04-136
6/9/06
9:56 AM
Page 835
Automated Techniques for Environmental Monitoring and Change Analyses for Ultra High Resolution Remote Sensing Data Manfred Ehlers, Monika Gaehler, and Ronald Janowsky
Abstract For monitoring environmental changes, new digital remote sensors have become available that allow monitoring and change detection analyses at resolutions and scales that were deemed impossible just a few years ago. The advent of airborne stereo scanners of ultra high spatial resolution offers the possibility of a complete digital remote sensing processing system. Current sensors include the High-resolution Stereo Camera (HRSC), the ADS-40, and the Digital Mapping Camera (DMC). For automated analysis, however, the new sensors require also new processing techniques. This paper presents results of change monitoring analyses for areas along the shorelines of the Elbe and Weser rivers in North Germany using integrated HRSC and GIS datasets. An automated procedure for highly accurate mapping was developed which is based on a hierarchical stepwise approach integrating GIS methods and digital surface information in this process. This approach allows the production of GIS maps that are more detailed and accurate than those that were previously produced by conventional means. Within the GIS environment, the multitemporal analysis also allows the exact quantification and location of changes of the protected biotope types.
deemed impossible just a few years ago. These sensors include the new very high-resolution remote sensing satellite programs such as Ikonos, Quickbird, or EROS (Ehlers, 2002a). In addition, the advent of stereo scanners of ultra high spatial resolution offers the possibility of a complete digital airborne remote sensing processing system: data acquisition, data preprocessing, data analysis, and data integration into a GIS can be managed in an integrated digital environment. These sensors deliver not only ultra high spatial resolution of about 5 to 30 cm but also 3D digital surface models (DSM). Accurate digital surface models (DSM) can also be obtained by direct recordings using Laser scanning systems (Lidar). The excellent geometric qualities of the new digital sensors allow easy integration with GIS and digital maps. The ultra high-resolution with its associated high in-class variances, however, have so far presented a severe impediment to a fully automated classification process often resulting in on-screen digitization as final option. Some progress has been made by employing segment-based and hierarchical image processing techniques. In this paper, we will present a method that is based on a hierarchical classification scheme and integration with GIS and height information.
Introduction
Digital Camera Systems
The number of remote sensing programs and systems has increased dramatically over the last several years serving the needs of geographic information systems (GIS) with high demands for geospatial data. New technologies such as coupled global positioning systems (GPS) and inertial navigation systems (INS) allow airborne sensors to produce digital data of excellent geometric accuracy and challenge standard large format aerial cameras. Multi-source remote sensing systems are creating data at higher spatial and temporal resolution than have been collected at any other time on Earth. GIS technology allows the efficient storage and management of spatial datasets in digital formats. Remote sensing systems, in turn, acquire current, accurate and synoptic data that can be used to update GIS databases. In combination with the appropriate data transfer and interoperability standards that are currently being developed, the technology is being put in place that will eventually allow standardized exchange, processing and dissemination of geospatial information. New digital remote sensors have become available that allow applications at resolutions and scales that were
After a long period of development, we now see the emergence of operational digital camera systems which challenge aerial frame cameras. Advanced technologies such as GPS coupled navigation systems and advanced digital sensor technologies have overcome the strongest impediment of aircraft scanners: the lack of geometric stability. Public and private research has concentrated on the development of digital line array or matrix scanners that will serve as successors to the classical airborne camera systems. Companies such as Leica Geosystems, Inc., Z/I, Imaging, or Vexcel offer commercial systems, research centers such as the German Space Center (DLR) operate their own prototypes. Such systems have to establish their market somewhere between the satellite image user seeking higher resolution and the air photo user seeking digital input and GIS compatibility. Consequently, airborne scanner systems have to offer stereo capability and multispectral sensing. Two different technologies are being employed to accomplish an airborne digital recording system. Z/I Imaging and Vexcel make use of two-dimensional arrays and a set of
Institute for Geoinfomatics and Remote Sensing (IGF), University of Osnabrueck, P.O. Box D-49069, Osnabrueck, Germany (
[email protected]; mgaehler@igf. uni-osnabrueck.de;
[email protected]). PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Photogrammetric Engineering & Remote Sensing Vol. 72, No. 7, July 2006, pp. 835–844. 0099-1112/06/7207–0835/$3.00/0 © 2006 American Society for Photogrammetry and Remote Sensing July 2006
835
04-136
6/9/06
9:56 AM
Page 836
