On the spatio-temporal representativeness of observations

Nick Schutgens, Svetlana Tsyro, Edward Gryspeerdt, Daisuke Goto, Natalie Weigum, Michael Schulz, Philip Stier

Abstract

The discontinuous spatio-temporal sampling of observations has a significant impact when using them to construct climatologies or evaluate models. This study provides estimates of the "representation error" for a range of timescales (semi-annually down to sub-daily) and length scales (300 to 50 km). We demonstrate that even after substantial averaging, significant representation errors may remain, often exceeding typical measurement errors. Our study considers various observations, including ground-site in situ data (PM2.5, black carbon) and satellite remote sensing (imagers and lidar). We find that observational coverage is not an effective metric to limit these errors. However, temporal collocation of data—matching model data to the exact timing of observations—combined with temporal averaging, is highly effective at reducing representation errors. We also show that ground-based and wide-swath imager satellite data yield similar representation errors despite different sampling patterns, while emission sources and complex orography create persistent errors that are difficult to mitigate.

Introduction: The Challenge of Representation Error

In atmospheric science, researchers frequently compare point-based ground measurements or narrow satellite "snapshots" to large-scale model grid boxes (typically 50–300 km wide). This is fundamentally an "apples-to-oranges" comparison because the atmosphere is highly variable. A satellite passing over a city at noon on a clear day does not necessarily represent the average aerosol concentration for that entire grid cell over a full month. This discrepancy is known as representation error. This paper quantifies these errors across different scales and observation types to determine if common averaging strategies are truly effective.

Methods and Study Regions

The researchers employed a "model-as-truth" approach, using high-resolution regional atmospheric models (10 km resolution) to create a four-dimensional "truth" cube of aerosol data. This synthetic truth was then sampled to mimic real-world observing systems:

  • Ground sites: Fixed points sampled continuously or on specific schedules.
  • Passive Satellites: Wide-swath imagers providing daily snapshots at specific local times, often hindered by clouds.
  • Active Satellites (Lidar): Extremely narrow "curtain" views with long repeat cycles (12–16 days).

By comparing these sparse "observations" to the full model grid averages, they isolated the error caused purely by sampling strategy, independent of instrument noise.

Map of the global study regions and the specific atmospheric models applied to each domain.

Datasets and Model Evaluation

To ensure results were not model-dependent, the study utilized several high-resolution simulations across diverse regions, including West Europe, Oklahoma, Congo, Japan, and the open ocean. The models' ability to simulate realistic driving factors, such as rainfall patterns which drive aerosol variability, was evaluated using the Global Precipitation Climatology Project (GPCP).

Comparison of rain rate frequency distributions between model simulations and GPCP observations for the Oklahoma region.

Key Insights and Ideas

  • Coverage is a Deceptive Metric: Simply increasing the amount of data (higher coverage) does not automatically reduce error. If sampling is biased (e.g., only clear-sky days), the error remains high regardless of the data volume.
  • The Power of Collocation: The most effective way to reduce representation error is temporal collocation. Instead of comparing monthly averages, researchers should sample model data at the exact times the satellite or ground station was active.
  • Geography is Persistent: Errors near emission sources or complex terrain are remarkably stubborn. Because spatial gradients are fixed by the landscape, temporal averaging does not smooth them out as effectively as it does for transient weather patterns.
  • Lidar's "Representativeness Gap": Satellite lidars have such narrow fields of view and long repeat cycles that their monthly representation errors are massive (often >100%), making them difficult to use for constructing monthly climatologies on small scales.

Key Results

  • Aerosol Optical Thickness (AOT): Monthly representation errors for ground sites and imagers typically range from 10% to 40%.
  • Daily vs. Monthly: Representation errors for wide-swath satellites and ground sites are quite similar on both daily and monthly scales, despite their different sampling patterns.
  • Lidar Extinction: Monthly errors for satellite lidar were found to be between 70% and 160%.
  • Optimal Averaging Windows: For hourly data to best represent a 210 km grid box, the authors recommend averaging observations over a 6-hour window to account for aerosol transport across the area.

Limitations and Future Work

The 10 km model resolution used here likely provides a lower bound for error estimates, as it cannot resolve sub-10 km atmospheric fine structures. Furthermore, the study assumes specific temporal profiles for emissions; if real-world emissions vary more sporadically, representation errors would be even higher. Future work may involve using Large Eddy Simulations (LES) to probe representativeness at the sub-kilometer scale.

Summary updated

Abstract

The discontinuous spatio-temporal sampling of observations has an impact when using them to construct climatologies or evaluate models. Here we provide estimates of this so-called representation error for a range of timescales and length scales (semi-annually down to sub-daily, 300 to 50 km) and show that even after substantial averaging of data significant representation errors may remain, larger than typical measurement errors. Our study considers a variety of observations: ground-site or in situ remote sensing (PM2. 5, black carbon mass or number concentrations), satellite remote sensing with imagers or lidar (extinction). We show that observational coverage (a measure of how dense the spatio-temporal sampling of the observations is) is not an effective metric to limit representation errors. Different strategies to construct monthly gridded satellite L3 data are assessed and temporal averaging of spatially aggregated observations (super-observations) is found to be the best, although it still allows for significant representation errors. However, temporal collocation of data (possible when observations are compared to model data or other observations), combined with temporal averaging, can be very effective at reducing representation errors. We also show that ground-based and wide-swath imager satellite remote sensing data give rise to similar representation errors, although their observational sampling is different. Finally, emission sources and orography can lead to representation errors that are very hard to reduce, even with substantial temporal averaging.
BibTeX
@inproceedings{schutgens2017spatio,
  author = {Nick Schutgens and Svetlana Tsyro and Edward Gryspeerdt and Daisuke Goto and Natalie Weigum and Michael Schulz and Philip Stier},
  title = {{On the spatio-temporal representativeness of observations}},
  booktitle = {Atmospheric},
  year = {2017},
  doi = {10.5194/acp-17-9761-2017},
}
Figure