Modelling of agricultural SAR Time Series using Convolutional Autoencoder for the extraction of harvesting practices of rice fields

Abstract:
We apply an unsupervised learning methodology to project SAR Time Series of growing rice fields onto a 3-dimensional space, where we explicit differences between the fields. The projection method used is a Convolutional Autoencoder, trained using a reconstruction task and a mean-square cost function. The chosen embedding space is of dimension 3, to provide the possibility to visualise it spatially using an RGB false colour composite. We compare two subsets of rice fields at both embedding space and original SAR time series levels to analyze the nature of the variations between the two subsets.