Test site 1. Sokolov - Change detection
Among anthropogenic environmental hazards, mining activities have a most serious immediate impact on their local surroundings. Formation of sinkholes, loss of biodiversity, dust, erosion, slope instability, and contamination of soil, groundwater and surface water by chemicals from mining processes are some of the harmful effects encountered in these areas.
Mining areas are some of the most rapidly changing areas on earth. The obvious changes are caused by extraction of soil and its deposition in waste dumps. This causes changes in terrain altitude and topsoil components, and mineralogical alterations caused by exposing deep compressed minerals to air, water and wind. In addition, indirect changes result from, for example, employment opportunities related to the mining industry (creating residential and social changes). Detecting and analyzing these various changes is of major importance for decision-makers, environmental-watchers and the local residents
Reflectance spectroscopy and imaging spectroscopy (IS), also known as hyperspectral remote sensing HSR or hyperspectral imaging, have been shown to provide good classification capabilities for minerals, soils, vegetation and various pollutants in different applications. Techniques for change detection from HSR data have been shown to be effective at detecting changes in a scene. Among the various change-detection methods, we chose to focus on linear chronochrome (LCC), linear covariance equalization (LCE), principal component analysis (PCA) and spectral angle mapper (SAM). The aim of this paper is to compare the results obtained by these algorithms and to validate them against ground truth information. The area selected for this study was the open-pit Sokolov mining area in the Czech Republic, using the airborne HyMap sensor. In the first section, we provide a brief summary of the change-detection algorithms. Then we describe the study area and the methods used for data acquisition and preprocessing (atmospheric corrections and geo-rectification). In the third section, the change-detection results and capabilities are presented.
There are basically several different methods for change detection:
- Change detection based on linear transformation and RX detector
- Chronochrome change detection2 (LCC)
- Change detection based on covariance equalization 3 (LCE)
- Change detection using principal component analysis (PCA)
- Spectral Angle Mapper5 (SAM)
Two HyMap images were acquired, in July 2009 and August 2010, over the Sokolov area. The 2009 flight campaign included 9 flight lines and the 2010 flight campaign 7 flight lines. The sensor configuration was set to 125 wavelengths across the 450–2500 nm spectral range. A ground team of 15 scientists and students collected the ground truth at the time of the overpasses. Atmospheric correction was performed using ATCOR4 code (Richter and Schlapfer). Geo-rectification of the HyMap strips was done using PARGE software, whereas mosaicking of the rectified images was performed by ENVI software.
The change-detection algorithms were implemented using Matlab software in the MathWorks environment. Both ENVI and ArcGIS software were used to archive the data to a precise geographic projection. Ground geo-tagging of photos, ASD measurements and metadata description were performed with GeoSciTag software from Novospec Ltd., which exported the data to Google Earth and Excel spreadsheets. For simplicity (to conserve run time and computer memory), we applied the algorithms over a representative subset image. The selected area was part of the active mining processes indicated as a major source for possible environmental impact within the Sokolov area.
Although most of the image geo-rectification is aligned, there are places with geo-rectification errors. These errors can be a result of flight fluctuations, different target angle view and shading of targets, which cannot be corrected. These errors must be taken into account when performing change detection between two images to ensure that the changes detected by the algorithm are spectrally and not spatially oriented. This problem can be dealt with in different ways, such as downscaling or averaging the image. In this study, we chose to average the image with a sliding window of P x P dimensions where P can be determined by the geo-rectification error. Since most of the image has one pixel error in geo-rectification, we used a 3 x 3 sliding window, which is one pixel error in each direction.
Coal mining area: changes and interpretation
As previously mentioned, an image subset was selected within the Sokolov active mining area shown in Figure 1. This figure provides the area along the basic strip (A), the georectified image (B) and the common area on a similar color composite. As noted from Figure 2C and D, significant changes occurred in the area during the studied year.
Figure 1: Subset of the Sokolov mining area as a test case. (A) 2009 and 2010 image strips. (B) Geo-rectification of the 2009 and 2010 strips; the overlapping area is demarcated by a red dotted line. (C) The coal-mining area in 2009 is demarcated with a cyan line. (D) The coal-mining area in 2010, where the mining progress is emphasized by comparing with the 2009 coal-mining area (cyan). The RGB images were composed from 634 nm, 543 nm and 466 nm, respectively.
Different detection capabilities
Figure 2A shows the subscene further selected to apply the change-detection algorithms. Changes in area ii were detected by SAM, PCA and LCE, but not by LCC (B,C,E and D, respectively). It is interesting to note that the applied change-detection algorithms show different performance with respect to the degree of spectral changes occurring in different areas, even as pertains to major changes. This can be seen in Figure 4 for areas i and ii. In Figure 3B (area ii in Figure 2), the same extension trend, with movement of water ponds to the west, is visible with SAM, LCC and LCE (see Figure 2 B,D,E) but not with PCA (Figure 2C).
In these examples, the SAM algorithm shows the best intensity separation between the easily seen spatial changes. In addition, these examples show how spectral information contributes to the detection of changes using the selected algorithm over “easily seen” cases.
Figure 2: (A) The subset to which the change-detection algorithms were applied. (B-E) The algorithms SAM, PCA, LCC and LCE, respectively. The ellipses in (B-E) outline the areas changed due to the progress of mining, as mentioned in Figure 1.
Figure 3: Examples of spectral changes over two areas. (A) and (B) correspond to ellipses i and ii in Figure 2.
Detection of changes which cannot be seen with the “naked eye”
As the best-performing change-detection algorithm in the previous stages was SAM, we selected it to examine further “hot spots” not visible to the naked eye using a simple color composite as shown in Figure 1. Figure 4 shows the SAM’s hotspots and spectral investigation of selected pixels within these spots. Significant spectral changes in selected pixels could be extracted by SAM, although no corresponding visible changes had been encountered. For example, changes from wet-clay or coal surface coverage in 2009 to water coverage in 2010 can be seen in Figure 4 A1 and A2. These materials are dark in the visible region but spectrally different when assessed by the entire sensor sensitivity. Moderate changes, such as in clay moisture (Figure 4 A3), and minor changes, such as in the amount of coal coverage (Figure 4 A4), appear as yellowish and blue areas. These examples show that a spectral-based algorithm can detect changes where a spatial-based algorithm cannot. Although these examples illustrate a very simple case, we believe that more spectral-based changes are occurring in the area and it is important that they be extracted.
Figure 4: SAM output map emphasizing the pixels in which spectral changes have occurred. Examples of such changes are shown: (A1) from clay to water (cyan rectangular), (A2) from lignite to water (yellow rectangular), (A3) different clay condition (magenta rectangular), (A4) minor change (gray rectangular). White arrows in (B) point to changes due to the progress of mining mentioned in Figure 1.