Column optical depth and breathing-height mass are different
How to translate a satellite aerosol optical depth map into meaningful surface PM2.5 information starts with one nonnegotiable distinction: a satellite optical retrieval describes light affected by the atmospheric column, while surface PM2.5 describes particle mass in a volume of air near the ground. These numbers have different units, vertical support and measurement mechanisms. Aerosol optical depth, abbreviated AOD, is dimensionless at a specified wavelength. PM2.5 is usually expressed as micrograms per cubic metre. A useful scientific explanation never labels AOD itself as measured surface pollution.
This matters when a map shows severe haze over a region. The optical signal may be influenced by near-ground particles, transported aerosol above the boundary layer, humidity-induced droplet growth, coarse dust, smoke composition and surface reflectance. Clouds may remove data altogether. To turn AOD into a ground-level estimate, researchers use physical transport and optical models, monitor-based calibration or data fusion. Every approach requires assumptions and validation; none is a simple universal multiplication that changes a colour scale into a pollution-health statement.
Consider a deliberately simplified equation for a well-mixed dry aerosol layer. AOD is approximately the layer’s average extinction coefficient multiplied by its thickness, and the extinction coefficient is approximately particle mass concentration multiplied by mass extinction efficiency. If AOD=0.30, height=1000 metres and efficiency=4 square metres per gram, inferred concentration is 0.000075 grams per cubic metre, or 75 micrograms per cubic metre. Doubling the assumed uniform layer depth halves the estimate to 37.5 micrograms per cubic metre. This is a dimensional demonstration, not a satellite retrieval algorithm or an observation of an actual city.
The next complication is water. Many atmospheric particles grow at high humidity and scatter more light, while dry gravimetric mass measurements condition or define particle water differently. A very humid column can thus look optically denser without a proportional increase in dry PM2.5. Equally, a dense layer of pollution trapped near the surface may be hazardous even if its total column optical thickness is modest. Reading these situations correctly calls for vertical evidence, humidity treatment and validation against surface monitors. We will translate the retrieval chain, explore the measurements, check the arithmetic and distinguish honest claims from overstatements.
1. AOD is not PM2.5
Aerosol optical depth, AOD, measures how a column of airborne particles weakens a light beam at a specified wavelength. Surface PM2.5 is a mass concentration for particles within a specified aerodynamic size range, usually reported in micrograms per cubic metre of air.
Do not translate a dimensionless optical-column value as though it were a surface mass concentration. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
2. Column integration
AOD is the integral of aerosol light extinction coefficient along the vertical optical path through the atmosphere. The same AOD can arise from a dense near-surface layer or a less concentrated aerosol layer extending over a much greater depth.
A satellite observing column extinction cannot locate every contributing particle at breathing height. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
3. The key mass conversion
Converting optical extinction into PM mass requires information about mass extinction efficiency and the vertical distribution of aerosol. The physical connection depends on particle composition, size, water uptake and wavelength, as well as where particles are concentrated.
A single universal multiplier from AOD to PM2.5 is not scientifically defensible. Frame any inference in three layers: what the optical sensor retrieves, what atmospheric assumptions the estimator adds, and what the external validation permits us to conclude. This structure gives readers useful information without making a column measurement sound like a street-level instrument.
4. A dimensional demonstration
For an idealized uniform layer, AOD≈mass extinction efficiency times dry aerosol mass concentration times layer height, with compatible units. If extinction efficiency is in m²/g, concentration in g/m³ and height in metres, the product is dimensionless.
This equation is a teaching model, not a complete satellite retrieval or proof of uniform real atmosphere. Do not replace unavailable data with an unqualified zero or treat a plausible map as an exact exposure record. The missing pixels, cloud screens, aerosol regimes and spatial averaging rules are part of the scientific result and deserve visible explanation.
