Automatically estimating the confidence levels of semantic segmentation predictions is important, as it can enhance accuracy and allow final users to have better insights into whether or not they can trust model results.

TerraTorch users should be able to add an (optional) confidence estimation step to semantic segmentation tasks to get confidence maps during inference (in addition to the standard outputs of the model).

One such method has been implemented in https://github.com/ESA-PhiLab/CAS_Confidence_Assessment, so I suggest we should integrate it into TerraTorch.