Spatial contextual feature extraction
iota2 uses radiometric pixel-based features only. This limits the ability to correctly identify classes for which the information of the neighbouring pixels is discriminant (discontinuous urban fabric, for instance). Simple contextual features can help here:
- Local statistics (mean, variance, etc over a local window)
- Textures (Haralick)
- Edge density
- Gabor Texture Index
- PanTex
The latter 3 are described in https://doi.org/10.1109/IGARSS.2009.5417530. All 5 approaches are (partially for Gabor) available in OTB.
Some of the features listed above could be computed on a superpixel neighborhood.
We should decide if a new ContextualFeatureExtraction application is needed or if we add options to the existing feature extraction application. Also, the contextual features don't need to be computed for all bands and all dates. Maybe it is wise to give the choice of a list of channels in the input image stack. Since it is interesting to compute features on derived indices like NDVI, if a new application is built, it should come after the existing feature extraction.