4.1. Geographic Object-Based Image Analysis
For the integrated analysis of the satellite imagery on multiple scales, Geographic Object-Based Image Analysis (GEOBIA) approach implemented in eCognition 8 using Cognition Network Language (CNL) was selected. With the increasing availability of high- and very high-resolution satellite sensors [36
], the images thus obtained provide small-scale details about the Earth’s surface and reveal additional facets which demand adequate image analysis and interpretation techniques [20
]. Along with the increasing resolution towards less than 1 m, new application fields became increasingly important. With regard to this sub-meter resolution where even smaller objects consist of several pixels, the question of why analysis should still focus on statistical values of single pixels rather than on spatial concepts arose [38
]. In conjunction with the rising demand for GIS-ready information—for which raster data is not well-suited [41
]—this provided the basis for transferring the concept of object-based image analysis already used in computer vision or biomedical imaging to the remote sensing domain in order to complement the more traditional pixel-based methods [38
Classification starting from image objects which are generated by grouping homogenous pixels can utilize additional features next to the spectral information and have the advantage of incorporating spatial concepts such as pattern, location and neighbourhood. Different algorithms have been developed [42
], first and foremost the multi-resolution image segmentation implemented in the software eCognition [43
]. This bottom-up, region-growing approach aggregates surrounding pixels according to pre-defined criteria of homogeneity through a combination of different parameters, i.e.
, scale, color and shape. Subsequently, spectral, contextual, textural as well as hierarchical information can be used to classify the features of interest [44
]. Owing to its bridging effect through remote sensing and GIScience concepts, it was recently proposed to extend the term Object-based image analysis (OBIA) to “Geographic Object-Based Image Analysis” (GEOBIA) [39
]. A review of the use of OBIA in remote sensing is given by Blaschke [39
], who leaves the question open of whether or not OBIA is a paradigm. However, most recently Blaschke et al.
] concluded that GEOBIA is a new and evolving paradigm.
The illustrated workflow contains a further development of the multi-scale image analysis approach conducted in [20
]. Schoepfer and Kranz [20
] focused on the conceptual design of the approach. In their study, a basic land cover classification on the HR data was performed to assess the potential of the workflow without performing a detailed analysis of the VHR data. Within this preliminary study the potential of the overall concept was proven, thus in the present study, the whole workflow was applied based on [48
, identification of potential mining areas within the HR data and a detailed assessment of two mining sites by the analysis of VHR imagery. Conducted in the rule-based classifier environment of eCognition8, the innovative aspect of the approach is the design of the rule-sets.
The first step is to analyze HR imagery to ensure the coverage of large survey extents for the detection of potential mining areas. Subsequently, VHR imagery is analyzed to investigate the previously defined “Region Of Interest” (ROI) in high detail. The final step is to conduct multi-temporal monitoring of the detected areas of potential mining on VHR data. This procedure is shown in Figure 4
The purpose of identifying hot spots of mining activity is to guarantee effective coverage of a large survey extent as well as the use of spatial resolution where mining sites are still detectable. One way to pursue this is to focus on spectral information only at this stage of the processing chain. Focusing on these potential mining sites, a ROI is created with a minimum size of 10 km × 10 km. All ROIs are additionally approved by experts to decide whether an order for VHR imagery is to be placed. A more detailed description of the identification of hot spot areas is provided in Section 4.3.
The detailed analysis of potential mining areas on VHR imagery is conducted on the previously determined ROI. The analysis of VHR data enables the initially identified hot spot of mining activities to be described much more precisely and to give more detailed indications of what might be happening in and around the mine itself. A more comprehensive description of this analysis procedure is given in Section 4.4.
The multi-temporal monitoring approach presents the final stage of the workflow. Its objective is to reveal changes over time concerning the extent of a potential mining hot spot area. The analysis itself resembles that of the preceding detailed analysis, enhanced by a post classification comparison. With eCognition8 software, two separate maps can be used to conduct the classification and subsequent synchronize the results for comparison. With respect to the primary objective of the study, only areas showing increase were considered during change detection analysis. A detailed description of this issue is provided in Section 4.5.
4.3. Identification of Hot Spots of Potential Mining Activity in Southern Masisi
At the beginning of the HR analysis, the creation of objects by means of a multi-resolution segmentation is carried out [43
]. Since the potential mining areas consist only of few pixels within HR imagery, small objects are considered to be more suitable for the detection of potential mining sites. Thus the segmentation scale is adjusted to create small objects. This also explains the reason why coarser, medium-resolution satellite imagery is not suitable for the purpose of this study. The detected areas of interest (AOI) are of fairly small size and therefore we assume that less than 10 m of resolution is about the limit for detecting these small objects. The subsequently elaborated classes are used to create areas to be excluded from further analyses, such as dense forest, clouds or water bodies (cf. Figure 5
Clouds are classified using an interface of spectral values, aiming at the high values of brightness inherent to clouds. Since vegetation and potential mining areas are mutually exclusive classes, a threshold of the Normalized Difference Vegetation Index (NDVI) is applied to classify vegetated areas. The threshold chosen is as close as possible to the values of the potential mining sites in order to maximize the exclusion effect and is assessed by an iterative procedure using expert knowledge. The classification of water bodies is also a key issue because proximity to water can be a significant indicator for the existence of potential mining sites.