coupled nadir-looking lenses to emulate a standard frame camera’s central perspective (Ferrano and Felix, 2003; Leberl and Gruber, 2003). Leica Geosystems, Inc. and the DLR employ triplet scanner technology with one-dimensional line arrays arranged in fore, nadir and aft looking modes (Sandau et al., 2000; Hoffmann and Lehmann, 2000). The advantage of a 2D matrix camera is that all standard photogrammetric techniques can be used in a digital environment. The advantage of a stereo triplet solution is that photogrammetric preprocessing (i.e., DSM and orthoimage generation) is performed before the user receives the data, thus alleviating the need to run sophisticated software at the users’ organization. The image data are provided in the required coordinate system and can be easily integrated into an existing GIS database. Which of the two approaches is more suitable will largely depend on the user demands and the price-performance ratios of the respective systems. Table 1 summarizes the characteristics of five selected ultra high-resolution airborne digital camera systems. The advantages of digital cameras are widely understood and include: no film, no photo processing, no scanning, better radiometric quality through direct sensing, non-aging storage and direct integration into GIS and image processing systems. The disadvantages of digital scanners, most notable, geometric distortions and monoscopic imaging mode, no longer exist due to the stereo capabilities of the new sensors and the use of integrated INS and differential GPS technology during image acquisition (Ehlers, 2004). For the present studies, the HRSC camera systems were applied. The HRSC-A (High Resolution Stereo Camera – Airborne) is a line array scanner with nine CCD line detectors mounted in parallel at the focal plane behind one single optics and perpendicular to the flight direction. Nine images of the same strip are simultaneously recorded as a result of the forward motion of the aircraft. Five of the nine CCD lines of the HRSC-A are panchromatic sensors and arranged at specific viewing angles (18.9° and 12.8° and nadir) to provide stereo viewing capability. Four of the nine CCD arrays are covered with different filters for the acquisition of multispectral images (see Table 1); they are also arranged at different viewing angles (3.3° and 15.9°). A fully automatic photogrammetric and cartographic processing system was developed at the German Space Center DLR, in cooperation with the Technical University of Berlin (Wewel and Scholten, 1999). Although this system can work without ground control points (GCPS), a few GCPS are usually required for validation purposes. For the generation of digital surface models (DSMS), a digital multiple correlation process is performed. The resulting identical image points are converted into object points by ray intersections. The DSM is derived by interpolation. Finally, based on the DSM, digital orthoimages are generated and multispectral orthomosaics are produced. As a spin-off, an extended narrow-angle system (HRSC-AX) has been developed by the DLR providing higher resolution and re-designed viewing angles (panchromatic: 20.5° and 12° and nadir; multispectral: 4.6° and 2.3°) and multispectral bands that are better suited for terrestrial applications (see Table 1; Neukum et al., 2001). This sensor has been operationally used by ISTAR (Sophia Antipolis, France) throughout Europe.
Problems for Automated Processing of High-resolution Data Of importance, however, is that the new sensors pose new challenges for automated interpretation. The homogenizing effects of comparably large pixel sizes are no longer valid. With these ultra/very high-resolution sensors, simple pixelbased analyses are no longer applicable because of the 836
July 2006
difficulty of classifying high-resolution data where each pixel is related not to the character of an object or an area as a whole, but to components of itself (Blaschke and Strobl, 2001). Instead, we have image pixels that might belong to the same class but exhibit totally different reflection values (high in-class variances). For example, a high-resolution image of the class house can be represented by hundreds of pixels that might belong to undesired subclasses window, chimney, sunlit roof, shadowed roof, front lawn, or driveway (Figure 1). It is well known that spectral classification of higher resolution data does not automatically lead to more detailed classification results (see, for example, Metternicht, 1999; Petit and Lambin, 2001). Further, using just multispectral information for classifications does not lead to accurate interpretation results because the differentiation between object classes is done not only with the help of spectral information but also with spatial (contextual) information of the image data. For example, using only multispectral information different objects like roofs and streets might not be separated in two object-classes because they are built with the same material (Hoffmann et al., 2000). Consequently, new intelligent techniques will have to be developed that make use of GIS integration, multisensor approaches and context based interpretation schemes (see, for example, Ehlers, 2000; Schiewe, 2003). Otherwise, the last step of an all-digital image acquisition and handling process has to consist of manual on-screen digitizing. It is therefore imperative that new techniques be developed that allow for automated processing of high-resolution and multisensor images. One promising approach is the use of auxiliary information in the processing steps. While the use of additional information is not new in the processing of remotely sensed images, the existence of geographic information systems (GIS) presents a formalized structure within which the additional information can be provided and preprocessed.
