the image look like there is not much dark areas inside. So you could start by computing the intensity of your RGB image (band 1 + band 2 + band 3) e.g. using raster calculator in ArcGIS, then you set a threshold (e.g. Con (intensity < 10 , 1, 0) using raster calculator). You can also do more complex testing such as directly Con(("band1" < 5) and ("band2" < 5) and ("band3" < 5), 1, 0). With gdal you can use "nearblack".
At this point, you have a nearly clean delineation of your area out of image. But for instance there could be some issues like on the top right of your image where you have something like 2 cloud shadows. One way to remove those artefacts is to combine a shrink and an expand (morphological mathematics). You erode your result by, lets say, 100 pixels, then you dilate it by 100 pixels. Small artefacts should disappear.
You could also convert to polygon and select only the polygon that are "large enough", but this will not correct the errors connected to the boundary. It would be quite fast to remove errors manually, though. Note that when you convert back to raster, you should specify in the environment seeting of the tools (processing extent) that you snap to the pixels of your image.