Case Studies | CoMet Toolkit

CoMet Toolkit vs. Atmospheric Weather Balloons

Tue, 13 May 2025 00:00:00 +0000

🎈 Background

Global Climate Observing System (GCOS) ensures access to long-term climate observations. Within GCOS, the GCOS Reference Upper-Air Network (GRUAN) plays a key role by providing reference-quality atmospheric balloon data.

GRUAN data includes measurements such as temperature, humidity and pressure. This data is traceable to SI standards, and published with detailed uncertainty and metadata. At present, GRUAN does not report detailed error correlation across its data products.

This article outlines a study which has successfully deployed CoMet Toolkit to acquire covariance information of GRUAN data.

🗸 Aims

Investigate GRUAN data product uncertainty components.
Obtain covariance matrices for temperature (T) and relative humidity (RH).
Assess how different data processing strategies impact covariance.

📋 Data

GRUAN data is gathered via radiosondes mounted on weather balloons, recording vertical atmospheric profiles at high resolution (1–2 seconds per point) reaching altitudes of 30–35 km.

We selected two radiosonde launches from Lindenberg, Germany (GRUAN’s lead centre) on January 1st, 2024 at 00:00 and 12:00 UTC. The data is available in NetCDF format, and includes full metadata and GRUAN-validated uncertainty components.

Each balloon ascent collects thousands of measurements. To reduce the number of data points, this data is typically layered. For instance, 137-layer ECMWF pressure coordinate system, is commonly used in atmospheric modelling.

🗸 Uncertainties

For each measurement point, GRUAN reports the combined uncertainty as well as three uncertainty components:

Uncorrelated: random, no relationship between points exists.
Spatially Correlated: data within a single vertical profile is spatially related
Temporally Correlated: data across different soundings and/or sites is temporally related (e.g., due to shared calibration).

GRUAN assumes full (r = 1) correlation for spatial and temporal components—an important simplification carried into the analysis.

☄️CoMet Setup

Covariance and correlation matrices were calculated using CoMet.

obsarray: used to re-structure weather balloon measurements, their uncertainty, and metadata (units, dimensions)
comet_maths: performed the mathematical operations to compute covariance matrices, correlation information

🗂️ Results

Using CoMet, raw covariance matrices were computed from the unprocessed midday profile.

Temperature and RH matrices show strong diagonals (variance) and structured off-diagonal patterns (systematic uncertainty).
The RH covariance scale is notably higher, especially due to sensitivity to solar radiation during daytime launches.

In layered heatmaps:

White regions indicate missing data beyond balloon burst.
Mean-layered data shows reduced variance (weaker diagonal).
Subsampled data retains more variance but still shows strong off-diagonal structure due to full assumed correlation.

Notably, the Upper Troposphere–Lower Stratosphere (UTLS) region exhibits distinct covariance patterns.

🗸 Layered Relative Humidity Covariance

Layered RH matrices differ significantly:

Covariance values are up to 100× greater than for temperature.
Heatmaps show a checkerboard pattern, indicating high inter-layer correlation.
Averaging reduces the variance, and temporal correlations dominate the structure, reflecting RH’s rapid environmental sensitivity.

These findings underscore the need to treat RH and temperature differently in uncertainty-aware models.

🗸 Summary of Findings

This project:

Developed and validated a method for computing covariance matrices from GRUAN RS-41 data using CoMet,
Tested multiple data processing strategies (mean, subsample, integration),
Provided practical recommendations through a GRUAN Technical Note.

The results help researchers better account for correlated uncertainty when working with radiosonde data.

🗸 Future Work

Planned next steps include:

Incorporating Tier 2 uncertainty sources,
Implementing partial correlation handling,
Expanding to other variables (e.g., pressure, wind, radiation).

These developments will enhance climate models’ treatment of uncertainty and ensure even greater accuracy in long-term climate monitoring.

CoMet Toolkit used in HYPERNETS Processor

Fri, 14 Jun 2024 00:00:00 +0000

🎈 Background

The LANDHYPERNET and WATERHYPERNET networks (which together make up the HYPERNETS network) consist of a set of autonomous hyperspectral spectroradiometers (HYPSTAR®) acquiring fiducial reference measurements of surface reflectance at various sites covering a wide range of surface types (both land and water) for use in satellite Earth observation validation and remote sensing applications.

The HYPERNETS_PROCESSOR is a Python software package to process the HYPERNETS in-situ hyperspectral raw data to reflectance and other variables.

🗸 Aims

In order to achieve fiducial reference measurement quality, and be optimally useful as a satellite validation reference, uncertainties need to be propagated through each step of the HYPERNETS processing chain, taking into account temporal and spectral error-covariance.

📋 Data

The HYPERNETS instruments measure downwelling hemispherical irradiance, and upwelling radiance. The raw data is provided as digital numbers, and combined with gains and non-linearity coefficients from Lab calibrations. The irradiance and radiance data are combined to calculate reflectance. These data are stored in different product files from L0-L2B.

🗸 Uncertainties

Three uncertainty contributions are tracked throughout the processing.

Random uncertainty: Uncertainty component arising from the noise in the measurements, which does not have any error-correlation between different wavelengths or different repeated measurements (scans/series/sequences).
Systematic independent uncertainty: Uncertainty component combining a range of different uncertainty contributions in the calibration. Only the components for which the errors are not correlated between radiance and irradiance are included.
Systematic uncertainty correlated between radiance and irradiance: Uncertainty component combining a range of different uncertainty contributions in the calibration. Only the components for which the errors are correlated between radiance and irradiance are included. This error-correlation means this component will become negligible when taking the ratio of radiance and irradiance (i.e., in the L2A reflectance products), which is why we separate it from the systematic independent uncertainty.

The two systematic uncertainty components are both based on lab calibrations. Since the same lab calibration is used within the HYPERNETS_ PROCESSOR for repeated measurements (scans/series/ sequences), the errors in the systematic uncertainty components are assumed to be fully systematic (errorcorrelation made up of ones) with respect to different scans/series/sequences. With respect to wavelength, we combine the different error-correlations of the different contributions and calculate a custom error-correlation matrix between the different wavelengths.

CoMet Setup

The HYPERNETS_PROCESSOR uses a Monte Carlo (MC) approach, to propagate uncertainties and error-correlations between product levels. This MC approach is implemented using punpy.

The HYPERNETS products themselves are stored as digital effects tables using obsarray.

Results

An example of the propagated uncertainties on the HYPERNETS radiances is shown below.

These uncertainties can be propagated throughout further processing chains (see e.g. this example)

Summary of Findings

The HYPERNETS_processor is operationally producing product files stored as digital effects tables. Uniquely for this type of processing, multiple different types of uncertainty (including error-correlations) are propagated through each of the processing levels.

These uncertainties have been very useful for the validation of satellite data (e.g. FLEX, CHIME, ENMAP, S2, …).

Future Work

There are a number of uncertainty components that are still missing from the HYPERNETS uncertainty budget (currently a placeholder uncertainty of 2% is added instead). Dedicated studies will be performed to better quantify each of the missing components.