The example dataset.
The file consists of a number of raster
data layers, which cover the same ground area in the Peak District. This
multi-layer data set has been produced for the purposes of the FLIERS project,
which evaluates the use and quality of fuzzy multi-spectral classifications. The
aim of the FLIERS study is to obtain good predictions of mixed ground cover
within pixels by using fuzzy landcover classes.
The Peak district data covers an area close to Froggat Edge, and contains:
What is in the example dataset?
Some of the principles behind the fuzzy classifications will be demonstrated below on a 1km by 1km area of urban Leicester. The pixels in each layer correspond to the pixels from a Landsat TM image, which would be included in the dataset. When the file is read in, each layer can be seen to have a name, which explains the data it contains.
These layers contain different types of
data, as follows;
Spectral reflectances: These are sections of a Landsat TM image. Each has 6 separate bands, which correspond to bands 1,2,3,4,5 and 7 of the TM image. They are still in their original 8-bit scaling, so that each layer contains values between 0 and 255. (These values represent the reflectance recorded by the satellite in that spectral band, at that location. 0 corresponds to no reflected radiation, and 255 to the maximum possible).
The grey-scale values show how much
reflected light the satellite picked up, in this particular waveband. Some areas
of the image reflect very weakly, and show as dark, particularly areas of water.
Others are bright, particularly areas of healthy vegetation. When all 6 layers
of TM data are combined together, a spectral signature can be derived for each
pixel, showing its score in all 6 bands.
This spectral signature can be compared to characteristic, known signatures for a variety of landcovers. The graph below shows two typical spectral signatures, derived from many samples of satellite data. It shows the typical mean score in each wave band.
The clustering process used in thematic classifications is described here.
Verification data: These layers show the amount of a particular cover type which is present in each of the Landsat pixels. These values are gathered as follows;
1) The ground study area is segmented into different cover types, using ground survey and photo interpretation. The different land covers used in the classification are listed here.
2) This vector coverage is transformed to the Landsat image, so that the digitised features correspond to the pixel values recorded in those locations.
3) The pixel boundaries of the Landsat image are overlaid on these vectors as a grid, and for every pixel, a ground cover proportion is calculated for each cover type.
The final verification layer is scaled between 0 (0% cover) and 1 (100% cover), and looks like this.
Mask layers: These are boolean raster layers which mask out areas which should not be included in analysis. These could be pixels where data is unreliable or unknown, areas outside the extent of the study, or known artefacts (e.g. doubled pixel stripes). A mask layer contains values 0 in pixels which should be ignored, and 1 for valid pixels. Mask layers are automatically referred to in many procedures, such as calculation of layer statistics and graph production, to ensure that invalid or background pixels do not bias calculations.
Fuzzy memberships: These are produced by fuzzy classification of the multi-spectral data for the area, (using the FLIERS classification software). The classifier is trained on known ground cover proportions, and then used to predict ground cover proportions from the Landsat reflectances. The result is a map of predicted memberships for each of the ground cover classes. These data are also scaled between 0 and 1. If the classification is accurate, there should be a 1:1 correspondence between a verification data layer and the corresponding fuzzy membership layer. In fact, for many reasons, there will be inaccuracies, and they may vary across space. These inaccuracies, and their pattern, can be visualised and explored using the graphs in this toolkit. The two images below show predicted proportion for grass and tarmac, produced using a fuzzy classifier. (Masked pixels are not included in the classification, hence the gaps in the image). It can be seen that these values differ from the verification data above.
Vector layers: Line or point vector layers can be
imported and displayed to add meaning to surface visulaisations. These layers
are not topologically structures, and cannot be used for point-in-polygon, or
other relational analysis. A point layer can be imported from an ASCII table of
values (see Vector point import), and allows the storage of spatially referenced
points, along with any number of attributes for each point. These points can be
coloured according to single attributes or combinations of attributes.
NEXT - Data reading