Can DEM Enhance The Digital Image Classification?

Can DEM Enhance The Digital Image Classification?

Apisit Eiumnoh and Rajendra P. Shrestha
STAR Program, Asian Institute of Technology
P.O. Box 4, Klong Luang, pathumthani 12120, Thailand
Tel: +66-2-5245588, 5245584; Fax: +66-2-5245597
Email :apisit@ait.ac.th

Abstract
Improving the image classification result has always been a concern while working with satellite data. Various techniques of image preprocessing, classification schemes and integration of ancillary data have shown positive results to improve the classification results. This study carried out to explore the use of Digital Elevation Model (DEM ) in Landsat -TM data classification for a watershed area of Thailand demonstrates the comparison of classification results from with or with out DEM. In our study, DEM helped to improve the classification accuracy from 58 to as high as 78 percent. The study concludes that besides band ratios and Principal Component Analysis, DEM is very helpful to improve the classification results, however there lies an important consideration that the selection of other input bands for classification be carefully considered.

Introduction
Satellite remote sensing has become a vital tool for the monitoring and management of natural resources and for environment monitoring. Improving classification accuracy of digital data has always been an important concern to extract the real world situation in the form of thematic maps. In machine classification, spectral value represented as digital number are grouped together to form certain classes to which a theme is assigned to produce a thematic map. One of the main problems when generating land cover maps from digital images is the confusion of spectral responses. The possibilities, that two or more different features having the same spectral behavior are eventually classified as the same class, not only creates the difficulties in extracting the valid information but also introduces the errors in the classification. Hence, the classification accuracy is influenced by the type of image and consequently the spatial and spectral resolutions, the spatial variability of the land cover types and the attributes to be determined among other factors (Campbell, 1987; congaltion, 1988).

Attempts have been made to improve the accuracy of image classification based on various approaches, such as use of multi-temporal imagery to individualize information classes (Conese and Maseli, 1991), piecewise linear classifier with simple post-processing (Kai-Yi and Mausel, 1993), GIS based methodology with ancillary information like soils, topography, bio-climates (Gastellu-Etchegorry et al, 1993), GIS rules with ancillary data on terrain mapping units, elevation data(Palacio-prieto and Luna-gonzalez, 1996) and so on.

Digital Elevation Model (DEM) can be created from either sterepairs derived from satellite data or aerial photographs or generated from digital terrain elevation data. DEM can be readily combined with image data (both optical and SAR) for a number of different purposes (Lillesand and Kiefer, 1987). Fore example, DEM was used to calibrate and geocode SAR data classification is very limited due to higher cost involved in preparing the DEM.

This study attempts to explore the usefulness of DEM in digital image classification combining it as a component band with various preprocessed band combinations of Landsat-TM image for the land cover classification.

The Study Area
The study area, Sakae krang river basin, is situated between 15o03' to 16o 05' N latitude and 99o 07' to 100o 04' E longitude in the central region of Thailand. The annual mean minimum and maximum temperature in the area range from 19.5o C to 33.6o C with an annual precipitation range of 950 to 1500 mm. The flat to gentle slope topography extends from east towards west which end up as a mountainous parts with more than 50 percent slope gradient forming the head watershed of the basin. The alluvial fan of the area is mostly under agriculture with the major corps, such as both irrigated and rained paddy, sugarcane, maize, orchards, plantation corps. Both mixed deciduous and dry Ditperocarp forest exists on the lower altitude limestone hills and dry evergreen forest in the mountainous area. The elevation of the area ranges from 20 to 1,641 m above mean sea level.

Materials and Methods

Materials
The primary data used in the study was the Landsat -TM data acquired for path 130/Row 49 on 21 February 1995. However, other maps, such as land use map of 1993 (1:250,000), topographic map (1:50,000) were also used for the study. Field based information gathered during the ground survey of the area was also used during training area selection and results verification. PC ARC/INFO was used to prepare the boundary of the study area and encoding elevation data to create the DEM where as image processing was done using ERDAS ver 7.5.

Image Preprocessing
Preprocessing, which is designed to remove any undesirable image characteristics produced by the sensor, refers to the initial processing of the raw data to calibrate the image radiometry, remove noise and correct geometric distortions (Schowengerdt, 1983).

The bulk format of TM data was acquired on two different CCTs which were downloaded as two different scenes. Radiometric calibration of the two different images was performed by comparing the Gray Level Histograms (GLH) of each image and sample pixel's spectral value for identified feature so that the same feature in two different images have the same spectral value.

