Create Map Predictions from ARTMO table

This recipe demonstrates how look-up tables generated with the ARTMO toolbox can be used to create spatially explicit maps of the parameters derived in the radiative transfer model.

This will include the following steps:
  • Import the ARTMO look-up table into the EnMAP-Box
  • Spectrally resample the wavelength information to the target sensor/image
  • Fit a Regressor
  • Apply the model to the image

This example uses the ARTMO table provided with the EnMAP-Box test dataset. You can find the look-up table and the metadata in the EnMAP-Box plugin folder:


Also open the other datasets from the test data (in the processing toolbox Auxilliary ‣ Open Test Maps).

First, lets have a look at the metadata (directional_reflectance_meta.txt):

CLASS:  Generic class
DATE:  2019-08-05
# BANDS:  2101.000000
# SAMPLES:  1000.000000
File created for ARTMO. Date: 2019/8/5 Time: 9:12:50
Line 1, Column 8 ... end: Wavelength
Column 1: Cab  min: 0.0550      max: 79.9352    count: 1000
Column 2: Cw   min: 0.0020      max: 0.0500     count: 1000
Column 3: Cm   min: 0.0020      max: 0.0500     count: 1000
Column 4: lai  min: 0.0001      max: 7.9968     count: 1000
Column 5: angle        min: 0.0264      max: 89.9445    count: 1000
Column 6: psoil        min: 0.0000      max: 0.9996     count: 1000
Column 7: vCover       min: 0.0001      max: 0.9995     count: 1000
Line 2, Column 8 ... end: Directional_reflectance
Fixed variables:
Model: Prospect 4
N:Leaf Structural Parameter:   1.5000
Model: 4SAIL
skyl:Diffuse/Direct light:     10.0000
hspot:Hot spot:        0.0100
tts:Solar Zenith Angle [deg]:  30.0000
tto:View zenith Angle [deg]:   0.0000
psi:Relative Azimuth Angle [deg]:      0.0000

As we can see, the first line in the directional_reflectance.txt file stores the wavelength information (starting at column 8, column 1-7 are set to 0). Column 1-7 store the values of the different indicators (Cab, Cw, Cm, lai, angle, psoil, vCover) which we are going to use as response variables. Starting in line 2 and column 8 we have the directional reflectance values.

Importing the ARTMO table

First we have to import the ARTMO table. Go to the Processing toolbox and start the algorithm Auxilliary ‣ Import ARTMO lookup table. This algorithm will convert the ARTMO table into two (pseudo) images with the dimensions n_samples x 1. This is necessary as the subsequent processing steps are based on the raster format as input.

  • In the algorithm dialog under ARTMO lookup table navigate to the file directional_reflectance.txt
  • The Reflectance scale factor allows rescaling the reflectance values in the ARTMO table on-the-fly. This might be necessary if the values of the target image have a different scaling. Here in the cookbook example this is the case, as the reflectance values of the EnMAP test image are scaled between 0-10000, but the ARTMO reflectance is in floating point between 0 and 1. Therefore set the value to 10000.
  • Now you just need to specify the output paths. Output Raster stores the directional reflectance values. Regression Output stores the corresponding indicator values (Cab, Cw, Cm, lai, angle, psoil, vCover), where every indicator is stored in a separate band.
  • Click Run to start the processing.

Spectral Resampling

Since the ARTMO data and the EnMAP image have different spectral resolutions we have to align them before we can fit a machine learning regression. Therefore the higher resolved ARTMO data will be resampled to EnMAP resolution. Open the processing algorithm Resampling and Subsetting ‣ Spectral Resampling.

  • Under Raster select the file with the ARTMO reflectance values (Output Raster from previous step).
  • Under [Option 1] choose select sensor
  • Under [Option 2] select enmap_berlin.bsq (i.e., the file with the desired target resolution)
  • Leave the Resampling Algorithm at Linear Interpolation
  • Specify output path under Output Raster
  • Click Run

Fit a regressor

In this step we will train a regressor where we, based on the ARTMO data, regress the (resampled) spectral reflectances against the indicators.

  • Open the algorithm Regression ‣ RandomForestRegressor


    You can also choose different machine learners here, the following steps will be identical.

  • Choose the resampled ARTMO raster as Raster (predictors) and the image containing the indicator values as Regression (response). Leave the Mask field empty.

  • You can leave the regressor parameters in the Code window at default.

  • Specify a output path for the Output Regressor. This is an intermediate .pkl file which will be needed in a subsequent step.

  • Click Run

Predict Regression

In this final step we will apply the model we fit in the previous step to the EnMAP image in order to receive a pixel wise prediction of the ARTMO indicators.

  • Open the algorithm Regression ‣ Predict Regression
  • Select enmap_berlin.bsq as Raster
  • Under Regressor specify the .pkl file from the previous step
  • Specify an output path (Output Regression)
  • Click Run

The result is a 7 band raster, where each band corresponds to one indicator: