Advances in Active Learning for Spectral-Spatial Classification of Hyperspectral Data

Zhou Zhang, Purdue University

Abstract

Remotely sensed hyperspectral sensors can acquired data composed of hundreds of images corresponding to different wavelength channels, which are useful for discriminating between spectrally similar materials. With the detailed spectral information, supervised classification is typically conducted to produce a land-cover map to discriminate different classes by using training samples. Although the fruitful spectral information is helpful for classification, augmenting spectral data with spatial information for hyperspectral data classification has recently gained significant attention, as classification accuracy can often be improved by spatial information extracted from either fixed or adaptive neighborhood systems. However, the resulting high dimensional feature space often imposes significant challenges in developing robust supervised classifiers, especially when the quantity of labeled samples is limited. Active learning (AL), which aims to find the most informative training set, has become an effective approach for dealing with the limited availability of labeled samples. However, most AL methods presented in the literature for classification of hyperspectral data deal solely with spectral information. Furthermore, common AL strategies work in a fixed feature space, and the feature space is not updated as AL proceeds with potentially suboptimal performance. In this dissertation, three adaptive active learning strategies are proposed for hyperspectral data classification using both spectral and spatial features: * Single view based active learning with adaptive spatial information Since hyperspectral data is defined in both spectral and spatial domains, combining spectral and spatial features for its classification helps reduce the labeling uncertainty when utilizing only the spectral information. A new strategy that combines spectral and spatial information is developed. First, a new supervised strategy is proposed for segmentation hierarchy pruning that allows a best map to be determined and updated at each iteration of the AL process, thus exploiting informative labeled samples provided by the user. Then the resulting map is used to incorporate the spatial information via two strategies: 1) enrich spectral signatures with spatial features (adding features); 2) add pseudo-labeled samples through a new semi-supervised approach (adding samples). * Multi-view based active learning with view updating Besides the limited number of labeled samples, high dimensionality of the input data resulting from stacking different types of features is another challenge for supervised classification. To address this issue, the proposed single view active learning framework is extended to a multi-view setting. In particular, the hyperspectral data is partitioned into multiple disjoint views, and for each view, the spatial features are extracted from the corresponding best segmentation map and combined with the spectral features. For AL, a maximum disagreement query strategy is adopted. Moreover, a dynamic view is generated at each AL iteration to further increase the disagreement level which is important for AL query. * Regularized multi-metric active learning with updated regularizer To simultaneously address the issues of high dimensionality and limited labeled samples, a new strategy is proposed which combines feature reduction for multi-type feature data and batch-mode AL into a unique framework. To accomplish this, a regularized multi-metric learning approach is developed to jointly learn a specific metric for each type of feature, and the regularizer is refined by exploiting the increased labeled information provided by the user as AL proceeds. In addition, a batch mode AL query strategy considering both uncertainty and diversity criteria is developed in order to select the most informative samples in the final low-dimensional feature space.

Degree

Ph.D.

Advisors

Crawford, Purdue University.

Subject Area

Remote sensing

COinS