Case Study: Biotope Type Monitoring The rationale for our study was several fairway deepening projects on rivers in northwestern Germany due to shipping requirements. Continuous environmental monitoring has been mandated at the conclusion of expansion projects with focus on mapping the tidally influenced riverside biotope types. Changes in composition and size of these biotopes should be documented over the long term to assess the impacts of hydraulic engineering measures. The term “biotope type” is being used as a more specific one than the term “land-cover class”. In contrast to “landcover class,” “biotope type” not only involves similar structures of vegetation or other surfaces but also includes abiotic parameters such as altitude or relative distance to water. The German Waterways and Shipping Administration was responsible for promoting the development of a generally applicable automated and cost-effective method for an operational, long term monitoring process. As stated above, simple pixel-based analyses are not applicable because of the difficulty of classifying high-resolution data where each pixel is related not only to the character of an object or an area as a whole but also to components of the object. In cooperation with the Institute for Geoinformatics and Remote Sensing (IGF) and the German Aerospace Center (DLR) in Berlin, an integrative monitoring concept was developed using a combination of GIS, image analysis and modeling software (Ehlers et al., 2003a and 2003b). Therefore, our objective was to determine, if an essential improvement could be obtained by using appropriately recorded digital data and applying an automated analysis process PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING ET AL.,
Own platform
Zeiss T-AS platform
Estim. Costs incl. Pos. system
Georeferencing
Data Recording
0.25 images/s 1:1,000
—
High speed recorded Applanix POS/DG with GPS and INS
CIR Modus
425,000 USD
80 GB exchangeable hard disk Applanix POS IMU with GPS and INS
510–600 (green) 600–720 (nor/NIR) 720–920 (NIR)
1640 lines/s 1:500
400–500 (blue) 500–600 (green) 600–680 (red)
Readout Frequency Largest Application Scale Stabilisation
RGB Modus
450–510 530–576 642–682 770–814
Resolution (nm)
(blue) (green) (red) (NIR)
520–760 (pan)
Spectral
9 m 12 bit
4092
—
Area CCD 2004 55 mm (color & CIR) 35 mm (only color) 37° 55.4° 1 4077
6.5 m 12 bit
2001; HOFFMANN
Applanix (Emerge) www.emergedss.com
DSS
2003C; HINZ
12172
29° 9
Line CCD 2000 151 mm
Sensor Type Introduction Focal Length
Field-of-view # CCD- Lines/Matrix Camera # CCDs across track # CCDs along track Sensor Size Radiometric Resolution
DLR www.dlr.de
Company URL
ET AL.,
64° 7
Leica Geosystems www.gis. leicageosystems. com/ Line CCD 2000 62.7 mm
ADS 40
LEHMANN, 2000; GRUBER
1,200,000 USD
Applanix POS IMU with GPS and INS
MM40 mass storage
LH platform
428–492 (blue) 533–587 (green) 608–662 (red) 703–757 (NIR) or 833–887 (NIR opt.) 800 lines/s 1:500
465–680 (pan)
6.5 m 12 bit
2 12 000 (pan) 12 000 (ms) —
AND
UltraCam-D
(blue) (green) (red) (NIR)
700,000 USD
Not specified
SCU (1 TB)
Not specified
0.75 images/s 1:150
390–470 420–580 620–690 690–900
390–690 (pan)
11 500 (pan) 4008 (ms.) 7 500 (pan) 2 672 (ms) 9 m 12 bit
Area CCD 2003 100 mm (28 mm multispectral) 55° 37° 9
2000) DMC
ET AL.,
824 000 680 000 12 12
(pan) (msl) (pan) (ms) m bit
(blue) (green) (red) (NIR)
POS Z/I 510 navigation system with GPS and INS 1,600,000 USD
Zeiss T-AS platform RAID disk system
0.5 images/s 1:150
400–580 500–650 590–675 675–850
400–580 (pan)
13 3 7 2
Area CCD 2002 120 mm (25 mm multispectral) 74° 44° 8
Z/I Imaging www.ziimaging.com
2003; MÖLLER, 2003; SANDAU
Vexcel Corp. www.vexcel.com
ET AL.,
9:56 AM
HRSC-AX
DIGITAL AIRBORNE CAMERA SYSTEMS (AFTER EHLERS
6/9/06
Sensor
TABLE 1.