5. Worked column example
With AOD=0.30, a uniform 1000 m layer and a mass extinction efficiency of 4 m²/g, estimated dry mass concentration is 0.30/(1000×4)=0.000075 g/m³. That equals 75 µg/m³ under the strong stated assumptions.
Change the aerosol distribution or extinction efficiency and the surface inference changes, even though AOD remains 0.30. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
6. Boundary-layer height
If the same AOD and mass extinction efficiency are spread uniformly over 2000 m rather than 1000 m, the illustrative concentration halves to 37.5 µg/m³. A deeper mixing layer can therefore produce lower average concentration for the same column optical burden.
Boundary-layer depth is not always equal to the vertical thickness of the full aerosol layer. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
7. Elevated transported smoke
Smoke carried above the boundary layer can create a large satellite AOD while surface PM2.5 at the location is relatively modest. A simple model attributing all extinction to near-surface aerosol will overestimate local mass in such circumstances.
Use vertical-profile evidence or model layers to distinguish overhead transport from near-ground exposure. Frame any inference in three layers: what the optical sensor retrieves, what atmospheric assumptions the estimator adds, and what the external validation permits us to conclude. This structure gives readers useful information without making a column measurement sound like a street-level instrument.
8. Trapped surface pollution
Conversely, a shallow polluted layer can generate high surface PM2.5 while total-column AOD appears only moderate. Small vertical thickness means fewer optical particles in the full column despite substantial concentration near the ground.
Do not dismiss an observed ground pollution episode because a satellite map seems light. Do not replace unavailable data with an unqualified zero or treat a plausible map as an exact exposure record. The missing pixels, cloud screens, aerosol regimes and spatial averaging rules are part of the scientific result and deserve visible explanation.
9. Humidity-driven growth
Many hygroscopic aerosols absorb water as relative humidity increases, altering their size and light-scattering response. A humid column can exhibit stronger optical extinction without an equal rise in dry gravimetric PM2.5 mass.
Separate ambient optical wetness from conditioning conventions used for dry mass monitoring. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
10. Humidity correction example
In an idealized example, AOD=0.40 through a uniform 1000 m layer, with dry mass extinction efficiency 4 m²/g and ambient humidity enhancement factor 2. The effective wet extinction efficiency is 8 m²/g, implying 50 µg/m³ dry mass rather than 100 µg/m³ if the enhancement is ignored.
The factor of two is only a teaching assumption; real aerosol humidity response is composition- and history-dependent. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
11. What f(RH) means
A humidity enhancement function compares optical response at one humidity with response under a specified drier reference condition. Its argument and reference humidity must be named; an f(RH) measured at a surface instrument is not necessarily representative of every height.
Applying a surface value to the whole column can conceal vertical variations in relative humidity. Frame any inference in three layers: what the optical sensor retrieves, what atmospheric assumptions the estimator adds, and what the external validation permits us to conclude. This structure gives readers useful information without making a column measurement sound like a street-level instrument.
12. Aerosol composition
Sulfate, organic material, mineral dust, sea salt and black carbon have different optical and hygroscopic behaviors. An optical retrieval sensitive to mixture type can change if transported dust replaces urban combustion aerosol.
Do not assume that one mass extinction efficiency applies throughout seasons and source regions. Do not replace unavailable data with an unqualified zero or treat a plausible map as an exact exposure record. The missing pixels, cloud screens, aerosol regimes and spatial averaging rules are part of the scientific result and deserve visible explanation.
13. Particle size
Particles scatter and absorb sunlight differently depending on their sizes relative to light wavelength. Fine-particle mass, coarse dust and ultrafine number cannot be inferred from a single optical measure without additional assumptions.
Specify that PM2.5 is a size-selective mass metric, not a synonym for total aerosol abundance. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
14. Wavelength
AOD is often reported at a chosen wavelength such as 550 nm; measurements at another wavelength are not automatically numerically equal. Spectral slopes help diagnose particle size tendencies, but spectral information alone does not uniquely determine PM2.5 mass.
