I have a large data set of RGB satellite data that classifies 64x64 pixel images with a spatial resolution of 10m per pixel into 10 classes (e.g. highway, industrial, river, forest). Now I want to train a model (e.g. using resnet34 with fast.ai) on it. So far my model has a error rate of 4% on the test data.
Now I want to classify chunks of an unseen larger scene. My approach is to iterate over the new image and create n x n pixel subimages that get classified and create a stitched 10-color image out of the classified n x n images.
From a visual standpoint I get good results down to about n=32. Below that the predictions are just wrong. I assume that below that mark the model cannot recognize shapes in the chunks.
My question as a novice in data science are:
- Is it even legitimate to classify chunks of n<64 considering that the model was trained on n=64 images
- Is there a way to classify my unseen picture with a model built on the dataset on a per-pixel basis (i.e. n=1) or do I have to find a different dataset / create my own dataset?