I am interested in training a neural network to perform pixel-based classification for land cover mapping of satellite data. In order to learn more about this interesting field I am trying to reproduce the results from this paper:
I currently have 30 satellite images of my area of interest per year for a total of 5 years, and am in the process of creating the training data set.
Pixel-based classification: Classification is done on a per pixel level, using only the spectral information available for that individual pixel (i.e. values of pixels within the locality are ignored). In this sense each pixel would represent a training example for a classification algorithm, and this training example would be in the form of an n-dimensional vector, where n was the number of spectral bands in the image data. Accordingly the trained classification algorithm would output a class prediction for each individual pixel in an image.
While understanding that the pixel is the fundamental unit of a satellite image and so it is natural to analyze spectral information contained within, I do not understand how this is used in order to process a whole region which may contain 50,000 pixels.
The authors use a 30m data set. This is represents 3 pixels from Sentinel satellite imagery.
My question is: If I train a neural network on a 30m data set to classify between vegetation and buildings, how would it be possible to test the network to classify an image containing, for example 10,000 pixels?
If I train a CNN on images containing only 3 pixels how can I test effectively on larger images such as the one below?
This is an example of a result I would like to replicate, however am unable to understand this concept which is vital for analyzing spectral information: