All the visualisations described above are automatically linked, once the beans that produced them have been linked using PropertyChange Events. Currently, the main effect of this linking is that once pixels are selected in a 2D display frame, an animation or a graph, the location and identity of those pixels will be communicated to any other surfaces and graphs which are open.
In the case of a fuzzy spectral signature, the spectral signature of any selected pixel will be overlaid on the plot, as shown. This allows you to see how closely the signature of that pixel conforms to the characteristic signature of that landcover class, and in which spectral bands it differs strongly.
Selecting numerous pixels can lead to many overlaid lines
Removed by reverting to the original plot
In the case of a scatterplot, each selected pixel is highlighted in red.
In animations, single pixels are highlighted with crosses, while digitised boxes and polygons are shown in red.
A group of linked visualisations can look like this;
In the visualisations above, the scatter plot (which shows genuine grass proportions against the predictions of a fuzzy classification) has been used to identify and select those pixels where correlation is particularly bad (the highlighted group which is strongly off the diagonal). In these pixels, the cover of grass has been severely underestimated by the classifier. When the pixels are located on the RGB image at top left, and the animation at bottom right, it can be seen that they fall at the edges of a large grassed area, and that this boundary mixing may be influencing the classification. The pixel signatures, overlaid on the fuzzy spectral signature at top right, are strongly consistent with the 'grass' signature in bands 1,2 and 3, but have lower reflectances in bands 4,5 and 6. These particular pixels are spectrally atypical of the pattern which the classifier has identified as its 'target' for classifying grassed areas, and this may have caused the errors in classification.
We can use fuzzy spectral signatures and surfaces to gauge which unsupervised spectral signature best represents a particular landcover. In the example below, pixels with very high concentrations of bracken (from the verification data) have been selected, and their spectral signatures mapped onto the 5 fuzzy classes in the example dataset. This helps to show the major deviations and similarities.
NEXT - FLIERS serialised file structure