04-136 Page 837
July 2006
837
04-136
6/9/06
9:56 AM
Page 838
Figure 1. High-resolution image of a house with several identifiable subclasses of different spectral reflectance (high in-class variance).
employing a hierarchical and rule-based procedure. We hypothesized that the selected approach would guarantee the best possible accuracy compared to traditional mapping and surveying methods that are always combined with extensive fieldwork. Study Sites The study sites are located along the tidally influenced areas of the rivers Elbe and Weser in northwestern Germany near the large cities of Hamburg and Bremen (see Figure 2). In both river areas, reeds and some relics of willow forests are of major interest for nature conservation and long-term monitoring. Datasets For our studies, complete HRSC datasets (i.e., panchromatic and multispectral images, and DSM data) were used. An example of various spectral band combinations is presented in Plate 1. In several projects, different HRSC sensors/datasets were applied (see Table 2). The data were recorded from flying heights between 3,000 m and 6,000 m and delivered with 15 cm to 60 cm ground pixel resolution. Absolute accuracies were given as 20 to 30 cm in horizontal and 30 to 50 cm in vertical direction, respectively. Methodology and Results The main objective of the projects was the development of automated and reproducible methods to extract useful information from remotely sensed ultra high-resolution
Figure 2. Study sites in Germany.
838
July 2006
digital data. The IGF research group developed an indexbased segmentation and pre-classification procedure as pragmatic data preparation approach for an automated hierarchical classification process. The analysis is implemented in several steps (see Figure 3). The first Level (1) is the computation of ancillary information such as texture and vegetation indices. For analyzing remotely sensed data, textural measures are used for increasing information content. In our studies the use of a variance filter (7 7 matrix) based on the panchromatic channel proved to be the most useful for a better differentiation of biotope types. Biotope types with relative smooth surfaces (e.g., water and grassland) can be distinguished from biotope types with relatively rough texture features such as tall reeds and trees. The computation of vegetation indices is a standard procedure for satellite remote sensing applications. Indices like the Normalized Difference Vegetation Index (NDVI) and its derivatives are commonly used for separating vegetation from bare soil as well as for estimating quality and vitality of vegetation stands (Jensen, 2004). In the field of airborne remote sensing, those computations have been used rarely, mostly due to the limited availability of an appropriate database. Scanned aerial photos have generally proven to be inappropriate for these methods due to their heterogeneous radiometric properties and their evident inconsistencies even within single scenes. Even the first models of operational digital scanners were limited for airborne applications computations of vegetation indices. For example, the HRSC-A offered an inappropriate spectral resolution (i.e., no true red band) for vegetation mapping tasks. The new HRSC-AX sensor includes bettersuited multispectral bands for terrestrial applications with higher radiometric resolution (see Table 1). Consequently the excellent quality of the HRSC-AX data enables the computation of vegetation indices. However, the applicability of different vegetation indices has to be examined thoroughly because spectral bands and scales of the airborne system differ from those of standard satellite systems. During our research, numerous indices have been tested for the HRSC image data. For the HRSC-A, best results were obtained with a combination of the near infrared, the panchromatic and a calculated virtual red band (Ehlers et al., 2003a). For the HRSC-AX, standard NDVI methods are sufficient. Of note, however, the use of vegetation indices combined with a hierarchical classification procedure improves the classification process with respect to speed and accuracy (Ehlers et al., 2003b). The NDVI vegetation index is used in the next step (Level 2) to identify and separate two coarse classes (nonvegetation/sparse vegetation/water/shadow and vegetation) (see Figure 3). With the additional incorporation of height information (ground level elevation plus vegetation height) from the DSM, the vegetation class is further divided into high vegetation (e.g., trees) and low vegetation (e.g., shrubs, grass) by using an appropriate threshold. This can be achieved by incorporating all data (original multispectral dataset, DSM, NDVI, and texture) in an integrated GIS/image processing environment using GIS and/or image analysis procedures. Thus, biotope types that do not show a difference in their multispectral reflectance characteristics but are of different height can be separated easily. This separation of information step, therefore, permits the detail and accuracy of the classification to be improved. The resulting segments are already pre-classified into the semantic layers non-vegetation/sparse vegetation/water/shadow, low vegetation, and high vegetation (for example, see Figure 4). Within this process for each segment and each semantic object class, a minimum size is defined. Smaller segments are PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
04-136
6/9/06
9:56 AM
Page 839
Plate 1. Various views recorded with only one HRSC-AX camera system: digital surface model (DSM) with orange high elevation, yellow and green medium elevation, and blue low elevation; color infrared (CIR); panchromatic; true color (RGB).
