GISdevelopment.net ---> AARS ---> ACRS 1998 ---> Digital Image Processing

Application of Change Detection Algorithms for Mine Environment Monitoring

Joseph Cacdac
9th Floor, Valero Tower
Valero St., makati City, Philippines
Tel : (63-2) 892-8026 Fax: (63-2) 817-9491
E-Mail: dames@portalinc.com

Abstract
Various changes detection algorithms are applied to multitemporal SPOT X S imagery of the Dizon large open pit mine area in San Marcelino, Zambales to investigate which techniques are suitable for mine environment monitoring. Techniques studies include image differencing, image-to-image ratioing, and principal components analysis. Change features are identified and detection by principal components analysis of a merged multitemporal data set is the most effective for mine environment monitoring.

Introduction
Remote sensing is proving to be suitable and cost-effective tool for certain components of environmental studies, especially those which are relevant in terms of content and scale, such as land cover and land use, water quality, landforms, drainage networks, and surface disturbance. Monitoring mine environments will require two basic information: baseline and change information. Baseline information pertains to information which future observations of the area can be compared with. Some of the environmental impacts of mining projects would be best indicated in change information.

Objectives
The objective of the study is to use straightforward image processing algorithms to generate change information from multitemporal data sets. The various methods are then evaluated for efficacy and suitability.

Materials and Methods
The following imagery were used: SPOT XS acquired 20 May 1998 and SPOT XS acquired 17 Decempber 1991. The overlap of the Dizon images cover images covers an area 16km x 12km with scene center located approximately at latitude 15°00' N and longitude 120°17'. Standard image processing procedures are employed. No attempt is made to devise new algorithms. Ground truth sources include a 1987 vertical air photo of the dizon minestite, government 1:50,000 topographic maps, a 1:250,000 land cover map plotted by the Swedish Space Agency in 1988, published geologic reports, company reports and other secondary sources. For the most part of the study, the PC-based ESIPP image processing system developed in the University of New South Wales was used.

Description of Study Area
The dizon Mines study area is on the western side of the southern portion of Zambales range and is bounded by latitudes 14°55' to 15°05' N longitudes 120°12 to 120°22' (see Figure 1). It covers an approximate are of 19,200 ha. The minestite lies of the western flank and is about 20 km south of Mt. Pinatubo whose eruption in 1991 caused heavy damage on equipment and facilities. The study area is drained towards the west by the Sto. Tomas and Santa Fe River. The former is fed by the Marella River, which drains the SSW slope of Mt. Pinatubo, and by the Mapanuepe River, which drains the catchment where the minestite is located.

Dizon is one of only two Philippine porphyry depsits inferred to possess an ophiolitic basement (Sillitoe and Gappe, 1994). Prior to the Mr. Pinatubo eruption, dizon Mines ranked as the Philippines' third largest copper and second largest by product gold producer, mining and milling 19,000 metric tons ore per day. The Dizon pit is oval in shape, about 1,3000 m across with a maximum width of 950m.

Average rainfall is 3,939 mm, one of the highest in the Philippines. Precipitation is characterized by extreme seasonality. December to April are extremely dry months and May to November extremely wet, with over 95% of total rainfall occurring during this period. Annual mean temperature is 27.1°C, January being the coldest (25.8°C) and may the hottest (28.5°C)

Lowlands are planted to rice, mango and coconut, and secondary crops like corn and vegetable. In the uplands, the indigenous Aeta propulation engages in swidden (slash-and-buran) cultivation, raising mostly root crops, corn and upland rice.

Pre-processing
Geometric rectification. Aquadratic ( second-order) transformation was performed on ground control points (GCPs) from the may 1988 SPOT XS image, given a root-mean-square error of ±1.24 pixels .pixel size was maintained equal to SPOT XS's normal 20m x 20m. Resampling was done using nearest neighbor so as minimize radiometric degradation.

Radiometric Correction. Dark pixel correction was used because data necessary to quantify the effects of atmosphere and instrument error on recorded digital numbers (DNs) were not available. Dark pixel subtraction is effected on each band by subtracting from all(DNs) the minimum DN for each band, the latter assumed an area on the ground with zero or near reflectance.

Image Rotation. In both Dizon images, solar illumination was coming from the sourtheast direction, resulting in topographic inversion when viewing the images north upwards. To facilitate visual interpretation, the image was rotated 180 degrees. The rotated 1988 and 1991 Dizon images are shown in figure 2 and figure 3 respectively. When referring to location maps, it should be noted that the upper part of the rotated Dizon image is actually oriented southward and the left side oriented eastward.

mage Registration and Radiometric Calibration
The December 1991 SPOT XS image was co-registered with the geometrically corrected May 1988 SPOT XS image. However, the two images have almost the same solar azimuth and elevation, allowing for reasonable radiometric calibration in lieu of absolute calibration using meteorological and other physical data.

