Introduction
Remote sensing is a powerful tool for the regional mapping of natural resources. With the use of imageries during the early stages of development of remote sensing in the mid-seventies, adequate progress has been achieved in the data interpretation. Digital processing of remotely sensed data has gained momentum in the last ten to fifteen years, especially with the availability of digital data. In India, with the establishment of remote sensing agency, attention is focussed on large-scale data processing for natural resource evaluation. One important aspect in remote sensing is the categorisation and classification of spectral measurements taken from satellite sensors into various features on land surface. Recognition of patterns for classification can be carried out if appropriate procedures are adopted. General classification methods have been developed using the image statistics, and their applicability to the processing of data is limited due to the spatial variation of natural resources.
The objective of the study is to describe and compare select image fusion techniques, namely: Band overlay, High Pass Filtering (HPF), Intensity-Hue-Saturation (IHS), Principal Component Analysis (PCA) and IMGFUSE. These techniques are both visual and statistical, in character. This study utilizes a multi-band data from:
-
IRS-1C LISS-III (26 February 1998 with 23.5 m spatial resolution)
- IRS-1C panchromatic (26 February 1998 with 5.8 m spatial resolution)
- RADARSAT-1 F1 Ascending (10 July 1999 with 6.25 m spatial resolution)
This is by way of determining the best image fusion technique for land cover mapping for natural resources management.
The present paper is part of a large-scale research, jointly conducted by the University of Madras, India and the University of Waterloo, Canada researchers. The discussion is very brief (Marino, 2001). Only select results are also displayed. The study assesses the utility of multi-band data for the study of land covers which are often difficult to accurately examine with remote sensed data and intended to examine the classification of land cover classes for different band combinations and the potential of the classification methods.
The rationale for classifying the enhanced images generated from the image fusion is two-fold:
First, it is necessary to classify the enhanced images because one of the goals of this study is to produce land-cover map for the area. The idea of classifying the fusion-generated products is based on the assumption that image fusion techniques do provide enhanced images with more information from a broader range of the electromagnetic spectrum than the original data, which can be used to better distinguish features during classification, especially with the combination of optical and radar data (Solberg et al., 1994)
Second, several researchers have used classification to evaluate the success of fusion techniques and to assess the quality of the fusion generated products by comparing classification accuracies to those obtained from the original imagery (for example: Franklin and Blodgett, 1993; Munechika et al, 1993; Haack and Slonecker, 1994; Dwivedi et al., 1997; Sunar and Musaoglu, 1998).
Classification Accuracy of Image Fusion Methods
Combination of Methods. Researchers used a combination of methods to determine which method improves classification accuracy. Franklin and Blodgett (1993) described the accuracy of maximum likelihood classifications (MLC) performed on the original multispecttral and panchromatic SPOT data and several fusion data sets. This was done for the discrimination of various land cover types in an alpine environment in southwest Yukon, Canada. Using a regression model, the PCA and the IHS, they have achieved classification accuracies of 54 per cent, 59 per cent, and 65 per cent, respectively. The IHS provided a 9 per cent increase in classification accuracy over the multispectral data alone and an increase of 25 per cent over the results from the panchromatic data. Based on these accuracies, they concluded that, quantitatively, the IHS method produced the best fusion image. It was therefore preferred over the original multispectral imagery for vegetation and the land cover discrimination.
Similarly, Sunar and Musaoglu (1998) used the IHS transform to merge a contrast-enhanced SPOT panchromatic image with Landsat TM data for land cover mapping in urban Istanbul. They performed maximum likelihood classification (MLC) on both the original multispectral Landsat data and the SPOT-TM merged data and found the merged data provided an increased classification accuracy of 6.5 per cent because of better spatial resolution. The merged data improved seperability of land cover classes and effective visual interpretability. Accuracy assessment of classification results is not possible because of inadequacy of ground truth information that was collected.
Evaluating the success of the fusion techniques is based on visual comparisons of the classified images and examination of the spectral signatures generated within each image relative to the band-overlay technique, which does not alter the spectral characteristics of the imagery.
K-Means Clustering Algorithm. In order to generate classified images from the image products generated using each of these image fusion techniques, an unsupervised classification procedure using the K- means clustering algorithm was used. An unsupervised classification approach was used, as there was inadequate ground truth information for selecting the training site data to be used for supervised classification. In addition, because the spectral characteristics of the enhanced images are under investigation, it is more desirable to examine clusters generated based on the spectral characteristics of the enhanced images. Because unsupervised classification is based on the groupings of the spectral properties of the pixels (Jensen, 1994), each cluster is statistically separable.
The K-means algorithm is one of the simplest and the most common clustering algorithms (Schowengerdt, 1997; PCI, 1998). According to Haack and Slonecker (1994:1256), when classifying complementary data sets, there is an inherent appeal in using a simple classification logic that can draw upon the basic synergistic nature of the fused individual sensors. When using the K-means algorithm, the analyst specifies the number of spectral clusters to be located within the image. Each pixel in the training set is assigned to the cluster whose mean vector is closest to the pixel vector. A new set of class mean vectors is calculated based on the results of the previous classification and pixels are assigned to new clusters. The iterations continue until there is no significant change in pixel assignments from one iteration to the next or until the maximum number of iterations as specified by the analyst is reached. The final cluster mean vectors are then used to classify the entire image using a minimum-distance classifier. The minimum-distance classifier assigns pixels to classes based on the shortest distance between each pixel and the class means. The analyst can specify whether class signatures are generated. Signatures can assist in the identification of the spectral classes or they can be used as input to a hybrid classification.