After clouds, vegetation and water bodies are classified, the exclusive areas are elaborated. The classification of potential mining sites—in the following called AOI—is then performed by an interface of firstly, a fairly low value of the NDVI, since there is hardly any vegetation, and secondly a high brightness value owing to their high albedo in all spectral bands typical for bare soil. However, because of some residual uncertainty, a further measure of refinement is conducted. By means of neighboring, distancing and geometric elements, the AOIs are further refined. Subsequent to this fully automatic detection, a more precise visual inspection of the resulting potential mining sites eliminates those sites that are bare soil, but too remote from settlements or at least dwellings, water bodies/rivers and roads, which all can be seen as necessary infrastructure to be expected in the vicinity of a mining area. A more specific analysis of the remaining potential mining sites is conducted on the VHR satellite images.
Building upon the AOIs, the exclusive areas become dispensable. The image is again tesselated using a rectangular segmentation algorithm where the tile size depends directly on the pixel size of the underlying satellite image. This ensures a typical tile size of approximately one kilometer, which is of importance to grow a ROI around the areas of interest classified.
The complete growing process is visualized in a series of stages in Figure 6
. Starting with the deletion of the exclusive classes in (1), a rectangular segmentation with a border length of approximately one kilometer is created. Those cells containing one or more AOIs (2) are defined as the starting point for the subsequent growing process (3). Region growing is performed in all directions resulting in a rhombus. In order to restore a rectangular ROI, all unclassified tile sizes which share at least 50% of their border with a tile size classified as ROI are also classified as such. Part (4) depicts the result: A rectangular ROI, approximately 10 km × 10 km in extent, including the areas of interest precisely in the centre. These squared ROIs constitute the basis of a request for VHR imagery, for which the minimal size to be ordered is 100 sq km. In terms of a service workflow, automated region deliniation is highly valuable.
It is important to note that the ROIs generated can still contain misclassified AOIs. Thus a visual validation of the ROIs by an expert is indispensable before ordering VHR data. Only if the result of this validation is positive should the next stage of the workflow be initiated.
4.4. Detailed Analysis of Mining Sites
The detailed analysis of mining sites was conducted for both sites, for (1) Mumba-Bibatama since it represents the ROI detected from the HR analysis, as well as for (2) Bisie, which acts as verification site within this study. The results of the analyses are presented in Section 5. In Section 4.4, technical and methodological aspects are explained in detail (cf. Figure 7
“Streets”, “rivers” and “settlements” are the three classes available from thematic layers mentioned in the Section 3. These layers are imported to an object-based classification rule set using the Cognition Network Language (CNL) [49
]. According to Tiede et al.
, CNL “enables addressing single objects and supports manipulating and supervising the process of building scaled objects in a region-specific manner” [49
]. During the following classification procedure, the thematic layers are used as indications and additional refinements [43
Initial classes with exclusive character—water and vegetation—are set up on the basis of spectral values. An automatically self-adjusting set of thresholds within the rule set ensures a highly objective classification of the aforementioned classes. Furthermore, a temporary class called “high reflecting areas” is derived. The latter contains potential mining areas as well as clouds and settlements since all three of them have significantly high reflectance values but cannot be subdivided on VHR data by spectral properties only. Thus the class “high reflecting areas” must be subdivided according to other properties. One option is to take textural properties of the objects into account.
eCognition8 provides several approaches for considering textural properties, whereby a well-established method is the Gray-Level Co-occurrence Matrix (GLCM) [50
]. This formula describes the neighboring gray-level values by means of a matrix of neighboring gray-level occurrences. For the purpose of subdividing the “high reflecting areas”, two different kinds of these GLCM features are applied in this study. GLCM-Entropy, which is a measure for the chaotic gray value distribution of neighboring pixels within one object, and GLCM-Mean, which focuses on a smooth and regular gray value distribution of neighboring pixels within one object.
In the course of the analysis, the GLCM-Entropy feature proved to be a suitable feature for the purpose of subdividing “high reflecting areas” as well as distinguishing between potential mining sites and adjacent vegetation areas. Vegetation is represented by a smooth and regular distribution of gray value, whereas a chaotic distribution is inherent to bare soil areas. The maximum value of the GLCM-Entropy feature is detected automatically. With expert knowledge, an alterable percentage of this maximum is calculated by an arithmetic expression. The exact formula is presented in Table 2
. The term “var” represents the adjustable lower value limit, whilst “x” stands for the value range between the maximum and “var”. Since the threshold between potential mining sites and sparsely vegetated areas is subject to variation and depends on the specifications of each satellite image, the percentage range needs to be adjustable. Once the value range has been determined, the “high reflecting areas” applying this range are classified as potential mining areas. Subsequent to the classification, the pixel-based growing algorithm is applied to ensure a highly accurate acquisition of the potential mining areas.