Significant Between-Data Events
Significant physical events occurred in the study between the dates of acquisition of the two SPOT XS imagery used in multitemporal analysis. Mt. Pinatubo erupted in June 1991,ejection an estimated 1 billion cubic meters of ash and phroclastic material that blanketed much of the surrounding area. Ash fall damaged mine facilities and equipment, significant enough lower mine production in 1991. Also, rainfall levels, after along drought, were returning to normal and triggering lahar's flow. Lahar's flow blocked Mapanuepe River, flooding the immediate downstream of Dizon.

Change Detection by visual Comparison
Many change features are discernible from a visual comparison of theDec. 1991 with the May 1988 images. The effects on the study area of the Mt. Pinatubo eruption. Lahar flows. Ash fall, and the flooded Mapanuepe River show very well. A cursory inspection of the mine area shows that pit has become deeper, although other mine features do not exhibit any conspicuous change from the may1988 scene.

Change Detection by Image Differencing
A direct approach to change detection is mage differencing-pixel of the earlier image are subtracted from those of a co-registered more recent one. Mathematically, the procedure can be expressed as : Dxilk=Xilk(t2)- Xilk(t1)+c where Xilk =DN fro band k,i and j are line and pixel number in the image, t1=first date , t2= second date and C=offset to produse positive DNs ( Singh,1989).

Band by band difference images for the Dec 1991 and May 1988 SPOT XS images were generated with one resulting single channel (NIR) difference image shown in figure 4. The more obvious changes are indcated as very light or very dark feature in the difference images are listed in table 1. Some features are unique in the Dec.1991 image while other are unique in the May 1988 image. Mid-grey pixel indicate relatively time-invariant features.

Image -to- Image Ratio
When ratioing co-registered multitemporal images band by band, we compute Rxilk=[Xilk(t1)]/[ Xilk(t2)] where Xilk (t2) is the pixel value of band k for pixel x at row i and column j at time t2 (Singh, 1989). For two absolutely calilutely calibarated multitemporal images, non-change pixels would have a ratio value of unit (1) (no scale factor applied). For non-calibrated image, the ratio for non-change pixel would deviate from unity.

Ratio of the Dec.1991 to the May 1988 Dizon data sets appear similar to the difference images but ratio images appear sharper and the degree of contrest between change and non-change features better indicates the degree of change.

Change Detection by Principal Analysis
The May 1988 and Dec. 1991 SPOT-HRV co-registered Dizon data were merged into one file and a standardized Principal Component Transformation (PCT) using eigenvector. loading derived from the correlation matrix was performed to produce six principal component (PCs). These are show as a single-channel images in figure 5 to 10. To aid in the analysis of information content , the eigenvectors loadings for each principal component are plotted and shown in figure 11. Values of eigenvectors and per qualitatively classified as either bright, medium or dark in each principal component and listed in Table3.

The eigenvectors show that PCI is a combination of all high positively loaded visible bands and subdued NIR. Thus features which are bright in the visible bands of both the May 1988 and Dec. 1991 images are highlighted. Because the corresponding bands have almost the same eigenvectors, PCI in effect does not offer much in discrimination of change features.

PC2 is a difference of the "average" albedos of the Dec.1991 and May 1988 images. Accounting for 27.3% of the data variance, PC2 would appear to be the main principal component of interest in terms of change detection. From the table of features, it can be noted that those listed as bright, medium and dark are very similar to those from image differencing, especially in the visible bands.

PC3 appears to be inverse of PC1, with visible bands of both the 1988 and 1991 images having low negative or negligible loading and NIR bands almost equal high positive loadings. However, unlike PC1, PC3 shows changes especially in terms of vegetation cover.

PC4,like PC2, is a difference image of corresponding bands, with positive high albedos changes showing bright together with negative vegetation changes. The plot of eigenloadings show that new high albedos features are emphasized against the old, with non-change features appearing as mid-grey. On the other hand, vegetation from the older image is emphasized against new vegetation. Again, non-change vegetation appears mid-grey.

PC5 and PC6 are also change images because each emphasizes, either negatively or positively, the visible bands of one date while the contribution of the other image is almost totally subdued. In PC5, the contribution of the Dec 1991 image is minimized (minimal eigenloadings) while the reverse is true for PC6. Together accounting for less than 0.5% of data variance, PC5 and PC6, however, prove interesting in terms of change features they highlight (see Table 3Table 11 describing which features are bright or dark in each PC). PC5 shows the tailings ponds as bright and the open pit and waste dump as dark, although these are supposed to be among the time-invariant features in the MAY 1988 AND Dec. 1991 images. Appropriate contrast stretching would reveal change features like the shallower areas at the east end of the ponds which appear slightly deeper in the Dec 1991 image due to the increase in water elevation at the ponds and dark portions of the pit and waste dump, indicating very local changes at the mine are. The discrimination in water depth is due to the very high visible green eigenloadings, green being the band (in SPOT XS) where water penetration is greatest. The very low May 1988 visible red eigenloadings on the other hand emphasize negatively the open pit and the water dump change features.

