Spatial and Temporal SAR Image Information Mining

Kurzfassung

In this paper, we focus on the development of new methods for spatial and temporal high resolution SAR image information mining. This paper proposes an entire and complete approach consists of mainly two parts, which are feature extraction and learning method for temporal evolution pattern indexing. Based on the intrinsic characteristics of VHR SAR images, Bag-of-Words (BoW) feature is developed for SAR image characterization, which can be applied on both static and temporal SAR images. As BoW feature is an intermediate feature, an appropriate local feature should be extracted. We show that, using only the basic local statistics of a mall neighborhood, i.e., local mean and variance, BoW feature could achieve better performance than Gabor texture features for SAR image classification. Evaluation of different local features leads us to an exciting finding that BoW feature using pixel values in a compact neighborhood as low level feature could achieve better performance than many other texture features. We develop as well a new features coding method, which is called incremental coding. Both this new yet simple features and incremental coding can achieve significantly better accuracy than state-of-the-art features and feature coding methods for SAR image classification. In addition, different aspects in BoW model have been evaluated and reliable conclusions are given based on the evaluation. The BoW feature has been extended to SAR ITS as well, giving a new Bag-of-Spatial-Temporal-Words (BoSTW), which has shown a better performance than the concatenation of other texture features. In the second part, a cascade active learning approach relying on a coarse-to-fine strategy for spatial and temporal SAR image information mining is developed, which allows fast indexing and hidden spatial and temporal pattern discovery in multi-temporal SAR images. In this approach, a hierarchical image representation is adopted and each layer is associated with a specific patch size. SVM active learning is applied at each level to obtain reliable samples and reduce the manual effort in labeling the images. Two components for classifier training using the labeled images and sample selection which selects the most informative samples for manual labeling work alternatively. When moving to a new level, all the negative patches are neglected and the learning at the new level focuses only on the positive patches. In this way, the computation burden in annotating large data set could be remarkably reduced while keeping the accuracy. In this method, we have solved another problem of training samples propagation between levels by multiple instance learning. In addition, we have proposed a new visualization method for SAR ITS using a simple color animation of the sequence. Three successive SAR images in the sequence are concatenated and represented as a color image, which is applied to all the successive images in the sequence. This simple color representation can significantly highlight the content variation while not distorting the information, which greatly facilitate the interpretation. Without any processing, we can easily observe many temporal patterns and content variation is completely visible. Through temporal pattern retrieval, the cascade active learning has been compared with a baseline SVM active learning operating only at the last level in terms of both accuracy and time complexity. We have demonstrated that cascade active learning can not only achieve better accuracy but also reduce remarkably the computation time.