A variance-based Bayesian framework for improving Land-Cover classification through wide-area learning from large geographic regions

摘要

Common to much work on land-cover classification in multispectral imagery is the use of single satellite images for training the classifiers for the different land types. Unfortunately, more often than not, decision boundaries derived in this manner do not extrapolate well from one image to another. This happens for several reasons, most having to do with the fact that different satellite images correspond to different view angles on the earth’s surface, different sun angles, different seasons, and so on.In this paper, we get around these limitations of the current state-of-the-art by first proposing a new integrated representation for all of the images, overlapping and non-overlapping, that cover a large geographic ROI (Region of Interest). In addition to helping understand the data variability in the images, this representation also makes it possible to create the ground truth that can be used for ROI-based wide-area learning of the classifiers. We use this integrated representation in a new Bayesian framework for data classification that is characterized by: (1) learning of the decision boundaries from a sampling of all the satellite data available for an entire geographic ROI; (2) probabilistic modeling of within-class and between-class variations, as opposed to the more traditional probabilistic modeling of the “feature vectors” extracted from the measurement data; and (3) using variance-based ML (maximum-likelihood) and MAP (maximum a posteriori) classifiers whose decision boundary calculations incorporate all of the multi-view data for a geographic point if that point is selected for learning and testing.We show results with the new classification framework for an ROI in Chile whose size is roughly 10,000 square kilometers. This ROI is covered by 189 satellite images with varying degrees of overlap. We compare the classification performance of the proposed ROI-based framework with the results obtained by extrapolating the decision boundaries learned from a single image to the entire ROI. Using a 10-fold cross-validation test, we demonstrate significant increases in the classification accuracy for five of the six land-cover classes. In addition, we show that our variance based Bayesian classifier outperforms a traditional Support Vector Machine (SVM) based approach to classification for four out of six classes.