Like PC5, PC6 highlights very localized changes, albeit these are positive, the PC6 eigenloadings appearing like the reverse of those of PC5. Positive changes highlighted are those which have very high reflectance in the visible red band. Thus, additional waste dump area, active mining area at the pit bottom, and the new access road all appear bright. It is worthwhile to note that it is only in PC6, the lowest order PC whose data content is often dismissed as noise, that these features are emphasized so well against the mid-grey background.

It is also interesting to note that in viewing PC5 and PC6 there is a sense of "thickness" in the image, although topography is actually subdued. This three -dimensionality in both PC5 and PC6 appears to be caused by the difference in solar elevation and angle, and possibly slight co-registration errors, which together introduce a parallax to the image.

Conclusion
The task of change detection is greatly facilitated by radiometrically calibrated multitemporal imagery. Because absolute, or even relative, calibration may not be practical under many situations due to the lack of ancillary data, the next best thing to do is to choose imagery with comparable solar illumination conditions, especially solar azimuth.

The application of various change detection procedures on the Dizon study areas shows the decided advantage of principal components analysis of a merged multitemporal data set. The method was able to isolate in lower order PCs localized change features not readily apparent or obscured in the image differencing or image ratioing.

Another chief convenience in change detection by PCA is that it relies less on the human interpreter's input, either in terms of threshold values or of training area, in discriminating change features from the non-change background. Compared to the other methods, PCA may be the closest approach to automated change detection. However, in tropical regions of rugged terrain, land cover is particularly dynamic and it may be risky to rely on automated change detection techniques for applications. In such settings change detection applications would continue to rely on human input the process.



Channel Bright Medium Dark

91G-88G and 91R-88R Ashfall,lahar, additional waste dump, new access road vegetation, tailings pond, open pit, waste dump clouds, lahar lake, Mapanuepe river, overflow weir
91NIR-88NIR Cloud shadows, lahar, ashfall vegetation, tailings pond, open pit, waste dump cloud shadows, lahar lake, overflow weir, lahar-covered rice paddies

Table1. Qualitative description of brightness of major features in the multitemporal Dizon images. (Bold characters represent features unique to the Dec.1991 image while those in italics are unique in the May 1988 image.)



Channel PC1 PC2 PC3 PC4 PC5 PC6

88Green 0.4659 -0.4230 -0.2162 -0.2467 0.6743 0.2044
88Red 0.4883 -0.3552 -0.2571 -0.2629 -0.6768 -0.2051
88NIR 0.1372 -0.5597 0.5220 0.6270 -0.0441 -0.0166
91Green 0.4840 0.4203

-0.0339

0.2960 0.2294 -0.6691
91Red 0.4852 0.4182 -0.0288 0.2966 -0.1811 0.6842
91NIR 0.2365 0.1724

0.7827

-0.5489 -0.0034 -0.0037
%VAR 48.1954 27.3204 18.8974 5.2164 0.2399 0.1305

Table 2. Egenvector loadings and % data variance for the principal for the component transformation of the merged co-registered Dizon data set.



Channel Bright Medium Dark

PC1 Clouds, ashfall, open pit, waste dump, new access road, marella river bed lahar, remobilised ash, bare ground tailings pond, lahar lake, vegetation
PC2 ashfall, lahar, remobilised ash, additional waste dump, cloud shadow, new access road vegetation, tailings pond, open pit, waste dump lahar lake, overflow weir, clouds, Mapanuepe flood plain
PC3 Vegetation vegetation at Mt. Balitog area, lahar, ashfall, open pit, waste dump lahar lake, Mapanuepe flood plain, streams, tailings pond
PC4 planted rice paddies, vegetation at Mt. Balitog area, ashfall, lahar lake, remobilised ash, lahar, additional waste dump, tailings pond fringe vegetation, tailings pond, open pit, waste dump, Mapanuepe, Marella and sto. Tomas flood plains cloud shadows, overflow weir, streams, Magara ridgeline
PC5 tailings pond, additional waste dump, clouds, cloud shadow, planted rice paddies vegetation, bare ground,lahar, ashfall new access road, Mapanuepe flood plain, open pit, waste dump
PC6 ashfall, lahar, active mining area, additional waste dump, new tailings in tailings pond, new access road vegetation, open pit, waste dump tailings pond, lahar lake

Table 4 Brightness of some features qualitatively described in each principal component of the multitemporal dizon set.