Blog/

Clear-sky bias in satellite AOD, and why it skews a gap-fill analysis

A lone tree stands at the edge where clear blue sky meets a wall of orange dust haze in the outback.

If you've ever pulled a satellite AOD time series to patch a gap between two monitors and noticed the record looks suspiciously clean, that's not luck. It's the cloud mask doing its job, and that job has a side effect most people filling network gaps don't think about until a reviewer asks why the interpolated values during a known haze episode look too low.

The cloud mask throws out more than clouds

Aerosol retrieval algorithms need a cloud-free line of sight to the surface to back out optical depth from top-of-atmosphere reflectance. So every overpass runs a cloud screen first, and any pixel that fails gets flagged and dropped from that day's scene. For a coarse multispectral sensor at 250 m-plus, that screen is coarse too, and it can't always tell a thick smoke plume from a cloud deck. Dense wildfire smoke or heavy regional haze frequently gets masked out right along with actual cloud cover. The pixels that make it through screening are, by construction, the ones where the atmosphere was optically simple enough for the retrieval to run, a reasonable requirement for the algorithm but also a selection filter on the days you get to see.

Why the clear days you keep aren't a random sample

This is the part that trips up a gap-fill analysis: the days that pass the cloud screen correlate with the synoptic conditions that produce both cloud-free skies and aerosol buildup. Stagnant high-pressure systems bring sunny, stable weather, weak mixing, and low-level inversions that trap particulate and let it accumulate day over day. Frontal passages bring cloud cover and rain that scrubs the column, and those are often the days the retrieval screens out anyway, so you'd have gotten low AOD readings from them if you could. Run the numbers for your own airshed and the skew shows up in a specific direction: a basin prone to stagnation events will have a clear-sky record tilted toward the buildup days, overstating the long-run average, while a basin where fronts do most of the scrubbing will have that same record tilted toward the clean days, understating it. A satellite-derived AOD climatology built only from cloud-screened overpasses describes fair-weather atmospheric states, a pattern the literature calls clear-sky sampling bias.

Layer the smoke-masking problem on top and you get a second failure mode pulling the other way: some of the actual worst days, the thick-plume days a ground monitor would register as a spike, never make it into the satellite record at all because the cloud algorithm mistook the plume for cloud. So your gap-fill input can be simultaneously biased toward fair-weather buildup conditions and missing the single most acute events. Both are forms of the same underlying problem, a representativeness error that comes from treating a cloud-screened sample as if it stood in for the full record.

What that means for the raster you're actually filling

For day-to-day monitoring-network work, read a clear-sky composite as a sampled subset of conditions and build your gap-fill around that limitation. A few things that help:

  • Track the clear-sky fraction for your grid cell alongside the AOD value itself. A cell with 40% valid-day coverage over a month tells a different story than one with 90%.
  • Flag suspected plume-masking events by cross-referencing fire detections or known transport corridors, since a masked day during an active smoke event is informationally very different from a masked day during a stable clear stretch.
  • Validate against whatever fixed monitors you do have, to check whether the satellite record's clear-sky bias runs in a consistent direction for your specific airshed.
  • Treat any interpolated value for a day that was cloud-screened across the whole region with real caution. There's no retrieval underneath it; the gap-fill is pure extrapolation from the clear-day pattern.

The coarse daily composite stays useful between stations when you frame it correctly. Air Quality Proxy builds its daily gridded AOD raster on that framing: a between-monitor proxy that shows where haze is building before a station catches it, delivered as a GeoTIFF and a per-cell time series meant to sit alongside your fixed network.

If you're trying to spot haze forming in the dead zones between stations, that's worth a look.

Get started

← Back to the blog