Deep Spatial-Aware Multimodal Regression for Traffic Field Strength Distribution

Abstract

With the rapid expansion of China's high-speed rail (HSR) network and sustained growth in aviation, the interplay between air transport and HSR has profoundly reshaped regional transportation patterns, manifesting both competition and complementarity. HSR frequently diverts short-haul air passengers while enhancing feeder connectivity to major airport hubs, significantly influencing the spatial distribution of multimodal traffic service radiation. Accurate regression modeling of traffic field strength distribution—which quantifies the service intensity and radiation range of airports and HSR stations—is essential for scientific planning of integrated transportation systems, resource optimization, and sustainable regional development. Traditional statistical models struggle with nonlinear multimodal interactions, spatial heterogeneity, and external disruptions (e.g., COVID-19-induced data sparsity), leading to biased predictions. This study proposes a novel multimodal regression model integrating Residual Networks (ResNet) and Coordinate Convolution (CoordConv). The model employs Principal Component Analysis (PCA) for dimensionality reduction of socio-economic and capacity features, introduces CoordConv to enhance spatial position sensitivity, utilizes ResNet for deep feature extraction, and applies attention mechanisms for dynamic fusion of airport and HSR data. Experiments on spatiotemporal panel data from Shandong Province (2015, 2019, 2021) show the proposed model outperforms baselines (e.g., standard CNN, ResNet without CoordConv), reducing MAE by ~40% (0.25→0.15), lowering RMSE, and achieving R² up to 0.92. Visualizations reveal spatiotemporal evolution patterns, economic-transport coupling, and pandemic impacts. This work provides an efficient deep learning paradigm for multimodal traffic field strength regression, addressing key limitations of traditional approaches.