The input remotely sensed image (natural) exhibiting strong textural
characteristics is given in Fig. 6.1. This image represents a small region in Texas and is acquired by Landsat at a spatial resolution of 30 m. Image spectral information is restricted to the visible region (R, G, and B).
This region is predominantly covered by dense and sparse vegetation, which
exhibit interesting forms of texture at different spatial frequencies and orientations. There are also some linear features in the form of roads.
Arid and fallow lands are present in this input image. Fig. 6.2 represents the ideal classification result for the given input image.
The ISODATA classification result using spectral information alone is given in Fig. 6.3. The classification parameters used are (No of classes = 6-8, Iterations = 3, minimum number of pixels per class = 50 and max. standard deviation in class = 2). This result is with a lot of salt and pepper noise, due to the confusion between various classes arising due to textural characteristics.
The classification result using GLCM indicators with parameters (mask size 7*7, d = 1, α = 0, -45) is given in Fig. 6.4. From Fig. 6.4, it is clear that a lot of the error in classification has been resolved by using texture indicators.
The classified result using Laws texture indicators on the input image is shown in Fig. 6.5. The L5E5, L5S5 and L5R5 convolution masks were used. Although Laws serves as an excellent textural edge detector its results are not very suitable for ISODATA classification.
Figure 6.1 – Input Landsat image
Figure 6.3 – Classification result (color only)
A series of Gabor filtered image channels are shown in Figures 6.6 (a) – (d). The Gabor filter parameters for Fig. 6.6 (a) are (frequency = 1/8 , angle = 0, phase = 0, standard deviation of the filter = π). The Gabor filter parameters for Fig. 6.6 (b) are (frequency = 1/10 , angle = 90, phase = 0, standard deviation of the filter = π). The Gabor filter parameters for Fig. 6.6 (c) are (frequency = 1/15 , angle = 45, phase = 0, standard deviation of the filter = π). The Gabor filter parameters for Fig. 6.6 (d) are (frequency = 1/20 , angle = 0, phase = 0, standard deviation of the filter = π).
The classification output using a Gabor filtered images with parameters
(frequency = 1/8 , and 1/20 orientation = 0, 45, phase = 0, standard deviation of the filter = π) is shown in Fig. 6.7. The classification result derived using GLCM indicators (mean and angular second moment) on a second level Daubechies wavelet
decomposition is shown in Fig. 6.8.
A natural image chip corresponding to Antigua Island is shown in Fig. 6.9. This image is an IKONOS image has a spatial resolution of 4 m. Image spectral information is available in visible bands and NIR bands.
This image consists of high spatial frequency texture components corresponding to coral reefs, dense vegetation and has a medium-low spatial frequency component corresponding to sparse vegetation. Some urban features in the form of an airport are present in the image.
The results of a supervised classification on the original IKONOS image corresponding to Antigua Island are shown in Fig. 6.10.
The classified result using spectral information of the image is represented in Fig. 6.11. The ISODATA classification parameters used are (no. of classes = 5-6, iterations = 3, min. no. of pixels per class = 100 and std. deviation within a class = 2).
The classified result using GCLM indicators (angular second moment and mean) on Band 3 of the IKONOS image is shown in Fig. 6.12. This result helps us differentiate between shallow and deep water bodies, urban areas and fallow land. However there is some amount of confusion between vegetation and water bodies.
The classification result using Laws texture indicators (L5E5, L5S5) on the input IKONOS image is shown in Fig 6.13. Although Laws masks are excellent for identifying textural discontinuities, the classification results are inferior compared to other texture indicators used.
The classification result using Gabor filtered images are represented in Fig. 6.14. The filter parameters used are (frequency = 1/15 orientation = 0, 90 and phase = 0, standard deviation of the filter = π).
The classified result using GLCM indicators (mean and angular second moment) on a first level Daubechies decomposition on the input Antigua image is shown in Fig. 6.15.
A natural image chip, corresponding to an urban scene in St. Johns, Antigua has been used test texture analysis of urban features. This image is shown in Fig. 6.16. The aim of the classification process in this image is to capture the macro scale urban texture. When this image chip is classified using color information alone, there is a salt and peppery nature to the classified output (Fig. 6.17).
The classification result using GLCM indicators mean, dissimilarity and angular second moment (window size = 11, d = 1, and α = 0, 90) is given in Fig. 6.18. Laws texture indicators are not used for the analysis of this image, as Laws masks are not suitable for analysis of macro level textures; given that they are specific to a given scale and orientation.
The classification result using Gabor filter (frequency = 1/25 orientation = 0 and phase = 0, standard deviation of the filter = π) is given in Fig. 6.19. The classification result derived from using 2 level Daubechies wavelet transform and post processed with non-linear texture energy function (M = 5, α = 0.25) is given in Fig. 6.20.
(a) (b)
(c) (d)
Figure 6.7 – Classification result (Gabor)
Figure 6.10 – Classification result (supervised)
Figure 6.12 – Classification result (GLCM)
Figure 6.14 – Classification result (Gabor)