Always preserve wavelength when quoting AOD. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
15. Ångström exponent context
The Ångström exponent summarises how AOD changes with wavelength and can provide evidence related to particle size distribution. An exponent estimated from two channels has uncertainty and can be affected by mixing of aerosol types.
Do not use it as a universal one-to-one lookup of particle diameter or chemical species. Frame any inference in three layers: what the optical sensor retrieves, what atmospheric assumptions the estimator adds, and what the external validation permits us to conclude. This structure gives readers useful information without making a column measurement sound like a street-level instrument.
16. AERONET comparison
Ground-based sun photometers in AERONET provide column optical observations that can help validate satellite AOD retrievals. Agreement in column AOD does not by itself validate the conversion from that AOD to breathing-level PM mass.
Separate retrieval validation from surface-concentration model validation. Do not replace unavailable data with an unqualified zero or treat a plausible map as an exact exposure record. The missing pixels, cloud screens, aerosol regimes and spatial averaging rules are part of the scientific result and deserve visible explanation.
17. Satellite reflectance problem
A satellite detects radiation influenced by the atmosphere and underlying surface, then uses a retrieval algorithm to estimate aerosol optical properties. Bright deserts, urban surfaces, snow and dark vegetation challenge algorithms differently because surface reflectance contributes to the same observed signal.
A coloured map of AOD is a retrieved quantity, not direct collection of aerosol mass. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
18. Dark-target approaches
Some retrievals exploit relatively dark surfaces to better separate atmospheric aerosol scattering from surface reflectance. That strategy’s applicability changes with land cover, season and view geometry.
Do not assume missing or low-quality pixels over bright surfaces mean clean air. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
19. Bright-surface approaches
Other algorithms are designed to recover aerosol information over surfaces poorly suited to dark-target assumptions. Their strengths and limitations differ with aerosol models, spectral bands and reflectance conditions.
When combining products, keep retrieval-method identity and quality flags visible. Frame any inference in three layers: what the optical sensor retrieves, what atmospheric assumptions the estimator adds, and what the external validation permits us to conclude. This structure gives readers useful information without making a column measurement sound like a street-level instrument.
20. View and sun geometry
Observation angle, solar position and scattering phase affect the signal arriving at a satellite instrument. Two observations of the same aerosol column with different geometry are not interchangeable without retrieval corrections.
Inspect geometry-related quality metadata when comparing sensors or extreme view angles. Do not replace unavailable data with an unqualified zero or treat a plausible map as an exact exposure record. The missing pixels, cloud screens, aerosol regimes and spatial averaging rules are part of the scientific result and deserve visible explanation.
21. Cloud screening
Clouds can obscure aerosol and complicate distinguishing fine droplets from thick humid aerosol. A high-quality aerosol product may exclude cloudy pixels, creating important gaps during storms or humid events.
No satellite AOD retrieval during a cloudy day does not indicate absence of particulate pollution. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
22. Cloud adjacency
Radiation from nearby bright clouds can influence measured reflectance around the cloud boundary. Such adjacency effects may bias retrieved aerosol properties even when a pixel itself passes simple cloud masking.
Check product-level quality guidance rather than accepting every coloured pixel. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
23. Missingness is not random
Satellite sampling opportunities change with cloud cover, season, geography and sensor orbit. An average computed only from clear-sky observations may not represent the all-weather average that health studies or regulators need.
Report the observation count and sampling pattern alongside a purported annual mean. Frame any inference in three layers: what the optical sensor retrieves, what atmospheric assumptions the estimator adds, and what the external validation permits us to conclude. This structure gives readers useful information without making a column measurement sound like a street-level instrument.
24. Overpass time
A sun-synchronous satellite samples a place around a particular local time rather than observing every hour. Morning traffic peaks, afternoon chemistry and evening fires may produce diurnal cycles missed by the overpass.
