Demographics-Informed Neural Network for Multi-Modal Spatiotemporal Forecasting of Urban Growth and Travel Patterns Using Satellite Imagery

Spatiotemporal forecasting of urban growth requires models that explain not only where change occurs but why, linking built form to demographic dynamics and mobility outcomes. However, most models treat these signals in isolation, relying on static projections. To address this problem, we present a Demographics-Informed Neural Network (DINN) that integrates multiyear satellite imagery with demographic data for spatiotemporal prediction of urban growth. The study also leverages these learned demographic-spatial representations for travel behavior forecasting. DINN couples a DenseNet-style image predictor with gated residual connections and a demographic encoder fused at the bottleneck; a separately pre-trained demographic predictor serves as a frozen consistency regularizer during training, while its encoder transfers to a travel-behavior head predicting 16 mobility features. A multi-objective loss balances image fidelity, demographic consistency, and semantic consistency. Using satellite images from 2012-2023 paired with county-level American Community Survey data, DINN improves image quality (SSIM $\approx$ 0.83) and demographic coherence (Demo-loss $\approx$ 0.14), achieves strong demographic prediction (overall $R^2 \approx$ 0.80, $>$0.93 for core population metrics), and delivers accurate travel behavior forecasts (overall $R^2 \approx$ 0.91). To validate the relevance of each architectural component in DINN, we conduct comprehensive ablation studies which effectively highlights the relevance of each model component. This study shows that the framework accurately forecasts spatiotemporal urban change and its associated demographics, linking where change occurs to its drivers and to resulting travel behavior.