Estimating Ground-Level PM2.5 over Beijing–Tianjin–Hebei from DQ-1 Wide-Swath Imager Signals and ERA5 Meteorology Using Physics-Guided Feature Engineering

The DaQi-1 (DQ-1) Wide-Swath Imager (WSI) combines an approximately 2300 km swath with 15 visible-to-shortwave-infrared channels at 75–600 m resolution, providing broad regional coverage and flexible spectral support. We developed an ExtraTrees estimator from WSI signals and ERA5 planetary boundary-layer height, winds, and relative humidity for Beijing–Tianjin–Hebei in 2024. H5 footprints placed verified daylight overpasses between 04:45 and 05:38 UTC, making 05:00 UTC the nearest available ERA5 field for all 96 dates. Multispectral screening retained 12,768 North China records; the strict benchmark used 4210 records from 65 stations. The pre-specified date split selected B3 (0.470 μm), but ten repeated balanced date-block allocations found no stable single-band winner (B3 mean R2 = 0.3373 ± 0.0311; highest mean, B14 = 0.3506 ± 0.0274). In a controlled date-grouped ablation, B3 plus meteorology and physics-guided terms achieved R2 = 0.1516; adding coordinates increased R2 to 0.1640, whereas adding calendar encodings increased it to 0.3919. This full contextual configuration was used only for within-2024 mapping. The primary coordinate- and calendar-free joint date–station benchmark yielded R2 = 0.1643 for ExtraTrees with WSI plus meteorology, compared with 0.1085 for meteorology alone and −0.2349 for WSI alone. Ninety-seven scenes produced 11 monthly composites. CHAP comparison yielded pooled pixel–month r = 0.799 and city–month r = 0.830, but May and July agreement was near zero and October differed by −16.01 μg m−3. Bias above 200 μg m−3 was −158.27 μg m−3 under date blocking. The results support a conditional WSI contribution within a meteorology-dominated estimator; they do not establish a stable optimal band, an independent radiometric retrieval, or reliable severe-event exposure estimates.