Satellite-based daily exposure estimates require temporal modeling, not simply relabelling one instantaneous retrieval as a daily mean. Do not replace unavailable data with an unqualified zero or treat a plausible map as an exact exposure record. The missing pixels, cloud screens, aerosol regimes and spatial averaging rules are part of the scientific result and deserve visible explanation.
25. Spatial footprint
A satellite pixel averages signals across a physical footprint that can cover different neighbourhoods and land covers. A roadside monitor may detect a hotspot much narrower than that pixel.
Avoid claiming street-level accuracy from an instrument whose effective resolution is kilometres. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
26. Averaging pixels
An area mean of valid AOD pixels can hide urban hotspots, coastline artifacts or retrieval-selection biases. Specify whether averages are weighted by area, valid retrieval coverage or another rule.
A pixel average is not necessarily a population-weighted exposure estimate. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
27. Surface-monitor matching
To evaluate a satellite-to-PM2.5 model, pair each satellite estimate with an independent surface monitor at compatible place and time. A comparison between satellite noon estimates and unaligned twenty-four-hour filter measurements introduces timing mismatch.
Define spatial tolerance, temporal window and monitor quality requirements before reporting accuracy. Frame any inference in three layers: what the optical sensor retrieves, what atmospheric assumptions the estimator adds, and what the external validation permits us to conclude. This structure gives readers useful information without making a column measurement sound like a street-level instrument.
28. Model calibration
Statistical models may combine AOD, meteorology, land cover, emissions indicators and ground monitors to predict surface PM2.5. The trained model can be useful but its equation is not a universal law of atmospheric physics.
Describe the calibration geography and period before applying it elsewhere. Do not replace unavailable data with an unqualified zero or treat a plausible map as an exact exposure record. The missing pixels, cloud screens, aerosol regimes and spatial averaging rules are part of the scientific result and deserve visible explanation.
29. A toy regression
Suppose an explicitly hypothetical local model predicts PM2.5=3+45×AOD in µg/m³ under otherwise held-constant conditions. At AOD 0.40 the model outputs 21 µg/m³, but those coefficients have no validity outside the invented example without empirical calibration.
Never publish such toy regression constants as observational results or an official conversion factor. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
30. Cross-validation by location
A model may perform well under random train-test splits because neighboring monitor data leak similar spatial information into both sets. Leave-one-site-out or geographically blocked validation tests transfer to unmonitored locations more honestly.
Report the exact validation design, not merely a headline accuracy statistic. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
31. Cross-validation by time
Testing on months not used for training probes sensitivity to new seasons, aerosol mixes and meteorology. Randomly splitting highly correlated daily observations may exaggerate how well the model handles future conditions.
Temporal holdouts are especially relevant when an estimate is to be used in later years. Frame any inference in three layers: what the optical sensor retrieves, what atmospheric assumptions the estimator adds, and what the external validation permits us to conclude. This structure gives readers useful information without making a column measurement sound like a street-level instrument.
32. Bias versus correlation
Predictions can have high correlation with ground measurements while systematically overestimating every polluted day. Inspect mean bias, absolute error, slope, extreme event performance and uncertainty, not only correlation coefficient.
A model can track day-to-day changes without returning accurate absolute mass. Do not replace unavailable data with an unqualified zero or treat a plausible map as an exact exposure record. The missing pixels, cloud screens, aerosol regimes and spatial averaging rules are part of the scientific result and deserve visible explanation.
33. Error arithmetic
If surface observations are 12, 20 and 28 µg/m³ and matching predictions are 14, 18 and 30, errors prediction minus observation are +2, −2 and +2. The mean bias is +0.67 µg/m³, while mean absolute error is 2 µg/m³.
The positive and negative errors partly cancel in bias but do not cancel in absolute error. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
34. High-pollution extremes
A model trained mostly on clean days may shrink extreme PM2.5 estimates toward the middle of its training distribution. Performance on severe smoke events matters when users rely on the result for event interpretation.