The penultimate classification step encompasses a separation of the remaining “high reflecting areas” class into the subordinate classes of “clouds” and “settlement areas”. This distinction is made by use of the GLCM-Mean feature according to [50
]. It takes into account the regular and smooth distribution patterns of neighboring gray-level values within the objects. This feature is important for cloud signatures, which are subject to highly regular patterns as opposed to settlement areas which are composed of different sizes of buildings and paths.
The calculation of the second alterable percentage range is also presented in Table 2
and should be carried out with expert knowledge. All objects in the superordinate class which apply the range between the maximum and the determined lower value limit are classified as clouds.
The last step in the scheme presented is to classify the remaining “high reflecting areas” objects into the class “settlement area”, since only objects that represent settlement areas remain in the superordinate class. At this point, the thematic layer “settlement” is taken into account. This makes it possible to reclassify distinct objects, firstly with regard to their directly neighboring objects such as rivers and streets, and secondly with regard to a distancing function, which is useful in the case of available settlement points. The refinements are to be carried out after visual interpretation of the classification result to avoid misclassifications.
With regard to the arithmetic expressions used during the detection of potential mining areas, a differentiation between arithmetic expressions for the vegetation classification and the textural calculations is obligatory. The former is a static expression which adopts the values of a given imagery and thus defines the thresholds for the classification in a fully automated manner. The latter, used for the classification of potential mining areas and clouds are expressions that require adjustments of the thresholds on the part of an experienced user.
The rule set described above applies to all four satellite imageries, i.e., Bisie (3 images) and Mumba-Bibatama. The procedure was carried out and validated repeatedly. After this mono-temporal analysis, a multi-temporal approach involving two distinct points of time can be acrried out. This enables the user to detect changes over time. The corresponding procedure is described in Section 4.5.
4.5. Multi-Temporal Monitoring
The rationale for the multi-temporal analysis carried out for the Bisie mining site was due to two considerations: Firstly, the site is well-known among experts and local reports as well as statistical numbers are available. Thus it is possible to compare and to verify the findings of this study using on-site information. Secondly, there were three satellite images available, which enables change detection over different points in time.
Several preconditions have to be considered in the case of a change detection analysis. First of all, the result of change detection is strongly influenced by environmental factors such as soil moisture or plant phenology. Secondly, different techniques to detect a change are available [51
]. For the purpose of this study, the post-classification comparison according to [36
] was chosen, thus the results of the above-mentioned mining site detection using different points of time were compared with each other. It should be mentioned that change detection is conducted fully automatically and not refined manually. Thus the changes revealed have to be interpreted carefully.
Change detection was only implemented for the three images of the Bisie mining site. The purpose was to assess whether the quality and extend of changes over time can be detected. For the most effective processing, a subset of the satellite images covering the mining area and its surroundings was created before applying the analyses. In this manner, the scenes to be compared are imported into two different maps, enabling the program to classify each scene separately. The analysis procedure for each map is the same but with adjusted values of each variable for the respective scene. Subsequent to the classification process, the post-classification comparison of both scenes is conducted on a third map. This approach ensures that it will be possible to return to a particular classification and to implement alterations at any time. On the change detection map, the class of potential mining sites from both input classifications is compared directly, and the change is detected through a simple scheme: Where both classes overlap, no change is detected. Where the former potential mining class and no later potential mining class is present, a negative change can be identified, i.e., a decrease in extent. In contrast, where the later potential mining class and no former potential mining class are apparent, a positive change has occurred, representing an increase in extent.
4.6. Accuracy Assessment
An evaluation of the quality of the classification of satellite imagery is elaborated on the basis of accuracy assessments. With regard to this study, a sample-based accuracy assessment according to [55
] for each classification was produced within eCognition8 software. The software offers the possibility to conduct a class-specific accuracy assessments, i.e.
, the samples collected are compared to the target class only which is in the present study the class “potential mining areas”. All other classes are considered exclusive classes and thus are not of major interest with regard to their accuracy. In the first place, the target classes are selected. This does not entail the fact that the samples have to be located inside these classes. The samples are selected in a second step, where the classification is set to invisible. In a third step, the selected class and the samples are compared and a number of accuracy is provided.
A distinction has to be drawn concerning the resolution of the imageries, which influences the number of samples collected since, with inferior resolution, the number of pixels representing the potential mining areas is smaller. With regard to the highest resolution of the GeoEye-1 data, 100 samples were collected to conduct the accuracy assessment. In terms of the resolution of the IKONOS imagery, 50 samples were collected. The RapidEye image of Southern Masisi provides a special case, since the object sizes cover only few pixels of ground data because of the coarser resolution of 6.5 m. Thus, the class of purpose is covered by only a few objects in the imagery. This explains why merely 27 samples were collected to conduct the accuracy assessment. Since samples cannot be selected randomly in eCognition8, an independent user collected the samples while the classification of the respective imagery was set to invisible. The total numbers of samples as well as the accuracy values are shown in Table 3