Measurement Vectors

The observation provides what we call the “Core Radiance Format”, or colloquially, L1 data. Often it is desired to not directly use L1 data in the retrieval process, but rather some transformation or subset of the L1 data. This could be something like a microwindow, high-altitude normalized radiances, wavelength normalized radiances, etc. We call the result of this transformation the measurement vector, and sometimes even refer to the process of going from L1 data to the quantities used in the retrieval the measurement vector.

Measurement vectors are passed into the Retrieval object as a dictionary of skretrieval.measvec.MeasurementVector objects.

Pre-defined Measurement Vectors

The simplest way to define the measurement vector is to use some of skretrieval’s pre-built measurement vectors.

skretrieval.measvec.Triplet(wavelength, ...)

A class that represents a measurement vector that is a weighted combination of log radiances, high altitude normalized

Measurement Vector Filtering

A MeasurementVector consists of a function that transforms the L1 data, and a parameter that is named apply_to_filter that defaults to “*”. The filter can be used to restrict the measurement vector to apply to one certain measurements. If the L1 data contains multiple RadianceGridded objects such as {"one": ..., "two": ...} then if we set apply_to_filter="one" the measurement vector will only apply to the first RadianceGridded object. This can be useful in synergistic applications where you may want to define multiple measurement vectors for different instruments.

Manually Defining the Measurement Vector

In some cases you may want to define your own measurement vector, in these cases you can supply your own user created function that transform the L1 data, however this can be more complicated than it seems. In addition to transforming the radiances, the transformation must also be applied to the Jacobian matrix, performing standard derivative propagation rules. The measurement error covariance matrix must also be propagated using Gaussian error propagation rules. To make this simpler, skretrieval provides convenience methods to handle these transformations for you. They can be composed to create more complicated measurement vector transformations.

See Measurement Vector API for more details.

Using Context in the Measurement Vector

In some cases, to define the measurement vector we may need to use information internal to the retrieval itself. For example, we may want to define a measurement vector that only uses measurements with tangent altitudes higher than the tropopause.

In this caes we can define a set of context variables that will eventually be passed to the measurement vector

from skretrieval.retrieval.processing import Retrieval

@Retrieval.register_context()
def context(cls, *args, **kwargs):
    tropopause_alt = ... # Calculate from any available ancillary information

    return {"tropopause_alt": tropopause_alt}

Then we can define our measurement vector using placeholder information that can access this context. For example,

from skretrieval.retrieval.measvec import Triplet

meas_vec = Triplet(
    wavelength=[470],
    weights=[1],
    altitude_range=["$tropopause_alt - 2000", 40000],
    normalization_range=[40000, 45000]
)

defines a triplet measurement vector that takes all measurements from 2 km below the tropopause altitude and above.