Seasonal shifts from runoff to groundwater dominance influence daily headwater stream temperatures, especially where local groundwater input is strong. This input buffers temperature during hot periods, supporting cold-water habitats. Recent studies use air–water temperature signal metrics to identify zones of strong stream–groundwater connectivity. While Previous studies used air–water signal ratios as proxies for groundwater influence but were limited to specific sites and periods, without dynamic forecasting. This study is the first to forecast daily A_r as a spatiotemporal signal using a Graph Convolutional Network–Long Short-Term Memory (GCN-LSTM) model. The model was trained using hydroclimate data (air temperature, precipitation, shortwave radiation, streamflow) and watershed physical features (e.g., sand content, slope). Results showed high predictive skill, achieving R² (NSE, RMSE) of 0.86 (0.73, 0.0004) for one-day-ahead to 0.52 (0.50, 0.0009) for seven-days ahead forecasts. Prior studies often have not explicitly incorporated spatial hydrogeologic drivers, but this model explicitly incorporates them to assess their impact on A_r forecasting and stream-groundwater connectivity. Feature analysis identified mean sand, elevation, slope, clay, and TWI as key predictors of A_r. Stronger groundwater signals appeared in hillslopes, elevations, and tributaries, highlighting watershed influence on streamflow. However, limitations include reliance on historical air–water temperature patterns for training and limited representation of extreme climate conditions. Despite these limitations, unlike previous studies relying on measured in-situ stream and air temperature, this study forecasts A_r directly from climate and physiographic features after training, avoiding in-situ data requirements. Findings aiding predictions of stream ecosystem resilience.