javaftw
commented
4 years ago Here's how it might work:
- Before starting the classification process, for each Macroclass, the user creates a vector "classification bias masking (CBM)" layer.
- On each layer the user draws shapes roughly around regions where the associate MC has either high (white) or low (black) probability of occurring. Default is grey.
- When finished, all vector layers are converted to singleband greyscale images and added as (special) bands to the Band Set
- The user is ready to start the supervised classification.
- When the ROI is created to sample pixels for a class, the special CBM bands associated with the MC is also taken into account when creating the spectral signature, while the special CBM bands for the other MCs are ignored
- The previous step is repeated for each MC, and this way, each signature will also contain a bias for its particular MC
- When the classification algorithm computes the probabilities, the bias is then also taken into account
There is currently no feasible way of doing this manually, as there is no way to selectively disregard certain bands when computing signatures for specific MCs
The vector layer, when converted to a raster, will of course have very sharp edges around the shapes. Ideally, some sort of "blur" convolution should be performed to soften (feather) the hard edges. The ideal would be draw/paint/spray tools such as those found in image manipulation software.
This ties in with a better approach to managing Cs and MCs
semiautomaticgit
The Classification Bias Tool (CBT) allows the user to create "areas of likelihood of occurrence" of classes and/or macroclasses so that the classification of pixels within those areas are weighted/biased by the user's settings. Within such an area, a pixel classified as a MC which has a low (or high) likelihood of occurrence is then reclassified as a MC which has a higher (or lower) likelihood (set by user) AND is closest (spectral signature) alternative MC.
The user must be able to define polygons which define the areas of influence of the classification bias. The user must also be able to use a single-band grayscale raster images whose pixel intensities determine the magnitude of influence of the classification bias.
Polygons as well as rasters may be stacked and may overlap. Pixels inside the polygon area are all influenced 100%, whereas pixels influenced by rasters are influenced between 0/255 to 255/255 (the pixel intensity values). A "feathering" of the hard edges of the polygons can be achieved by (silently, background) converting the polygons to rasters, applying a "blur" convolution, and then using the resulting image as a raster influence source.