I am doing a supervised classification in ArcMap to classify land cover types in a study area in Africa. I am using RapidEye satellite imagery (Planet Labs) and they have 5 bands: blue, green, red, red edge and Near infrared.
I can classify soil and forest cover (see Classified Image) successfully but I'm unable to successfully separate out houses in the classification. I only want to classify the houses that have Iron sheet roofs (they are the small white dots circled in red on the satellite image). However, these houses are being misclassified as soil and road pixels (see classified image). Ideally, I want to classify 4 land cover types: soil, forest, house, and paved road.
Can anyone suggest a better way to separate out the spectra from human-made surfaces such as the iron sheet house roofs and paved roads? I would prefer to stick to supervised classification if possible, but am willing to try other classification methods. I tried object-based classification (segmentation) in ArcMap I couldn't get the 16 bit RapidEye image to classify properly in ArcMap and it kept showing a black square. And the unsupervised classification methods (e.g. ISO cluster and Class probability) had terrible results in ArcMap.