AOD to PM2.5 conversion explained: why column aerosol isn't surface concentration

If you've pulled an AOD product and tried to line it up against your reference monitor's PM2.5 reading, you've already found the problem: the numbers don't match, and they're not supposed to. Aerosol optical depth is a column measurement. It tells you how much sunlight gets scattered or absorbed by particles in the entire vertical slice of atmosphere the satellite is looking through, from the sensor down to the surface. PM2.5 is a ground-level mass concentration measured by a sensor sitting a few meters off the pavement. Converting one to the other means estimating a surface value from a column signal, with several layers of atmospheric physics sitting in between.
What's actually standing between AOD and a PM2.5 number
Three variables do most of the damage to a naive AOD-to-PM2.5 regression:
Boundary layer height. AOD integrates aerosol through the whole column, but PM2.5 only cares about what's mixed into the layer near the ground. On a day with a shallow boundary layer, the same AOD value can correspond to a much higher surface concentration than on a day when the mixing layer extends a kilometer up. This is why a lot of agencies pull in a reanalysis-derived boundary layer height as a covariate before trusting any AOD-based estimate.
Relative humidity. Aerosol particles are hygroscopic. They swell as humidity rises, which changes how much light they scatter without changing their dry mass. A hazy, humid afternoon can show elevated AOD that overstates the actual PM2.5 load, and a dry, dusty day can do the reverse.
Aerosol type and vertical profile. Smoke, dust, sulfate, and sea salt scatter light differently per unit mass, and they don't all sit at the same altitude. An AOD retrieval over a region with a mixed smoke-and-dust event is averaging across particle types with different optical properties, which is part of why empirical AOD-PM2.5 relationships built for one airshed often need recalibrating for another.
AOD keeps its value for surface air quality work as long as the conversion has ground truth behind it, and the quality of that ground truth usually comes down to how many reference monitors are feeding the regression.
Where satellite AOD ground truth actually comes from
Most operational AOD-to-PM2.5 approaches lean on one of two paths. The statistical path fits a regression (often with boundary layer height, humidity, and sometimes land use as covariates) against collocated monitor readings, then applies that fit to grid cells without a monitor. The chemical transport model path runs a model like CMAQ or a similar framework and uses AOD as an assimilation input to nudge the simulated surface field. Both approaches depend on having enough fixed-station data in the region to validate against, which is exactly the resource that's thin between monitors in the first place. That's the structural tension in this whole exercise: the areas where you most want a converted PM2.5 estimate are usually the areas with the least ground truth to calibrate it against.
That scarcity caps what you can defend from the conversion. A coarse-resolution daily AOD raster won't give you a defensible PM2.5 number for a monitoring report between two stations twenty miles apart. What it will do is show you where the column is loading up with aerosol before any station catches it, which is a different and still useful kind of information for a network with real gaps in it.
Using AOD as a between-station signal, not a station replacement
If the goal is catching a haze event forming in the space your fixed network doesn't cover, you don't necessarily need a converted PM2.5 value. You need a daily, gridded view of column aerosol that flags where it's rising relative to its own baseline, so you know where to look first. That's a narrower claim than "satellite PM2.5," and it's the one that holds up without pretending a coarse multispectral pass can substitute for a reference-grade monitor.
This is the gap Air Quality Proxy is built around: a daily AOD raster gridded over your region, delivered as GeoTIFF with a per-cell time series, built to sit between your fixed stations rather than replace any of them. If you're trying to see a haze event building before it reaches a monitor, it's worth a look.