Semisupervised Hyperspectral Image Classification Using Spatial-Spectral Information and Landscape Features

In hyperspectral image classification, the foremost task is that: how can we apply limited labeled samples to achieve good classification results? Spatial–spectral classification methods, which assign a label to each pixel regarding both spatial and spectral information, are effective to improve classification performance. Moreover, semisupervised learning (SSL) focuses on the scenario that the number of labeled data is rather small while a large number of unlabeled data are available. To complement spatial–spectral classification methods and semisupervised learning for each other, we propose a novel learning landscape features semisupervised framework (LLFSF) based on M-training algorithm and weighted spatial-spectral double layer SVM classifiers module (WSS-DSVM). In this novel framework, we first propose a SLIC (simple linear iterative clustering) based non-local superpixel segmentation algorithm to initially learn landscape feature and spatial composition. Then, we apply WSS-DSVM module to obtain initial classification maps. To better characterize complex scenes of hyperspectral images, we quantizes both the landscape diversity and separability from the initial classification map, which increase availability of spatial details and structural information of objects. Finally, we put some patches with lower accuracy into Multiple-training algorithm for further classification. In order to achieve an unbiased evaluation, we have evaluated the performance of LLFSF on three different scene hyperspectral data sets and compare it with that of three state-of-the-art hyperspectral image classification methods. The experimental results confirm the efficacy of the proposed framework.