Validate tails separately and avoid assuring accuracy at concentrations barely represented in training. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
35. Domain shift
An algorithm developed for humid temperate cities may encounter dust-dominated arid skies or high-altitude aerosol elsewhere. Changes in aerosol type and meteorology break the learned correspondence between optical depth and dry PM2.5.
External validation is required before exporting the same model to a different atmospheric regime. Frame any inference in three layers: what the optical sensor retrieves, what atmospheric assumptions the estimator adds, and what the external validation permits us to conclude. This structure gives readers useful information without making a column measurement sound like a street-level instrument.
36. Topography
Mountain valleys can trap pollution while nearby high terrain samples another part of the atmospheric column. Surface altitude, boundary-layer variation and complex winds complicate interpretation of coarse satellite pixels.
A regionally reasonable model may be unreliable for a single steep valley. Do not replace unavailable data with an unqualified zero or treat a plausible map as an exact exposure record. The missing pixels, cloud screens, aerosol regimes and spatial averaging rules are part of the scientific result and deserve visible explanation.
37. Wildfire smoke aging
Fresh and aged smoke differ in size, composition, humidity response and vertical distribution. Even a plausible AOD-PM2.5 model may change bias over the course of a large smoke transport event.
Monitor-based and vertical comparisons should be included rather than assuming one stable conversion through the plume. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
38. Dust events
Large mineral particles can strongly affect optical signals, but PM2.5 collects only particles below a specified aerodynamic cutpoint. An AOD spike during dust can coexist with changes in PM10 that exceed those in PM2.5.
Know which particle-size metric a surface validation target represents. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
39. Marine aerosol
Sea-salt droplets may take up large amounts of water and produce strong optical effects near coasts. A land PM2.5 monitor can respond differently to a moist sea-salt plume depending on inlet and conditioning.
A simple optical-to-mass relationship may need explicit marine and humidity treatment. Frame any inference in three layers: what the optical sensor retrieves, what atmospheric assumptions the estimator adds, and what the external validation permits us to conclude. This structure gives readers useful information without making a column measurement sound like a street-level instrument.
40. Inverse modeling
Chemical-transport models simulate aerosol emissions, transformations, humidity and vertical mixing to estimate the relationship between column extinction and surface mass. Their output depends on input emissions, meteorological analyses, boundary conditions and parameterizations.
Calling the product ‘satellite measured PM2.5’ obscures the modeled components. Do not replace unavailable data with an unqualified zero or treat a plausible map as an exact exposure record. The missing pixels, cloud screens, aerosol regimes and spatial averaging rules are part of the scientific result and deserve visible explanation.
41. Data fusion
Modern surface PM2.5 products may fuse satellite retrievals, models, ground measurements and learned predictors. That combination can outperform a single sensor but also transfers biases and dependence among input datasets.
List the role of each data source before evaluating claims of fully independent validation. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
42. Uncertainty intervals
A point estimate of 25 µg/m³ could have a wide conditional error distribution depending on region, clouds and aerosols. Reported intervals should reflect calibration and model uncertainty and be tested for actual coverage, not only produce narrow-looking bands.
Distinguish formal statistical intervals from unknown errors caused by new aerosol regimes. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
43. Exposure interpretation
A surface PM2.5 map estimates ambient outdoor concentration at a certain spatial and temporal support. Personal dose depends on indoor infiltration, location changes, activity and breathing rates.
Satellite PM2.5 products do not directly measure individual inhaled dose or diagnose health consequences. Frame any inference in three layers: what the optical sensor retrieves, what atmospheric assumptions the estimator adds, and what the external validation permits us to conclude. This structure gives readers useful information without making a column measurement sound like a street-level instrument.
44. Regulatory comparison
Air-quality standards generally specify particular averaging periods, measurement definitions and compliance methods. An estimated satellite-based hourly or annual mean should not be presented as official regulatory compliance without appropriate authority and method.