TABLE 2.
Sensor Dataset provided by Flight Height Ground Pixel Resolution Radiometric Resolution
SENSOR
AND
DATASETS USED
IN THE
STUDIES
Pilot Project (River Elbe 1999)
River Elbe 2000 a (Area North of Hamburg)
River Elbe 2000 b (Area South of Hamburg)
HRSC-A DLR
HRSC-A DLR
HRSC-AX DLR
3000 m 15 cm 50 cm (DSM) 8 bit
6000 m 30 cm 50 cm (DSM) 8 bit
6000 m 60 cm 100 cm (DSM) 12 bit
eliminated using majority filtering. Level 1 and 2 are conducted completely without human intervention. In Level 3, the separated layers were treated with appropriate classification algorithms (i.e., isodata clustering for the non-vegetation/sparse vegetation/water/shadow layer and supervised classification for the herbaceous vegetation layers) (see Figure 3). With this approach, the level of detail in the biotope type classification could be significantly improved. Finally, a GIS-based post-processing (Level 4) is involved to produce the final classification result. GIS operations such as majority filtering, logical overlay, definition of neighborhood relations and minimum area functions were used to compute the correct classes for the shadow areas and to combine the individual information layers: The individually classified layers were combined in a hierarchical process considering a predefined priority order (see Table 3) to obtain one final resulting layer. (For example, biotope types of herbaceous vegetation had a higher priority than types of non-vegetation.) Within this process, a minimum size filter for each biotope type was used (Ehlers et al., 2003a). PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
River Elbe 2002
River Weser 2002
HRSC-AX TerraImaging/ ISTAR 6000 m 25 cm 100 cm (DSM) 12 bit
HRSC-AX TerraImaging/ ISTAR 6000 m 32 cm 100 cm (DSM) 12 bit
The shadow areas were eliminated by GIS-based filter techniques, such as majority and altitude-based filtering. Shadow pixels/regions that were higher than 12 m (derived from the DSM) were filled up only with biotope types of tree vegetation in the direct neighborhood and not with biotope types of non-vegetation or low vegetation and vice versa (see Figure 5a through 5c). In addition, neighborhood relations between the different biotope types were considered. For example, if a road is located between two water objects it is defined as a bridge. A further neighborhood application is the automated recoding/combination of shrubs which are located directly beside trees or forests stocks into tree or forest: because of the interpolated DSM and the fixed threshold value for classifying high vegetation the edges of trees or forests are often classified as shrubs instead of tree/forest. In Figure 5c and 5d this automated step and, in addition, the rare necessity of final checking is illustrated. A flowchart-based modeler tool was used to formalize and automate this hierarchical classification procedure July 2006
839
04-136
6/9/06
9:56 AM
Page 840
Figure 3. Hierarchical classification process for high-resolution image data.
Figure 4. Separation of information into different independent layers: (a) Full HRSC-AX data set with 8 bands (original data plus additional information); (b), (c), and (d) are pre-classified semantic layers-also containing 8 bands; (b) ‘non-vegetation/sparse vegetation/water/shadow’, (c) ‘low vegetation’ and (d) ‘high vegetation’.
TABLE 3. Priority Class
840
PRIORITY ORDER
AT THE
COMBINING PROCESS
High
Medium
Low
Classified Biotope Types from Herbaceous Vegetation
Classified Biotope Types from High Vegetation
Classified Biotope Types from Non-Vegetation/ Sparse Vegetation/ Water/Shadow
July 2006
which was then implemented in the software. For simplifying the operability, furthermore, the IGF developed a software module “Biotope Type Classification for Ultra High-resolution Data.” The final output was a GIS layer with biotope types for the study sites. A comprehensive synopsis of the results is given in Table 4. The differences between the final classification result and older maps created by fieldwork and photo interpretation are presented in Figure 6 and Table 5 for a small test site located at the Elbe River. The richness of PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
04-136
6/9/06
9:56 AM
Page 841
Figure 5. Steps of the GIS-based post-processing: (a) RGB, (b) combined classified biotype types, (c) filling shadows, and (d) automated recode: shrubs-forest and manual editing.
TABLE 4.