The K-mean algorithm was applied to each image-fusion product specifying that ten spectral clusters be generated for each classification. By keeping the specified number of clusters constant, performance of the classification based on the spectral characteristics of the image fusion results could be analysed. Spectral response curves were generated for each cluster identified by the clustering algorithm.
Classification Results of Six Images
The classification results from each of the six images generated from the image fusion techniques using the panchromatic data are:
-
HPF Technique. Compared to the signatures generated based on the band-overlay classification, the signatures from the HPF classification are quite different. The HPF technique has resulted in changes to the spectral characteristics of the transformed data set, which are exemplified in both the classified image and the spectral response curves (Figures 1 and 2).
- IHS Technique. In order to investigate the effect that stretching the panchromatic band to match intensity component has on classification, both the non-stretched and stretched fusion images were classified. The results obtained from classifying the IHS image with panchromatic band stretched to match the intensity component is quite similar to the band overlay image with the panchromatic data and is a considerable improvement over the results of that used the non-stretched panchromatic data (Figures 3 and 4).
Stretching the panchromatic band to match the intensity component prior to converting the IHS components back to RGB space has a positive impact in terms of classification. Because the descriptive statistics of the stretched IHS-transformed data are more similar to the original data than the non-stretched data, the classification results are very similar to the results using the original non-enhanced data and are therefore considered to be more accurate compared to the non-stretch results.
Since the spectral data are used to generate the signatures are so similar to the band-overlay image and it is this information on which classification is based, the signatures for each land-cover class are very similar in shape to the band-overlay signatures. The mean brightness values are slightly lower than band overlay (Figures 5 and 6).
- PCA Transformation Technique. The classification results based on the image generated using PCA transformation technique is very dissimilar to the classified image based on the band overlay fusion technique (Figures 7 and 8).
Evaluation
The classification based on the IHS image (with the panchromatic stretch) was very similar to the band overlay image. Because the results from the two techniques are so similar, there is no advantage to use this technique over the simpler band-overlay technique.
Evaluation of the RADARSAT-1 data revealed them to be complementary in terms of separating the villages from the land cover class. Apart from the village area, the rest of the images were classified well using band overlay. A ‘sequential masking’ classification procedure introduced by Ehrlich and others (1994) was selected to classify the whole area. The sequential masking approach to classification is based on photo interpretation technique whereby the most distinct image features are classified first. This sequence continues until the entire image is classified.
Conclusions
The classification results show that the band-overlay image was considered accurate in terms of its spectral information. The classified results are used as a basis to compare the other results. The results from the HPF technique were not as good as band overlay. The slight changes in spectral characteristics were exemplified in both the classified image and the spectral response curves. Better classification results were obtained using the IHS image with the stretched panchromatic band. The mean brightness values for each land cover type are very similar to those associated to the band overlay image. Classification of the PCA image yielded poor results. The class signatures reflect the radiometric distortion resulting from the use of this technique. The classified image and spectral signatures resulting from the IMGFUSE image are very similar to those of the band overlay image because the radiometric information of the original data is maintained.
Acknowledgements
The authors acknowledge with deep gratitude and a sense of thanks the munificent grant from the Shastri Indo-Canadian Institute – Canadian International Development Agency for the research reported here under the Partnership Program Phase II, during 1999-2001.
References
-
Dwivedi, R.S., A.B. Kumar and K.N. Tewari (1997): The utility of multi-sensor data for mapping eroded lands, International Journal of Remote Sensing, 18(11): 2303-2318.
-
Ehrlich, D., J.E. Estes, J. Scepan and K.C. McGwire (1994): Crop area monitoring within an advanced agricultural information system, Geocarto International, 4: 31-42.
-
Franklin, S.E. and C.F.Blodgett (1993): An example of satellite multisector data fusion, Computers and Geosciences, 19(4): 577-583.
-
Haack, B.N. and E.T. Slonecker (1994): Merged space-borne radar and Thematic Mapper digital data for locating villages in Sudan, Photogrammatric Engineering and Remote Sensing, 56(12): 1631-1641.
-
Jensen, J.R. (1994): Introductory Digital Image Processing, New Jersey: Prentice Hall.
-
Marino, L.A. (2001): Examining Image-Fusion Methods for Land Cover Mapping in the Thevaram Basin, Southern India, Master’s Dissertation, Faculty of Environmental Studies, University of Waterloo, Waterloo, Canada (unpublished).
-
Munechika, C.K., J.S. Warnick, C. Salvaggio and J.R. Schott (1993): Resolution enhancement of multispectral image data to improve classification accuracy, Photogrammatric Engineering and Remote Sensing, 59(1): 67-72.
-
PCI (1998): Using PCI Software, Volumes I and II, version 6.3, December 4, 1998.
-
Schowengerdt, R.A. (1997): Remote Sensing: Models and Methods for Image Processing, San Diego: Academic Press.
-
Solberg, A.H., A.K. Jain and T. Taxt (1994): Multisource classification of remotely sensed data: fusion of Landsat TM and SAR images, IEEE Transactions on Geoscience and Remote Sensing, 32(4): 768-778.
-
Sunar, F. and N. Musaoglu (1998): Merging multisectoral SPOT P and Landsat TM data: the effects and advantages, International Journal of Remote Sensing, 19(2): 219-224.