Image noise is any unwanted signal or disturbances in an image. It can be grouped as random, isolated, stationary and non-stationary periodic noise. For the given image, few bad lines as an isolated noise were suppressed by identifying the horizontal bad lines during visual inspection and replacing them with the average value of adjacent two lines.

The geometric correction of the image was performed registering the image to 1:50,00 scale topo map sheets by selecting 45 Ground Control Points (GCPs). The Root Mean Square (RMS) error accepted was less than 1 pixel (30 m) at the first order and nearest neighborhood transformation. Two geometrically corrected scenes were then stitched together from which only the basin are was clipped with a vector layer.

Of the data reduction technique, band ratios, NDVI of NIR and R bands, and Principal Component Analysis were carried out to enhance the image as the quantity of information carried by satellite data is not necessarily same as the amount of data.

DEM Creation
To create the DEM some 35,000 elevation points were digitized along the contour from topo map using PC ARC/INFO. For the plain area, the contour interval digitized were 10 m interval where as it was 20 m interval for the hilly area. This vector based digital terrain elevation data wee rasterized using ERDAS software ERDAS software. After testing several DEM created, a smooth DEM of 30 m pixel size was created using the following algorithm.

e**(-0.5) x (5Q))**2 where, Q = calculated distance/search radius
Image classification
Both unsupervised and Supervised classification were Employed to classify the image With combination of various original bands, single ratio bands, NDVI, principal component bands And DEM to test the classification Results. A total of 12 combinations for unsupervised and one for supervised classification were used (Table 1). Due to lower spatial resolution of band 6 of TM data, except this band, all other bands were used in the classification process.

Table 1. Band Combinations used.

Band Combination (BC)	Bands	Classification Technique
1 2 3 4 5 6 7 8 9 10 11 12	X2.X3.X4 X1,X4,X5 X2, X4, NDVI X4,X7/ST, NDVI X2/X1, X7/X5,NDIV X2,X4, DEM X2/X1, X7/X5, DEM X2, NDVI, DEM X4, NDVI, DEM X2/X1,XY/X5,NDVI, DEM PCI,NDIV, DEM X7/X5,NDVI,PCI,DEM	Unsupervised
13	NDVI,PC1,DEM	Supervised
Note: X = Band number PC1 = Principal component band 1

ISODATA clustering of Unsupervised approach was used to Examine the various band combinations. The approach is relatively simple and has considerable intuitive appeal (Vanderzee and Ehrlich, 1995), however, the output of this technique could be affected by the choice of initial parameters and their interactions with each other (LAS, 1990). In this case, the parameters assigned to each band combinations were kept same including number of cluster which was 40. The clusters formed were regrouped with Ward's method which first calculates the means for each variable within each cluster. Then, for each case, it calculates the squared Euclidean distance to the cluster means. These distances are summed for all of the cases. At each step, the two clusters that merge are those that result in the smallest increase in the overall sum of the squared within-cluster distances.

A supervised classification was also run on only one band combination that gave the better result during unsupervised classification. A maximum likelihood classifier with nearest neighbor transformation was used. Results of each band combinations were evaluated creating error matrices and finally, two descriptive statistics, namely producer's and user's accuracy.

Results and Discussion

DEM creation
The interpolated DEM output from the elevation point data wee evaluated randomly selecting 100 sample points and comparing the pixel value with the topo map. The closeness of fit between the extrapolated and topo value wre around 90 percent. A 3-D view of DEM is presented in the form of wire mesh in Figure 1.

Figure 1: A wire-mesh representation of the area
Unsupervised Classification Results
Spectral values for the different land cover classes in different band combinations are presented in Figure 2. For BC1, only harvested agriculture area and mixed deciduous forest could be separated. A mixed class situation was observed between rainfed agriculture, mountainous area and irrigated paddy in the lowland plain were classified as same class due to similar spectral behavior. In BC2, there was no significant improvement in the result. The situation was similar to the previous one.

Figure 2. Spectral values for different land cover classes in different band combinations.
AG= Agriculture, H= harvested, BA=Burnt area, DF=Deciduous forest, DEF=Dry evergreen forest, MDF=Mixed deciduous forest, WB=Waterbody, IP=Irrigated paddy, DDF=Dry dipterocarp forest, OV=Other vegetation, F=Forest, LLV=Lowland vegetation, SCL(P)= Scrubland in the plain, SP(FH)=Sparse forest at foot hills, LO=Lowland, UP=Upland
In BC3, NDVI helped to separate waterbody and burnt paddy field from other land cover types, however there is still mixed class situation between agriculture, forest and other vegetation types BC4 and BS5 gave the similar. In both cases, ratio band could differentiate the waterbody and burnet area. Similarly, harvested agricultural area was also distinct, however other land cover classes were still in mixed class situation.

