Skip to content

Subpixel classification

The aim of this exercise is to explore the possibilities of subpixel classification, namely the different methods of endmember collection, regression-based spectral unmixing, and evaluation of the resulting species maps.

  • Prerequisites

    • Installed EnMAP-Box plugin for QGIS (manual)
      By now you should have EnMAP-Box installed on your computer as a QGIS plugin and be familiar with the basic software operations. If not, please refer back to Exploration of hyperspectral data using EnMAP-Box.
    • Downloaded data (module4/theme4_exercise_subpixel_classification)
      The dataset consists of:
      • preprocessed (geometrically and radiometrically corrected) hyperspectral images: BL_hyperspectral_image.dat (data file) and BL_hyperspectral_image.hdr (header file). More information about the acquisition of the data can be found here.
      • training polygons for all classes (species): BL_training_polygons.shp (shapefile)
      • pixel purity index image: BL_ppi_image.dat (data file) and BL_ppi_image.hdr (header file):
      • spectral library with semi-automatically pre-extracted endmembers: BL_spectral_library.gpkg (GeoPackage).

  • Tasks

    • Manually extracting endmembers from hyperspectral images
    • Comparing the acquired spectra with the provided spectral library
    • Spectral unmixing – subpixel classification
    • Results evaluation

0. Data preparation and exploration

The area of interest is Bílá louka meadow, a patchwork of alpine grasslands located in the Krkonoše mountains (50.729N, 15.683E). Matgrass (Nardus stricta), reedgrass (Calamagrostis villosa), and purple moor-grass (Molinia caerulea) form monodominant stands that are occasionally mixed with other grasses and herbs. Several isolated Norway spruce (Picea abies) trees can be found. Despite the grass stands low species diversity, they play an important role from an environmental point of view. Thus, the development of methods for reliable monitoring of vegetation change is crucial. Our exercise goal will be to experiment with subpixel species classification.

Species Abbreviation
Calamagrostis villosa cv
Molinia caerulea mol
Nardus stricta nard
Bílá louka species and their abbreviations

Start with visualizing the hyperspectral image BL_hyperspectral_image.dat in EnMAP-Box, mapping Band 24 to Red, Band 14 to Green, and Band 8 to Blue, and stretching the values to Min and Max using a Cumulative count cut should give you an image resembling true colors. It is also possible to use a different band combination for the RGB composite.

Figure 2

Band Rendering settings to achieve true color visualization

Overlay the image with the training polygons BL_training_polygons.shp. Using "Layer Properties-> Symbology" and the Value stored in the field CLASS classify the polygons, so each species is visualized with a different color. Feel free to play around with the data. 

Determine the spatial and spectral resolution of the image from metadata. What is the wavelength range?

1. Endmember extraction

“A hyperspectral endmember (also known as ‘pure pixel’) is an idealized, pure signature of a spectral class. The pure spectral signature signifies the complete reflectance of a pixel exclusively occupied by a single surface material.” (Karbhari et al., 2015)

Endmembers can be acquired from an existing spectral library, by measuring spectra in the field or lab, or extracted from hyperspectral/multispectral images. The first task is to manually extract endmembers from the hyperspectral image.

Add the Pixel Purity Index (PPI) image BL_ppi.dat as an available data source and visualize it. Using the knowledge from Exploration of hyperspectral data using EnMAP-Box create a labeled spectral library with at least one endmember for each species.

The training polygons should give you an idea where the species are located, and the PPI image serves as a measure of pixel purity (the higher the value the purer). You can link multiple Map windows in “View -> Set Map Linking -> Link Map Scale and Center” to easily identify and select pure pixels located inside the training polygons.

Figure 3

Imported spectral library view BL_spectral_library.gpkg 
Compare the extracted endmember spectra with the semi-automatically pre-extracted endmembers located in the spectral library BL_spectral_library.gpgk  
Do they differ significantly and if yes, then why?

To obtain the provided PPI image, first the Minimum Noise Fraction (MNF) transformation was applied on the hyperspectral image. Based on the eigenvalues, only the first 11 MNF bands were considered. To reduce spatial dimensions, the PPI was then calculated on the MNF transformed data using a threshold factor of 2.5 and 20 000 iterations.

