A Spatial Semi-supervised Learning Method for Mining Multi-spectral Remote Sensing Imagery

Abstract

Supervised learning, which is often used in land cover (thematic) classification of remote sensing imagery, has two limitations: first these techniques require large amounts of accurate training data to accurately estimate underlying statistical model parameters and secondly, the independent and identically distributed (i.i.d) assumptions made by these techniques do not hold true in the case of high-resolution satellite images. Recently, semi-supervised learning techniques that utilize large unlabeled training samples in conjunction with small labeled training data are becoming popular in machine learning, especially in text data mining. These techniques provide a viable solution to small training dataset problems; however, the techniques do not exploit spatial context. In this paper we explore methods that utilize unlabeled samples in supervised learning for classification of multi-spectral remote sensing imagery, while also taking into account the spatial context in the learning process. We extended the classical Expectation-Maximization (EM) technique to model spatial context via Markov Random Fields (MRF). We have conducted several experiments on real data sets and our classification procedure shows an improvement of 10% in overall classification accuracy. Further studies are necessary to assess the true potential and usefulness of this technique in varying geographic settings.

Keywords: MAP, MLE, EM, Spatial Context, Auto-correlation, MRF, semi-supervised learning, mixture models