Identify the standard’s species, period, jurisdiction and measurement protocol before comparison. Do not replace unavailable data with an unqualified zero or treat a plausible map as an exact exposure record. The missing pixels, cloud screens, aerosol regimes and spatial averaging rules are part of the scientific result and deserve visible explanation.
45. Communicating absence of pixels
White areas or missing grid cells on a satellite map often mean retrieval failure, cloud masking, inadequate quality or no satellite view. They should be labeled unavailable rather than zero pollution.
This is both a graphics issue and an important scientific distinction for readers unfamiliar with remote sensing. Before turning this observation into an air-quality headline, identify the quantity, vertical extent, wavelength and time support. The most important error is often not in arithmetic but in moving without notice from one physical measurement domain into another.
46. Writing the safe conclusion
A defensible statement can say a calibrated model estimated surface PM2.5 using AOD and other variables, validated against independent monitors with reported errors. It should not claim a satellite directly weighed particulate matter at street level.
The strongest conclusion states the predicted quantity, observed validation evidence and remaining limitations. An effective independent check is to compare against a surface monitor or vertically resolved observation that was not used to fit the proposed conversion. Without that separation, an apparently excellent agreement may only restate the model’s training data.
Six checked transfer problems and a defensible report
The reliable interpretation path begins with the satellite sensor’s observed radiance, the algorithm that retrieves AOD, the wavelength and quality mask, and the cloud and surface conditions. The next stage is a model connecting AOD to dry surface PM2.5 using vertical profiles, aerosol extinction efficiency, meteorology and sometimes ground-monitor measurements. Then come validation against independent monitoring sites, error statistics and restrictions on time and geography. Only after those stages should a result be described as a surface concentration estimate. Each stage answers a different question.
Independent exercise one: a toy one-kilometre layer has AOD 0.24 and mass extinction efficiency 6 m²/g. Dry concentration under uniform conditions is 0.24/(1000×6)=0.000040 g/m³, or 40 µg/m³. Independent exercise two: a satellite sees the same AOD over an aerosol layer 2 km deep with otherwise equal extinction efficiency. The uniform concentration becomes 20 µg/m³. Neither value is a claim about where the aerosol actually sits.
Independent exercise three: two hypothetical days have the same observed dry PM2.5 concentration but day B’s humid optical enhancement is twice day A’s. Would AOD necessarily remain unchanged? No. If other conditions are equal, extra water may increase aerosol optical response. Independent exercise four: cloud masking removes all observations during a smoky rainy week, leaving clean clear-sky days for the monthly average. Can the valid-pixel average be declared an unbiased all-weather monthly pollution mean? No, because the missingness is related to weather and may correlate with pollution.
Independent exercise five: an estimated surface series has predictions 8, 18, 28 and observations 10, 20, 20 µg/m³. Errors are −2, −2 and +8; mean bias is 4/3=1.33 µg/m³, while mean absolute error is 12/3=4.00 µg/m³. The outlying overprediction matters even if the average appears small. Independent exercise six: an AOD retrieval agrees with a sun photometer but disagrees with surface monitors. That result does not necessarily discredit optical retrieval; it may instead show flaws in the optical-to-surface conversion. Diagnose the stages separately.
Technical translation is most useful when it teaches what a measurement cannot say. An AOD map cannot uniquely measure surface mass, and a well-calibrated estimated PM2.5 map cannot uniquely reveal individual dose. Precise language protects readers from false certainty while preserving the considerable scientific value of remote sensing, meteorology and independent air-quality observations working together.
First-party scientific sources and related reading
Read NASA’s aerosol optical depth explanation and the NASA DISCOVER-AQ research on surface PM2.5 inference and vertically transported aerosol to understand the retrieval limitations. eduKate companion articles explain AERONET optical-depth measurement and quality screening, aerosol lidar and vertical optical structure and surface particle-mass measurement principles. All numbers in worked examples are hypothetical teaching values.