Cold-Start Hyperlocal PM2.5: A Leakage-Controlled Comparison of Interpolation, Multimodal Fusion, and Contrastive Representation Learning

Abstract

Estimating PM2.5 at unmonitored locations during a wildfire smoke episode is usually framed as a fusion problem. The goal is to produce a single estimate. This estimate combines three data sources. The first source is sparse regulatory monitors. The second source is dense, low-cost sensors. The third source is satellite aerosol retrievals. We test that framing under a leakage-controlled cold-start protocol over the 2020 San Francisco Bay Area smoke season, holding out each of the 15 reference monitors in turn and scoring the transfer to the held-out site. The result is deflationary for the learned approaches. Classical spatial interpolation is the strongest estimator: ordinary kriging reaches an RMSE of 6.10 ug/m^3, against 8.39 for the best supervised tabular fusion and 7.39 for learned graph kriging under the same protocol, and the ordering persists as the network is thinned and on smoke days. Richer inputs do not narrow the gap. Variance-optimal reduction is statistically indistinguishable from the supervised bottleneck at 15 folds, so no representation headroom is detectable, an absent gap rather than proven equivalence. A reference-anchored contrastive pilot aligns the modalities but does not become target-predictive: it fails to encode concentration even in-sample, across eight seeds, which places the failure in the objective and not in generalization. Among the inputs, the low-cost network carries the predictive signal while satellite aerosol optical depth adds little, a non-confounded attribution from train-time ablation and permutation importance rather than a missing-modality floor effect. We present these as calibration findings for the monitored near field, not the far field, and release the dataset assembly, protocol, and experiments for reproduction.

Cite

@inproceedings{antar2026coldstart,
  author    = {Antar, Siam Shibly and Ador, Syem Shibly},
  title     = {Cold-Start Hyperlocal {PM2.5}: A Leakage-Controlled Comparison of
               Interpolation, Multimodal Fusion, and Contrastive Representation Learning},
  booktitle = {Proc. 2nd IEEE International Conference on Data Science and
               Geoinformatics (ICDSG)},
  year      = {2026}
}

Version and rights

This is the accepted author version, not the final published version. © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting or republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. The final published version will be available via IEEE Xplore.