Unsupervised Quaternion Feature Learning for Remote Sensing Image Classification

Bag-of-words image representations based on local descriptors are common in image classification and retrieval tasks. However, in order to achieve state-of-the-art results, complex hand-crafted feature filters and/or support vector classifiers with nonlinear kernels are needed. Compared with hand-crafted features, unsupervised feature learning is a popular alternative, which results in feature filters adapted to the problem domain at hand. Although both color and intensity are important cues for remote sensing image classification and color images are commonly used for unsupervised feature learning, most of the existing algorithms do not take into account interrelationships between intensity and color information. We address this problem using quaternion representation for color images and propose unsupervised learning of quaternion feature filters, as well as feature encoding using quaternion orthogonal matching pursuit (Q-OMP). By using quaternion representation, we are able to jointly encode intensity and color information in an image. We obtain local descriptors by soft thresholding and computing absolute values of scalar and three vector parts of the quaternion-valued sparse code. Local descriptors are pooled, power-law transformed, and normalized, yielding the resulting image representation. The experimental results on UC Merced Land Use and Brazilian Coffee Scenes datasets are comparable or better than the state of the art, demonstrating the effectiveness of the proposed approach. The proposed method for quaternion feature learning is able to adapt to the characteristics of the available data, and being fully unsupervised, it emerges as a viable alternative to both hand-crafted representations and convolutional neural networks, especially in application scenarios with scarce-labeled training data.