|
|
|
Fusion of ASTER image data for enhanced mapping of landcover features
Detailed inspection of the fused images reveals that the major categories of rocks has been very well demarcated in comparison to the original VNIR & SWIR images. The individual outcrop boundaries have been distinctly displayed due to the improved spatial resolution and spectral content in the fused images. Thus pockets of charnokite, gabbro and pneiss are very well represented in the fused images, leading to improved information extraction about these features.
The barren patches of soil are represented with better contrast in the fuse images due to increased reflectence of soil in the SWIR region compared to VNIR region.
Vegetated features such as plantation, agricultural areas, dense and moderately dense forest, are better represented due to fusion with SWIR data.
Because of the relatively higher reflectance of water in visible region and almost total absorbtion in the SWIR region, enough contrast between water and land features is provided in the fused images.
Having observed that multi sensor fusion of ASTER image data enhances the land cover features, an attempt was made to prepare landuse/landcover map by classification of fused image data.
This exercise was attempted despite the fact that many studies have demonstrated that classification of fused images does not result in appreciable increase in accuracy and that certain features tend to get disturbed in terms of their spectral content Musa (2000). Apart from these limitations, synthetic pixels/artifacts are generated Schowengerdt (1997)
Supervised classification of VNIR, SWIR and fused images was attempted using training sets that were common to VNIR and fused images. A separate training set was generated for SWIR image because of its coarser resolution. The training sites comprised of classes belonging to water bodies, agricultural vegetation, forested vegetation, built-up area, four different rock types and mined area. The sites were selected using a region growing algorithm to avoid error in choice of pixels for training class. The number of pixels in a training site for a given class was decided based on the proportional representation of the class. Care was also taken to adhere to the rule of 3N pixels per training class where N represents the number of bands.
Classification of visible and NIR resulted in a map (fig.4) that showed the eight land cover classes, while water body and vegetation classes were displayed without much misclassification, the other categories of classes such as urban/built-up and rock types were displayed with some amount of misclassification. This is perhaps due to the lower signature separability of urban and rocky classes in the VNIR bands. The overall accuracy and kappa coefficient values are 58.84 % and 0.54 respectively. Such a low accuracy is due to the mis-classification of the rock type and urban classes in the VNIR image data. The overall accuracy and kappa statistics for SWIR are 53% and 0.47 respectively. This is mainly due to larger inaccuracies contributed by water and vegetation classes and perhaps due to the lower spatial resolution (30m) (Fig. 4a-4b).
Since lower accuracy was obtained while classifying the VNIR and SWIR images, classification of fused images was attempted. The fused image was obtained by merging VNIR (15m resolution) and SWIR (30m resolution) bands to result in a 18m VNIR + SWIR images with 15m resolution. Classification was attempted on this image using the same training sites as that for VNIR image. The resulting classified image proved to be better than the classified image obtained from VNIR image. The four rock types were represented with greater accuracy in the same manner as water body and vegetation classes. Improved accuracy may be attributed to the complementary information provided by VNIR and SWIR images. While the information from SWIR image contributed to differentiate between rock types, information from the visible and NIR bands contributed to distinguish between waterbodies and vegetation classes. Thus an overall accuracy of 66% and kappa value of 0.61 was obtained .
In this study, classification was attempted not as a tool to evaluate the quality of the fusion. However, the training sites used for classification of VNIR image were also used for classification of the fused image. This was done to verify the accuracy of classification of fused images. It was found that accuracy of classification improves on fusion of images. The classification exercise was performed not to achieve very high levels of accuracy but to demonstrate that classification accuracy improves in fused images.
Conclusion
This study has evaluated the potential of ASTER VNIR and SWIR sensors for landuse / landcover mapping. The capabilities of individual sensors and the fused outputs were evaluated by visual observation and by image classification.
While the individual VNIR and SWIR images provide little information about land cover classes, it was observed that fusion of images obtained from these two sensors resulted in enhancement of land cover features for improved mapping of the same. Information in the VNIR image contributed to the enhancement of vegetation and water classes. Rock and soil units were enhanced due to the contribution by the information in the SWIR images. Improvement in spatial resolution of the SWIR images and the contribution of complementary information present in the SWIR images resulted in enhancement of soil and rock units. Improved contrast between waterbodies, vegetation and rockunits was also observed. Thus the potential of ASTER data and fusion of the VNIR and SWIR images for enhanced landcover mapping has been demonstrated in this paper. Further study is needed in the area of classification of fused ASTER images.
Acknowledgement
The authors thank the ERSDAC (Earth Remote Sensing Data Analysis Center), Japan for providing the ASTER images to them for ARO on ASTER data use. The rights on image data belong to the Ministry of Economy, Trade and Industry, Japan. Institute of Remote Sensing and Centre for Geoscience and Engineering, Anna University are thanked by KV for providing the necessary support for this research.
References
- Chavez, P. S. Sides, S. C. and Anderson, J. A. 1991. Comparison of three different methods to merge multi resolution and multi-spectral data: Landsat TM and SPOT panchromatic. Photogrammetric Engineering and Remote Sensing. Vol 57. No. 3. pp. 295-303.
- ERSDAC, 2000. Advanced Spaceborne Thermal Emission and Reflection Radiometer, URL: http://asterweb.jpl.nasa.gov/
- Goward.N., Huemmrich.F and Waring.H 1994, Visible-near infrared Spectral reflectence of landscape components in western oregon : Remote Sensing environment Vol.47, Pp190-203
- Lakshmi, K., 2000. Improved information extraction using image fusion techniques. Unpublished M.Tech Thesis. Anna University.
- Musa.M.K.A, Hussin.Y.A. 2000 Multi-data fusion for sustainable forest management : A case study from northern part of selangor, Malaysia, International archives of Photogrammetry and Remote Sensing Vol.XXXIII, part B7
- Pohl, C. 1999. Tools and methods for fusion of images of different spatial resolution. International Archives of Photogrammetry and Remote Sensing. Vol 32. Part 7-4-3. Valladolid. Spain.
- Pohl. C and J.L.Van Genderen,1998. Multisensor image fusion in remote sensing : Concepts, methods and applications. Int.J.Remote Sensing,19(5),pp.823-854
- Schowengerdt, R. T., 1997, Remote sensing models and methods for Image processing, 2nd edition, Academic Press, USA.
- Vani, K. 1999. Merging and Analysis of multi-resolution satellite images. Unpublished M.E Thesis, Anna University.
|
|
|