<p style="font-weight: 400;">Under global warming, accurately predicting regional temperature and understanding the response mechanisms of multi-dimensional environmental variables are crucial for climate adaptation. This study integrates multi-source remote sensing, reanalysis, and ground observation data to construct a monthly provincial-scale environmental dataset for China covering 2005–2022. Long-term trend analysis reveals a widespread increase in Mean Annual Temperature (MAT), which has enhanced vegetation activity, evidenced by rising NDVI and GPP, with rapid responses (0–1 month lag) especially in eastern humid regions. Conversely, deep soil moisture (100–289<span> </span>cm) has declined in northern China, exhibiting lags of up to 5–6 months, indicating prolonged soil drying under sustained warming. Shallow soil moisture (0–7<span> </span>cm) shows variable lags, concentrated in the Yangtze River basin and Southwest China, while soil temperature responds within 0–3 months. Groundwater levels display weak direct correlation with air temperature. To capture complex spatial dependencies and non-linear interactions, we construct a provincial graph based on real geographic adjacency and develop a Graph Convolutional Network (GCN) coupled with Kolmogorov-Arnold Networks (KAN). The KAN-GCN model achieves state-of-the-art performance on the 2021–2022 test period (R<sup>2<span> </span></sup>=<span> </span>0.9889, RMSE<span> </span>=<span> </span>1.18<span> </span>°C), reducing MAE by 15<span> </span>%<span> </span>–<span> </span>30<span> </span>% in ecologically fragile regions such as Xinjiang and Qinghai compared to conventional MLP-GCN. Feature importance analysis consistently identifies surface soil temperature (0–7<span> </span>cm) as the most critical predictor, highlighting the pivotal role of soil-atmosphere thermal coupling. Furthermore, Local Indicators of Spatial Association (LISA) error clustering confirms that KAN-GCN eliminates the persistent high-error clusters observed with MLP-GCN over the Qinghai-Tibet Plateau, demonstrating superior spatial robustness and predictive reliability.