Temporal Dimensionality Reduction for Land Cover Map Production Using High Resolution Image Time Series

Abstract

Land cover mapping using high resolution image time series faces the issue dealing with high volumes of data which can threaten the ability of supervised classifiers to learn pertinent decision boundaries. Although dimensionality reduction approaches have been applied to hyperspectral imagery for a long time, their use with dense time series has not yet been explored. We study the usefulness of dimensionality reduction as a pre-processing step for high resolution optical image time series supervised classification for land cover mapping. Principal Component Analysis (PCA), Autoencoders and Ko-honen's Self Organising Map are compared over 3 dimensionality reduction approaches: global, per date and per band. Applying PCA to each date of the time series yields the best results in terms of classification accuracy.