Functional Sensor Stratification and Internal Air Quality Reconstruction in the Fisciano District

Low-cost air-quality networks require a distinction between functional context, measurement validity and reconstruction performance. This study examines twenty devices in Fisciano, Italy, assigned to pedestrian/academic, green and vehicle-access classes (11, three and six devices). Device-reported daily PM2.5 and PM10 from May to August 2026 support descriptive class analysis, class-agnostic reconstruction and a computational siting illustration. Crossed date/device mixed models, green-device exclusions, common-date fits and autoregressive residual models assess uncertainty. Three chronological held-out-device tests compare the peer median: inverse-distance weighting, nearest neighbor and fixed random forests. The peer median has the lowest observed pooled error in every period; final-period MAE is 1.43 and 1.58 µg/m3. Its paired advantage over inverse-distance weighting is not consistently distinguishable from zero. The PM2.5 green–pedestrian coefficient changes from −0.340 to −0.003 when D06 is excluded, showing substantial device influence. Reduced-network results differ between omitted-site prediction and retained-site gap filling. Geometric comparisons depend on the demand boundary and unverified candidate assumptions. These findings describe the available network without testing a reconstruction benefit from functional labels. Unknown upstream processing and missing reference co-location limit measurement interpretation.