Figure 4

First five bands after applying MNF transformation

2. Spectral unmixing

Open the Regression-based unmixing application in EnMAP-Box. As the "Endmember dataset" choose either the provided spectral library BL_spectral_library.gpkg or the one you created during the previous steps. The library can be added using the “Create Classification dataset (from categorized spectral library)” option. The other input is the “Raster layer” on which the spectral unmixing is going to be performed, in our case that is the hyperspectral image.

Next, the Regressor algorithm needs to be defined. Both linear and nonlinear options are available. EnMAP-Box uses the Python scikit-learn library. More information on the different machine learning (ML) algorithms can be found in the documentation. Use an ML algorithm of your choice, but keep in mind that some algorithms are more costly than others, thus, to reduce running time, one of the following is advisable: LinearRegression, RandomForestRegressor, SupportVectorRegressor.

The process of creating the synthetically mixed spectra as training data from the endmember library is affected by the mixing parameters. “Number of mixtures per class” defines the total number of mixtures per class to be created. “Mixing complexity probability” consists of a list of probabilities for using 2, 3, 4, ... endmembers in the mixing models. Other options decide whether to include the original library spectra into the dataset or allow mixtures with profiles belonging to the same class. For the first run, you can leave all of these values at their defaults.

Figure 5

Regression based-unmixing workflow (Figure by Janz et al. 2023/ GPL-3.0)

Finally, the outputs need to be specified. Fill in the file destination for the “Output class fraction layer” and the “Output classification layer”, which is derived from the class fraction layer and corresponds to the maximum class fraction.

Now you are ready to run the application. 

Pick one or more of the mixing parameters and determine their changing effect on the results in terms of visual inspection of the output images, accuracy etc.

Figure 6

Regression based-unmixing application settings

Figure 7

Create Classification dataset (from categorized spectral library) settings

3. Results evaluation

Visualize the classification layer and the class fraction layer. You can create an RGB image from the class fraction layer using Symbology or the EnMAP-Box Processing Toolbox, respectively the “Create RGB image from class probability/fraction layer”. The RGB pixel color of a single pixel will then be given by the weighted mean of the given class colors. The weights are given by class fractions/probabilities (i.e. values between 0 and 1). For example, pure pixels with cover fractions of 1 appear in its pure class color. A mixed pixel with 50% fractions in two classes colored in red and green, appears in a dull yellow.

Figure 8

Create RGB image from class probability/fraction layer settings

A different way to assess the fraction layer is to use single band visualization. In "Layer Properties -> Symbology -> Band Rendering" choose Singleband gray as “Render type”. Each band corresponds to a class/species. The fraction of a given class/species in each pixel will then be defined by a grayscale value.

Figure 9

Single band visualization settings

Accuracy assessment of classification results should generally include computation of various accuracy metrics, such as overall accuracy, confusion matrix, user and producer accuracy, F1 score. However, visual comparison will suffice for the purpose of this exercise.

Examine the original hyperspectral image, all the unmixing outputs, the reference map below obtained using pixel classification, and try to answer the following questions: 

What does the class fraction layer represent?  
Describe the resulting RGB image, are there mixed/pure pixels?  

The northwest edges of compact mol areas seem to be mixed with cv, but not the other edges.  
Can you determine what is causing this effect?  
(hint: Carefully looking at the original hyperspectral image might help.)

The reference map contains the class desch which we did not classify.  
How are these areas manifested in the class fraction layer and the classification layer? 

Compare the classification layer and reference map.  
Do both maps show a good agreement in terms of spatial patterns? Discuss all the possible effects on the outcome.

What kind of data might be more suitable for subpixel classification and why?

Figure 10

Reference map obtained using pixel classification

References

Karbhari, K., Solankar, M., Nalawade, D. (2020): Hyperspectral Endmember Extraction Techniques. 10.5772/intechopen.88910.

Janz, A., Jakimow, B., Thiel, F., Linden, S., Hostert, P. (2022): EnMAP-Box 3 Documentation. enmap-box.readthedocs.io.

Exercise solution

Proceed to example solution Subpixel classification - report

Back to theme

Proceed by returning to Machine learning in imaging spectroscopy