General questions on the training data files
Q1: could provide further details about the given data or, if the data is provided in some standardized format?
The training data is given in hdf5 (.h5) file format. This hierarchical data file, similar to netcdf, allows storing several variables and metadata/attributes within one single file. It's broadly used in the remote sensing community and there are libraries for operating with such files in languages such as Python, Matlab, R, or C++.
The inputs (X) for your model are stored within the variable called "LUTheader" with dimensions dxn (d=number of input dimensions; n=number of samples), whereas the outputs (Y) are stored in a variable called "LUTdata" with dimensions lxn (l=number of output dimensions). There is another variable of interest, "wvl", which stores the wavelength grid (in nm) of the spectral data in LUTdata.
Q2: How many radiative transfer functions are stored in scenarios A and B?
Both scenarios contain the same number of radiative transfer functions (6), as shown in the global attribute outnames: Lp0, Edir0, Edif0, S, tdir, and tdif. You can confirm this by looking at the dimensions of LUTdata (k=25230 in scenario A and k=14700 in scenario B) and the wavelength grid length in the wvl variable (n_wvl=4205 in scenario A and n_wvl=2450 in scenario B).
Q3: What do these radiative transfer functions represent?
The functions correspond to various atmospheric radiative processes. As described in the outnames attribute: Path radiance (Lp0), direct and diffuse at-surface solar irradiance (Edir0 and Edif0), spherical albedo (S), direct and diffuse surface-to-satellite transmittance (tdir and tdif). For more details, you can refer to this publication: DOI: 10.1080/01431160802438555.
Q4: What is the order of the transfer functions in the target data?
The transfer functions are concatenated across their entire spectra i.e. first Lp0, then Edir0 and so on. The order is as follows:
(f1_wvl1, f1_wvl2, ..., f1_wvl_k, f2_wvl1, ...)
Q5: What does one data point (n_comb dimension) represent? Are these resolved spatially or temporally?
Each of the n_comb samples (e.g., n_comb=500 in the train500.h5 file) represents a random combination of atmospheric/geometric input variables (stored in LUTheader), obtained via Latin Hypercube Sampling. These samples do not represent spatial or temporal data, but rather "potential conditions" that could occur in nature. For more context, see the publication: DOI: 10.1109/TGRS.2023.3300460.
Q6: where can I find more information about the input variables? The meta-data of the group 'dynamic' seems to match the input features. Is that correct?
You can find the ID name of each parameter used in the input dimension (n_param) within the attribute "varnames" in the variable "LUTheader". For example, for the scenario A, the "varnames" attribute contains the following: O3STR (O3 in atm-cm), H2OSTR (columnar water vapor in g/cm2), VIS (aerosol optical thickenss at 550 nm), G (henyey-greenstein asymmetry parameter), ASTMX (Angstrom exponent), SSA (aerosol single scattering albedo), PARM2 (solar zenith angle), OBSZEN (viewing zenith angle), PARM1 (relative azimuth angle).
You can find more details if you explore the content of the "dynamic" group, which contains the min-max values of these variables and, more importantly, the attributes with "description" and "units" for further details. A priori you don't need using the information contained in the 'dynamic' group.
Additionally, the 'dynamic' group provides the min and max values of each input variable, along with further details like descriptions and units. However, for model training, you likely don’t need to use the 'dynamic' group itself; the main data for training is contained in LUTheader.
Q7: What are the two values associated with each dynamic variable in the meta-data, and what does the 'static' group refer to?
The two values represent the min and max values of each dynamic variable. For example, you can refer to dynamic/H2OSTR to find the min-max values for columnar water vapor (in g/cm²). However, the actual LUTheader variable contains all n_comb combinations of the input variables for your model training.
The static group contains MODTRAN variables that were held constant (with default values) throughout the dataset. Generally, you won’t need to use the static group for model training.
Final Note:
For your model, we recommend sticking with the variables in LUTdata (outputs Y), LUTheader (inputs X), and, if needed, wvl (wavelength grid, for splitting the 6 transfer functions).