SYNOPSIS OF THE RESULTS FOR UNDER INVESTIGATION
THE
TOTAL AREA
Area Sum [hectares]
Percentage
Fields Forests/Trees/Wooded Areas Gardens Shrubs Trenches Wet Grassland Medium Grassland Oligotrophic dry grassland Bare Soil Pioneer Vegetation Reeds Ruderal Vegetation Residential Area Impervious Tideland
30 467 43 343 19 96 1521 238 324 2 340 441 21 265 2440 6590
0,45 7,08 0,66 5,20 0,28 1,46 23,09 3,61 4,92 0,03 5,16 6,70 0,32 4,02 37,03 100,00
(Water Surface)
8595
TABLE 5.
CLASSIFICATION RESULTS
Class Name Shrubs (Salix) Reed (Phragmites) Single Trees Tidal Creek/Tideland Sands Ruderal Vegetation Willow Forest
Area Sum 1995 [m2]
Area Sum 2002 [m2]
10629 39333 0 26545 0 0 2331 78838
692 31309 4829 25520 1134 1559 13796 78838
detail of the classification results corresponds well with the structures in the original image. The visual interpretation result readily shows the generalization that was performed by the human operator. The older reference map shows the subset mapped with only a few polygons whereas the new classification result consists of more and better fitting polygons described by over 1,400 vertices. Even single trees, shrubs, or open forest areas smaller than 100 square meters can be detected over large areas.
Accuracy and Change Analysis
Figure 6. Comparison of reference map and result after hierarchical classification: (a) classification result (2002), (b) HRSC-AX image data (2002), and (c) reference map (1995).
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Considerations for Accuracy Assessment Considerations have to take into account that an accuracy assessment of the resulting high-resolution biotope type maps requires contemporary field data and cannot be made based on outdated and generalized reference data like the map of 1995 (see subset in Figure 6). Furthermore, previous classification results of high-resolution data (e.g., two years earlier) could not be used for accuracy assessment because natural and human induced changes to biotope types would be characterized as mapping inaccuracies. This is especially true in this study area, where for example, different water levels have an enormous effect on tidally influenced classes (e.g., tidal land, tidal creeks). Hence, field measurements and visual interpretation based on high-resolution datasets are the only possibilities for accuracy assessment. Because of the large study area, our accuracy assessment is based mostly on visual examination and some exemplary field measurements which focused on the specific biotope types of interest (e.g., different types of reed, trees/forest, and ruderal vegetation). Due to the fact that the classification results are the basis for change analyses and the assessment of the impacts of hydraulic engineering measures, a visual checking and manual correction of the automatically-generated classification results July 2006
841
04-136
6/9/06
9:56 AM
Page 842
are indispensable. For the statistical accuracy assessment of the automated classification, the classification result without manual editing was visually checked using the original image data. Based on this visual verification, the final edited result was created by manual correction. In Plate 2a, a comparison of the original HRSC-AX data (true color display), the classification
result (before final editing), and the final result is illustrated. For this area, a thorough statistical accuracy check was conducted. Classical analyses such as random sampling, random stratified sampling and equalized random method yielded overall accuracies between 70 percent and 91 percent
Plate 2. (a) Area of the accuracy check and comparison of original HRSC-AX data in RGB display (left side images), classification result (without manual editing) (center images) and final result (classification after final manual editing) (right side images); (a) the classification result before final editing and the final result, (b) areas which could not be properly classified.
842
July 2006
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
04-136
6/9/06
9:56 AM
Page 843
TABLE 6.
