Scenario-Adaptive Visibility Level Retrieval via Multi-Source Synergy: Enhancing Physical Traceability and Scene Decoupling Within a Tree-Routed TabPFN Framework

Accurate retrieval of visibility grades is critical for transportation safety. Due to the highly complex meteorological backgrounds, traditional global deep learning models frequently struggle with limited physical traceability and feature heterogeneity. To address these challenges by enhance physical traceability and reduces heterogeneity, this study proposes a scenario-adaptive visibility retrieval framework based on multi-source synergy, namely TabPFN-ExtraTrees (TabPFN-ET), targeting major transportation routes in Anhui Province, China. Fusing Fengyun-4 (FY-4A/4B) satellite multispectral observations with ground meteorological data, this framework utilizes the divide-and-conquer routing mechanism of ExtraTrees to decouple the complex, heterogeneous feature space into highly homogeneous sub-scenarios. Subsequently, the TabPFN model conducts high-precision inference within each specific subspace. Evaluations on a class-balanced benchmark demonstrate that TabPFN-ET achieves an Overall Accuracy of 0.681, outperforming baseline models such as SAINT across various metrics. Furthermore, this paper conducts a physically consistent analysis of the framework. Feature importance and node profiling corroborate its physical consistency: the FY-4 upper-level water vapor channel (Channel 09) and near-surface humidity act as the macroscopic atmospheric stability and microscopic thermodynamic constraints, respectively, driving the model’s scene decoupling and inference. Cross-regional tests in Jiangsu provide preliminary indications of context-specific transferability.