In BC6, DEM showed the capability of differentiating 3 distinct forest types but other land cover class wee still observed to be in the mixed class. In BC7, ratio bands with DEM helped to differentiate major land cover classes in the area, however, there was mixed class situation between lowland forest and other tree vegetation. For BC8 and BC9, similar results were observed. Besides 3 types of forest, other land cover classes like harvested agricultural area and scrubland in the plain were also separately grouped due to NDVI and DEM bands. BC10 comprising ratio bands, NDVI and DEM gave improved classification results, specifically to differentiate between irrigated paddy and dry evergreen forest which otherwise were found to be classified as the same class in all previous combinations.

In BC11, the inclusion of the first order band of principal component showed superior results to previous combinations by differentiating major land cover classes, however, there was little bit mixed class situation between irrigated paddy and other lowland vegetation. In BC12, the additional ratio band (X7/X5) did not add much towards improving the result compared to that of BC11.

Supervised Classification Results
The land use map of the study area as an output of the supervised classification is presented in Figure 3. The result was much superior to the unsupervised classification results as the mean spectral values of the training area selected for classification showed satisfactory trend of separation for the land cover classes identified (Figure 4). Altogether 13 class could be identified for the given image.

Figure 3: Land use map of the study area.

Figure 4: Spectral values of the training area selected.
Two descriptive statistics, namely producer's accuracy and user's accuracy were calculated to estimate the Overall Classification Accuracy (OCA) for each band combinations. For the unsupervised technique, OCA ranged from about 58 to 78 percent (Table 2). Among 12 band combinations, DEM included combinations showed better results most of the time. For supervised classification, an OCA of OCA of 82.3 percent was obtained.

Table 2. Overall classification accuracy for different band combinations.

Band Combination	Overall classification Accuracy (%)
1 2 3 4 5 6 7 8 9 10 11 12	58.1 60.7 60.2 64.2 65.3 67.1 66.3 70.4 71.9 72.4 77.5 76.1
13	82.3

Conclusions
Any image classification results is influenced by data itself, preprocessing and enhancement techniques and classification schemes being used. Supplementing with ancillary data during image classification has been reported to improve the classification results. This study also concludes that DEM as one type of ancillary data integrated in image classification can improve the classification result, however proper band combinations for specific purpose is always taken to be in consideration.

Reference

CAMPBELL, J.B., 1987. Introduction to Remote Sensing (New YorkL Guildford).
CONGALTION, R.G., 1988. Using spatial autocorrelation analysis to explore the errors in maps generated from remotely sensed data. Journal of Photogrammetric Engineering and Remote Sensing, 54, 587-592.
CONESE, C. AND MASELLI, F., 1991, Use of Multitemporal information to improve classification performance of TM scenes in complex terrain. ISPRS Journal of Photogrammetric Engineering and Remote Sensing, 4 187-197.
GASTELLU-ETCHEGORRY, J.P. ESTREGULL, C. AND MOUGIN, E., 1993, A GIS based methodology for small scale monitoring of tropical forests - a case study in Sumatra. International Journal of Remote Sensing, 14 2349-2368.
GUNERIUSSEN, T. AND JOHNSEN, H., 1996, DEM Corrected ERS-1 SAR data for snow monitoring. International Journal of Remote Sensing, 17 181-195.
KAI-YI HUANG AND MAUSEL, P., 1993, Spatial post-processing of spectrally classified video images by a piece wise linear classifer. International Journal of Remote Sensing, 14 2563-2573.
LAS, 1990, ISOCLASS User's Guide. Land Analysis System, Ver. 5, Vol. 3, U.S. Geological Survey, EROS Data Center, Sioux Falls, South Dakota (Unpublished manual)
LILLESAND, T.M. AND KIEFER, R.W., 1987, Remote Sensing, and Image Interpretation. John Wiley and Sons, Singapore, Second edition.
PALACIO-PRIETO, J.L. AND LUNA-GONZALEZ, L., 1996, Improving spectral results in a GIS context. International Journal of Remote Sensing, 17 2201-2209.
SCHOWENGERDT, R.A., 1983, Techniques for Image Processing and Classification in Remote Sensing. Academic Press, Inc., Florida.
VANDERZEE, D. AND EHRILICH, D., 1995, Sensitivity of ISODATA to changes in sampling procedures and processing parameters when applied to AVHRR time-series NDVI data. International Journal of Remote Sensing, 16, 673-686.