Simple fusion scheme at region boundaries
In issue #361, iota2 provides data to study the behaviour of classifier at boundary by extracting the points.
In this issue, a simple fusion scheme for the boundary area is proposed. The files produced in #361 could be produced during steps used to produce the land cover map instead of be independent tools.
The current workflow should be modified:
- The boundary area must be automatically defined from the region file and a boundary size parameter (new step, exists in #361)
- The classification step must no longer apply the original region mask
- Sum the chosen metric: Probabilities (costly to obtain, limited to some algorithms), confidence or other
- Weight the sum by the distance to the boundary (optional).
- Take the decision for each pixel in boundary
- Write the reconstructed tile
- Back into the standard workflow for produce the mosaic using all tiles and run until the last
validation
step.
In this case, the algorithm is pixel oriented, and there is no relation between tiles where a pair of boundary is seen. That means that the reconstruction of each tile merging all regions can be done once the classification by each model is done. It's quite different than the issue#361 where the matrix agreement should merge all pixels from all tile for a given pair of regions.
Some particular cases (not exhaustive):
- chunks: all algorithms using chunks (scikit learn, external features, DL and GP) use the region mask to not process masked area. Then the mask used should take into account the boundary are to classify the region + the boundary and not the entire tile.
- When more of two regions intersects one tile. Should the algorithm consider only one pair of region at time ? And how handle the case when several regions intersect. In the central zone, each pair of regions provides pixels.
- Maybe the total area of a region over all tile considered should be take into account. In the case of the following image, the green region covers only the bottom right corner of T31TCJ tile. If only this tile is considered, the amount of classes for this model is very small (two classes), then in this case, the region should be ignored ? In the case of multiple tiles, the region model contains 10 classes which over a better land cover description.