Bag-of-Words model based image classification and evaluation of image sample quality in remote sensing

In this paper, we propose to use Bag-of-Words (BoW) model based methods for the remotely sensed images classification, and also attempt to do research about the image quality and algorithm evaluation. We introduce the Locality-constrained Linear Coding (LLC) for the remotely sensed image classification, and the hard vector quantization and soft vector quantization are adopted for experiments. Regards to the algorithm evaluation, we propose to use a hierarchical model for the image sample quality evaluation as a starting point. The model includes global, class and image level quality, which considers both the image quality and the class separability. Experiments are conducted on the image database selected from google earth and three synthesized databases. Experimental results show that the BoW model based methods are suitable for the remote sensing image classification. However, the advantage of LLC is not as significant as in nature scene image classification. Results of sample evaluation show that the number of middle quality images in a database is the largest, and the quality distribution of a database is consistent with the classification results. The hierarchical model provides an effective quality measurement of remote sensing image samples.