussion

The road-network classification has a small positive association with urban vitality, driven mainly by distinguishing areas with no clear pattern. At very high vitality levels, gridiron and organic networks may outperform radial networks, although this result is not fully proven.

  • 5.1. Understanding the effectiveness of road network classification: The classification effect is small but positive, with no-pattern areas showing lower average vitality than gridiron, organic, and radial categories.Differences among the three patterned categories are subtle, while the no-pattern category shows a clear left shift in its vitality distribution.
  • 5.1. Understanding the effectiveness of road network classification: The proportion of no-pattern areas drops sharply from vitality scores 0–20 to 40–60, while the three patterned categories increase at similar rates.This indicates that the classification effect primarily identifies areas without clear road-network patterns.
  • 5.1. Understanding the effectiveness of road network classification: No-pattern areas tend to have fewer people and activities because sparse or poorly structured networks provide less interaction space and movement efficiency.Possible sources include suburban locations, land uses requiring sparse networks, and geographic obstacles such as hills, lakes, and parks.
  • 5.2. The influence of urban morphology on urban vitality: Development intensity is the predominant global determinant of urban vitality, while identifying no-pattern areas is the primary source of the classification effect.The result links the classification’s predictive value mainly to differentiating areas with weaker network structure and lower vitality.
  • 5.2. The influence of urban morphology on urban vitality: At vitality scores 60–80 and 80–100, gridiron and organic networks dominate, whereas radial and no-pattern networks approach zero proportions.Top-vitality comparisons associate gridiron and organic networks with higher vitality despite lower building density than radial networks, but the result remains insufficiently proven.
  • 5.3. Limitations and future studies: The model recognizes overall road-network patterns but not detailed information, motivating larger labeled datasets and variable-scale object detection in future work.The authors specifically identify Faster R-CNN and YOLO as possible alternatives for detecting patterns at variable scales within larger CRHDs.
  • 5.3. Limitations and future studies: Future studies should add more cities and test other applications because the vitality case study may not fully confirm the classification method’s effectiveness.Suggested applications include energy consumption, heat islands, and mobility.

6. Conclusions

The paper introduces an automated deep-learning framework for visually classifying road networks and applies it to urban vitality prediction. The classifier achieves 0.875 overall accuracy, while classification has a small positive effect on vitality prediction, mainly through identifying no-pattern areas.

  • 6. Conclusions: The study proposes Colored Road Hierarchy Diagrams and an automated pipeline that predicts road-network categories and their probabilities.The approach supports machine cognition of road morphology and is designed for scalable comparative studies.
  • 6. Conclusions: The framework identifies four road-network subgroups among nine cities and provides openly released tools for future large-scale urban morphology studies.The released tools include the Morphoindex generator, CRHD generator, and road-network classification model.
  • 6. Conclusions: Overall classification accuracy reaches 0.875 for distinguishing gridiron, organic, radial, and no-pattern road networks.The method uses a ResNet-34 road-network classification model applied to CRHD images.
  • 6. Conclusions: Road-network classification affects urban vitality prediction positively but only to some degree, with the strongest morphology association observed for built-environment vibrancy.Population density and nighttime vibrancy follow built-environment vibrancy in their reported relevance to urban morphology.
  • 6. Conclusions: The detected vitality-prediction effect primarily reflects differentiation of no-pattern areas, while gridiron and organic networks may be more suitable for high vitality than radial networks.The authors explicitly state that the latter comparison has not been fully proven.
  • 6. Conclusions: The case study contributes both improved statistical measures and new ways to examine links between categorical urban morphology and urban vitality.The authors propose extending the technique to other urban studies.
Loading 2105.09908v3…