ACCURACY CHECK STATISTICS
Random Method Number of Points Overall Classification Accuracy Overall Kappa Statistics Class Name Reed Reed, disturbed Reed with Shrubs Canary Reed Grass Ruderal Vegetation Ruderal Vegetation with Shrubs Mesophile Grassland Intensive Grassland Sands Coastal Protection Buildings Alley Sport Facility River Harbor Tidal Creek/Tidal Land Shrubs Single Tree Willow Forest
Random Stratified Method
256 91.41% 0.89
256 86.00% 0.81
Equalized Random Method 360 (20 per class) 70.00% 0.68
Producers Accuracy
Users Accuracy
Producers Accuracy
Users Accuracy
Producers Accuracy
98.36% 50.00% 100.00% — 85.71% — 100.00% 100.00% 100.00% — 100.00% — 100.00% — — 88.89% 100.00% 100.00%
96.77% 100.00% 100.00% — 100.00% — 100.00% 100.00% 71.43% — 100.00% — 78.57% — — 100.00% 100.00% 100.00%
98.94% 100.00% 100.00% — 71.43% — 100.00% 90.00% 50.00% — 100.00% 100.00% 100.00% — — 81.82% 100.00% 83.33%
94.90% 75.00% 100.00% — 100.00% — 80.00% 90.00% 66.67% — 83.33% 100.00% 77.05% — — 90.00% 100.00% 90.91%
95.00% 80.00% 70.00% — 100.00% 75.00% 100.00% 85.00% 60.00% — 80.00% 80.00% 100.00% — — 90.00% 95.00% 95.00%
(see Table 6). Accuracies for classes of interest (e.g., different types of reed, ruderal vegetation, and willow forest) could be mapped in most cases with an accuracy exceeding 90 percent (see Table 6, except “Canary Reed Grass” – see explanation below). It was possible to identify 14 different classes automatically instead of 18 biotope types by visual interpretation. The four classes which could not be detected are “Canary Reed Grass,” “Coastal Protection Building,” “Harbor,” and “Tidal Creek.” Apart from Canary Reed Grass these biotope types cannot be differentiated by spectral, textural or height information from other biotope types, because the differences are only based on the knowledge of the experts in the field. Examples for areas which could not properly classified are demonstrated in Plate 2b. Because of the relatively high water level the class “Tidal Creek” (yellow ellipse) is classified as “River.” The whole class “Tidal Creek/Tidal Land” with 54,773 m2 and also the class “Harbor” with 4,562 m2 in this test area were classified as “River.” This accounts for the area difference for the class “River” (58,014 m2) between the automated classification result and the revised classification result (see the difference column in Plate 2). For the class “Coastal Protection Building,” examples are marked with white ellipses. The automated classification assigned these regions to the classes “Sand” and “Ruderal Vegetation.” In this case, only the surfaces of the “Coastal Protection Building” could be classified (instead of the whole object). The class “Canary Reed Grass” could not be identified as a separate class and was classified as “Reed.” The reasons are (a) the spectral, textural and height similarity to “Reed,” (b) the rare existence of this biotope type (one sample only), and (c) the occlusion by shadows. Because of these reasons, the low user’s accuracy of the class “Reed” in the Equalized Random Method is understandable. The discrepancy in the blue ellipse is also caused by the spectral similarity to another class (here “Alley” versus “Grassland”). The problem with shadows is illustrated by the area marked with a red ellipse in Plate 2b. This area shows some misclassification in its shadow parts that could not be corrected by automatic means. Overall, however, the accuracy for the biotope type classes to be monitored proved to be quite sufficient to the PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Users Accuracy 44.19% 94.12% 100.00% — 100.00% 100.00% 83.33% 89.47% 42.86% — 76.19% 100.00% 31.25% — — 85.71% 100.00% 73.08%
German Waterways and Shipping Administration. The level of detail and the excellent geometric accuracy exceeded all previous analyses based on photointerpretation and fieldwork. Consequently, this final result in GIS format will serve as the basis for future change detection analyses. Change Analysis Preliminary results suggest that the new methods may be useful for change analysis. These estimates can be given on the basis of two comparisons. On one hand, new results of the automatic classification can be compared to the older reference maps of 1995 (see Figure 6 and Table 5). On the other hand, the two stages of the automatic classification results 1999 and 2002 can be compared with each other (for an excerpt, see Figure 7). Both comparisons reveal a variety of differences. Especially for the first case, some of the differences result from different analysis methods, e.g., the distinction
Figure 7. Reed development between 2000 and 2002 for a subset showing a total loss of 1,040 m2.
July 2006
843
04-136
6/9/06
9:56 AM
Page 844
between “forest,” “single trees,” and “shrubs” or the distinction between “reed” and “ruderal vegetation.” In the second case, some differences may result from different water levels during data collection. Due to these problems, examination of our results often has to be reduced to single phenomena to exclude errors based on different environmental conditions. Apart from these methodically determined variations a set of true differences occurs, especially concerning the reed vegetation adjacent to the waterline. Figure 7 shows gain and loss of reed directly along the waterline resulting from a direct GIS-based comparison between the classification results of 2000 and 2002. These differences are based on true changes because the two classification results have been achieved using the same methods and technology. Furthermore, the changes were verified in field. These changes demonstrate the application potential and the high mapping accuracy of the methods presented here. Within a small subset of the study area (40,000 m2) a loss of more than 1,000 m2 of reed within two years could be estimated. To extrapolate this to the entire area under investigation would result in a loss of up to 1,000,000 m2 of reed. This large-scale investigation is currently being conducted by the German Waterways and Shipping Administration.
