My task is to perform a supervised classification of landcover across a huge (20,000+ hectare) property. I need to classify into three classes: forest, regenerating trees, and soil. We use s Sentinel-2 imagery for this task.
However, both the randomforest classifier and maximum likelihood classifier get confused as some regenerating trees are as small as 1m2 and the image is 5m2. so they're surrounded by soil and we can't classify them accurately.
Is there a way to calculate the amount of tree in each pixel?
My current thought process is that we could make training samples that are as accurate as possible (and close as possible to spectrally pure). Then we could set a really narrow threshhold in the classification so only pixels which we are certain are tree/ground are classified. Then with the remaining unclassified pixels we perform some form of calculation based on colour to find how much vegetation could be in the remaining pixels.