I am implementing a fully convolutional neural network (specifically U-Net) over a time series of Landsat 8 images to predict land cover change. I am following the demo notebook for the predictions. However, this notebook only trains CNN over a single image (using median over multiple years) to predict single image output of impervious surface.
What I want is to input multiple months of imageries for training to predict the respective monthly land cover change (that varies spatio-temporally).
I have found
Dataset.window() that can create windows from a time series of datasets (here), but have no idea how can it be implemented for my case of time series of Landsat imageries.
Any ideas or examples where this is implemented?