Evaluation of the Textural Statistics of the Gray Level Co-Occurrence Matrix Performance for Change Detection

Geomatics science and technology is a main source of geospatial information providers in variant forms. In producing such information, Remote Sensing and Photogrammetry with high potentials as different sensors with variant data types have an essential role. In recent years, change detection has been important for many organizations in attention to the nature of the change. Until today, many change detection methods as direct comparison, Transformations, classification based and so on have been used. Each one of these methods with considering to the final goal has been employed variant type of data and different results have been obtained. The image in the visible and near infrared part of the electromagnetic wave has the most usage as an input data for change detection. According to the nature of the change detection method, different feature spaces as textural and morphological features for increasing the accuracy of the produced results have been tested. Textural statistics extracted from the Gray Level Co-occurrence Matrix (GLCM) with high variant has different effects on the classification and change detection obtained results. But using all kinds of the textural features for increasing the accuracy of the produced results will make problems because of the correlation between classes and sometimes because of the high noise values and also in some cases with more increasing the feature spaces and so decreasing the processing speed. In this paper, comprehensive performance evaluation of each textural statistic: Mean, Variance, Homogeneity, Entropy, Dissimilarity, Second order moment, Contrast and Correlation on improving the change detection accuracy have been done. For this, firstly three spectral bands of Landsat 8: 2, 5 and 7 bands, in the visible, near infrared and mid infrared region of EM wave scince 2013 and 2015 years selected as input data. Then, mentioned textural statistics in 4 directions (0, 45, 90 and 135 degree) on the each considered bands and years are extracted. After that, for eliminating the direction effect on the features the mediocre of all the 4 directions for each statistic are estimated. Afterward, differential image is estimated by subtraction of corresponding textural bands of the 2013 and 2015 years. Also, the differential images of the spectral bands: 2, 5 and 7 are produced. Then, every differential image of the textural statistics (in each band independently) is integrated with differential image of the spectral bands and are fed to the Maximum Likelihood classifier. The obtained results are shown, in visible and near infrared region, Mean statistic increased the overall accuracy about 15 and 16 percent respectively, and improved the Kappa coefficient value about 30 and 31 correspondingly and have the most influence on increasing the accuracy of the change detection output results. Also, in the both of mentioned EM regions, Second order moment, Contrast and Correlation statistics have the least effect on the change detection accuracy improvement. In the mid infrared region, approximately all statistics have the same performance. Furthermore, all features of the GLCM are combined and fed to the Principal Component Analysis (PCA) band reduction technique and first band selected. Then, the first band of the PCA is employed as other features for change detection. Obtained results have been shown the efficiency of this strategy for accuracy improvement. The achieved results of this paper on tests data have been approved that Mean, Entropy and Homogeneity have the highest improvement performance and Variance, Correlation and Contrast have the lowest improvement performance on the change detection accuracy.