Sentinel 1 and 2 fusion produces higher overall accuracy but unable to distinguish grasslands


I am carrying out classification of certain landscape types. Using Sentinel 1 alone, the overall classification is around 76% but user and producer accuracy of grasslands is around 50%. Now when I fuse Sentinel 1 and 2 together, the overall accuracy is higher at 88% but user and producer accuracies of grasslands goes down to 0%! Then when I use sentinel 2 alone, I get a user accuracy of about 25% for grasslands.

Does anyone have any idea why?



Using S-1 coherence should be useful over grasslands, especially if your area of interest is covered by 6-day repeat coverage.

Some ambiguity here. It depends on what the other classes are to understand producer and user accuracy. Also, the classes “grassland” and “certain landscapes” are rather fuzzy. Grassland can be anything from an intensively managed flat meadow to some swampy peat-bog in the Highlands. If your other classes include stuff like “semi-natural grazing areas”, “shrubland”, “fallow land”, etc. it’s little wonder that accuracies are not going to be stellar. Also, your choice of the classifier, the dates of the S2 imagery, etc. may play a role. Too many unknowns, therefore.

1 Like

Thanks. Can it work with 12 day repeat cycle?

Thaks. I didn’t give all the details: I actually have major classes: Built-up, water bodies, bare soil, grassland, hedges, woody vegetation, cropland. Grasslands haven’t been subdivided into lower classes. I suspect signal confusion between hedges and woody vegetation e.g since a hedge can be made up of the same components as the woody vegetation, and possibly croplands.