Most of common classification algorithms in remote sensing are based on spectral characteristics of pixels. These approaches ignore the spatial information of data, such as texture, in classification process. Simultaneous usage of texture and spectral information is a new trend in remote sensing image classification which has been considered in this study. We have evaluated the efficiency of gray-level co-occurrence matrix texture features (GLCM) extracted from panchromatic (PAN) image of ALI detector in improving the classification accuracy of Hyperion hyperspectral (HS) data in urban regions of Tehran. Classification is performed using a support vector machine (SVM) classifier with a Gaussian kernel. In our experiments, we have considered three different cases: a) classifying original Hyperion data, b) classifying Hyperion data pansharpened by color normalized transform (CNT), and finally, c) simultaneous use of GLCM texture features of panchromatic data and the spectral features of pansharpened data in classification process.
In case b and c, for pansharpening HS data we have performed the following steps:
- Registering HS data with PAN data using
- nineteen ground control points,
- polynomial warping of second order, and
- Nearest neighbor interpolation.
- Selecting a subset of HS bands which spectrally overlap with PAN image.
- Fusing the spatial information of PAN image into the HS subset bands, obtained in step 2, using CNT method.
Moreover, as GLCM features, we have extracted 8 texture features from GLCM matrices: mean, variance, homogeneity, contrast, dissimilarity, entropy, angular second moment, and correlation. In order to assess the influence of the size of GLCM extraction window on the quality of texture features, we have considered various window sizes: 3×3, 5×5, 7×7, and 9×9.
At the first phase of our experiments, we compared classification results obtained using the original HS data with the results obtained from the pansharpened HS subset – see Table 1. The results showed an increase of about 15% in the average classification accuracy when using the pansharpened data. At the second phase, we combined each of the texture features individually with the pansharpened HS subset. The results are given in Table 2. As the table suggests, regardless of the type of texture feature and the size of the GLCM extraction window, the combinations improve the overall classification accuracy (OA) of data. However, texture features show better quality when extracted from GLCM matrices obtained using 9×9 neighborhood windows. In addition, we observe that regardless of the size of GLCM window, dissimilarity feature delivers the best results.
To summarize, by using the pansharpened HS subset instead of the original HS data, we achieved about 15% gain in the classification accuracy. Moreover, combining dissimilarity texture features –extracted from GLCM matrices obtained using 9×9 neighborhood windows– with the pansharpened HS subset improved classification results. In our experiments, we achieved about 5% increase in OA compared to that of using pansharpened HS subset alone.
Table 1. Comparison of classification accuracies obtained using original HS data with those obtained from pansharpened HS subset.
Table 2. Comparison of overall classification accuracies (OA) of “pansharpened HS subset” with those of “pansharpened HS subset + GLCM texture feature*”.