Hybrid RidgeGAN: Predicting Transportation Indices in Small and Medium-Sized Indian Cities

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Dec 18 2023

In an article published in the journal Nature, researchers addressed the complexities of predicting road network density in small and medium-sized Indian cities coming under the Integrated Development of Small and Medium Towns (IDSMT) project. They introduced a hybrid Ridge Generative Adversarial Network (RidgeGAN) model, incorporating Kernel Ridge Regression (KRR) and the City Generative Adversarial Network (CityGAN) model and successfully generated hyper-realistic urban patterns for these cities.     

Background

Rapid urbanization comes with various challenges, such as environmental degradation, air pollution, and more. The transportation system's complexities within urban and regional road networks have been analyzed using transportation models. However, predicting road network density in small and medium-sized Indian cities presents challenges. Existing models often lack the ability to simulate realistic urban patterns, especially in regions with limited data on spatial covariates.

These models face limitations, prompting the integration of deep learning, specifically CityGAN, for generating synthetic urban patterns resembling real Indian cities. However, the quantification of urban patterns of Generative Adversarial Network (GAN)-generated cities and the prediction of transportation indices for new urban regions remain inadequate.

To bridge these gaps, the RidgeGAN model was proposed, leveraging the creativity of GAN-generated cities and the data-centric KRR model to predict transportation indices, specifically road network density. The research involved 503 real Indian cities, aiming to establish relationships between Human settlement indices (HSIs) and transportation indices. The innovative approach enabled visualization and evaluation of different urban development scenarios, aiding urban planners in optimizing layouts for specific objectives like minimizing traffic congestion.

Methods

The proposed two-step pipeline involved using CityGAN for the unsupervised generation of urban patterns and KRR for the supervised prediction of transportation indices. Correlation analysis, including Pearson's correlation coefficient (PCC) and Chatterjee Correlation Coefficient (CCC), was employed to assess relationships between transportation indices and HSIs.

The researchers addressed the absence of empirical research on the connection between transportation indices and human settlements, particularly for Indian cities. The RidgeGAN model integrated CityGAN and KRR, aiming to predict road network density for GAN-generated urban patterns. The research highlighted the significance of evaluating GAN-generated images to ensure their representation aligned with real-world urban areas before practical applications.

The contributions of RidgeGAN included its application to modeling urban morphology, simulating spatial structures of cities, and establishing a spatial relationship model between spatial metrics and topological indices. The study emphasized the importance of understanding the relationship between spatial metrics depicting human settlement patterns and transportation indices, particularly in the context of developing countries like India.

Despite the contributions, the research acknowledges limitations. Evaluating GAN performance lacks a standardized objective loss function, and the proposed evaluation measures may be insufficient for time-varying urban patterns and extensive data analysis. Additionally, the validation of the relationship between HSIs and transportation indices cannot be performed on simulated data lacking network density information.

Study Results

The study area encompassed 503 cities in India with populations ranging from 20,000 to 500,000. Settlement footprints and transportation network data were collected and pre-processed using satellite imagery and Geographical Information System (GIS) software. HSIs were calculated, including metrics such as Total Class Area (CA), Number of Patches (NP), Largest Patch Index (LPI), Clumpiness Index (CLUMPY), Aggregation Index (AI), and Normalized Landscape Shape Index (NLSI). Additionally, OpenStreetMap (OSM) data was used to calculate road network density. Validation metrics, including Average Radial Profile (ARP) and correlation coefficients, were employed to assess the accuracy of the CityGAN model in simulating realistic urban patterns.

Furthermore, the authors used correlation analyses to explore the relationship between human settlement characteristics and transportation networks in real cities. The findings highlighted significant correlations between certain human settlement indices and transportation indices, such as CA, showing a strong correlation with road length (RL) and network density (ND).
To predict road network density, a KRR model was implemented and compared with other regression models. The KRR model outperformed alternative models in terms of accuracy metrics, indicating its efficacy in predicting transportation indices for given settlement patterns.

Discussion

The CityGAN model was validated using average radial profiles, and K-means clustering optimized its performance. Supervised machine learning models were then compared, with KRR outperforming others in predicting transportation indices from human settlement data. The proposed hybrid model proved effective in forecasting road network density for CityGAN-generated urban patterns, offering a decision-support system for sustainable planning. However, challenges could arise in larger cities with diverse urban landscapes, necessitating extensive and varied datasets for optimal CityGAN performance. Future work could explore connectivity measures in conjunction with settlement indices for a more comprehensive analysis.

Conclusion

In conclusion, this study presented a hybrid model, integrating CityGAN simulations and KRR, to predict road network density for small and medium-sized Indian cities. Validated through average radial profiles and K-means clustering, the model demonstrated accuracy in generating realistic urban patterns. As a decision-support system aligned with the IDSMT scheme, it aided in the optimal allocation of funds for transportation infrastructure development. While acknowledging potential challenges in applying CityGAN to larger cities, the study offered a practical tool for sustainable planning and development, with future research avenues suggested for exploring connectivity measures alongside settlement indices.