Performing Supervised Classification on Sentinel Image using ArcGIS Desktop?

Question

Is it possible to perform supervised classification on Sentinel Satellite Image with the minimum error. I compared two raster, one being the manual classification of an area using Photoshop CS6 and the other being Supervised Classification with Interactive Supervised Classification (ArcGIS). The pixels that didn't match represented more than 70% (Confusion Matrix). So, this is the farthest I've gone. I need to know if there is more that I can do with only ArcGIS Desktop.

This is the Manual Classified Image with Photoshop CS6.

This is the raster using Interactive Supervised Classification (This tool uses training polygons)

As you see there is a big space that should be colored with violet in the second image. Obviously in the future using CS6 should not be what we always use but this is only used by compare and see if Interactive supervised classification works.

Answer 1

You are likely observing the difference between pixel based classification and object oriented classification. The Photoshop algorithm likely incorporates some form of image segmentation (i.e. a type of object oriented image analysis), whereas the ArcGIS pixel based classifier uses the Maximum Likelihood classification algorithm. Esri defines the difference between the two methods as follows (Source):

Pixel-based is a traditional approach that decides what class each pixel belongs in on an individual basis. It does not take into account any of the information from neighboring pixels. It can lead to a salt and pepper effect in your classification results.

The object-based approach groups neighboring pixels together based on how similar they are in a process known as segmentation. Segmentation takes into account color and the shape characteristics when deciding how pixels are grouped. Because this approach essentially averages the values of pixels and takes geographic information into account, the objects that are created from segmentation more closely resemble real-world features in your imagery and produces cleaner classification results.

Generally speaking, image segmentation based classification approaches achieve much higher accuracy than pixel based approaches. In fact, ArcGIS has a Mean Shift segmentation algorithm which works pretty well. Note that you will need to select a classifier such as Random Trees to classify your image objects created during the segmentation.