Conclusions The potential of digital airborne scanner data with excellent geometric fidelity was demonstrated for a high-resolution mapping project. Digital image data in combination with an integrated GIS/image processing environment, allowed the development of an automated classification procedure for detailed and accurate biotope mapping. This automated hierarchical technique facilitated the documentation of dynamic processes in environmental monitoring projects. The original images as well as the image classification approach and results can be integrated easily in a GIS environment. This allows operational analysis and the measurement of changes over time. Of particular importance was the effective integration with the digital surface model that was simultaneously produced from the HRSC sensors. With this, it is possible to differentiate between high and low vegetation biotope types despite their similarity in spectral reflectance. Using GIS operators such as majority filtering and rule-based overlay techniques, shadows can be eliminated and individual classification layers can be combined. The hierarchical classification procedure was formalized and presented in a flowchart environment. Classification accuracies prove to be very good, especially for the relevant biotope types that were most importance to the user community. Future investigations will incorporate of direct surface measurement from laser scanning technology for more accurate DSM generation.
References Blaschke, T., and J. Strobl, 2001. What’s wrong with pixels? Some recent developments interfacing remote sensing and GIS, GeoInformations-Systeme, 6(01):12–17. Ehlers, M., 2000. Integrated GIS – From data integration to integrated analysis, Surveying World, 9:30–33. Ehlers, M. (editor), 2002a. Remote sensing for environmental monitoring, GIS applications, and geology, Proceedings of SPIE, Volume 4545, Bellingham, Washington, 330 p. Ehlers, M., 2002b. Fernerkundung für GIS-Anwender: Sensoren und methoden zwischen Anspruch und Wirklichkeit, Fernerkundung und GIS: Neue Sensoren – Innovative Methoden (T. Blaschke, editor), Wichmann Verlag, Heidelberg, pp. 10–23.
844
July 2006
Ehlers, M., 2004. Remote sensing for GIS applications: New sensors and analysis methods, remote sensing for environmental monitoring, GIS applications, and geology III (M. Ehlers, H.J. Kaufmann, and U. Michel, editors), Proceedings of SPIE, Volume 5239, Bellingham, Washington, pp. 1–13. Ehlers, M., M. Gaehler, and R. Janowsky 2003a. Automated analysis of ultra high-resolution remote sensing data for biotope type mapping: New possibilities and challenges, ISPRS Journal of Photogrammetry and Remote Sensing, 57:315–326. Ehlers, M., R. Janowsky, and M. Gaehler 2003b. Ultra high resolution remote sensing for environmental monitoring, Earth Observation Magazine, 12(9):27–32. Ehlers, M., J. Schiewe, and M. Möller 2003(c). 3D city modelling using ultra high resolution and multi-sensoral remote sensing, Geo-Informations-Systeme, 6(03):30–37. Ferrano, G., and C. Feix, 2003. The Z/I Imaging DMC – Digital Mapping Camera system status, configuration and calibration, ASPRS 2003 Annual Conference Proceedings, unpaginated CD-ROM. Gruber, M., F. Leberl, and R. Perko, 2003. Paradigmenwechsel in der Photogrammetrie durch digitale Luftbildaufnahme, Photogrammetrie, Fernerkundung, Geoinformation (PFG), 4:285–297. Hinz, A., C. Dörstel, and H. Heier, 2001. DMC - The digital sensor technology of Z/I-Imaging, Proceedings of Photogrammetric Week 2001, (D. Fritsch, and R. Spiller, editors), Wichmann, Heidelberg, pp. 93–103. Hoffmann, A., and F. Lehmann, 2000. Vom Mars zur Erde – die erste digitale Orthobildkarte Berlin mit Daten der Kamera HRSC-A, Kartographische Nachrichten, 50(2):61–71. Hoffmann, A., J.W. van der Vegt, and F. Lehmann, 2000. Towards automated map updating: Is it feasible with new digital data acquisition and processing techniques?, International Archives of Photogrammetry and Remote Sensing (IAPRS), Amsterdam, 33/B2, pp. 295–302. Jensen, J.R., 2004. Introductory Image Processing: A Remote Sensing Perspective, 3rd Edition, Prentice-Hall, Upper Saddle River, New Jersey, 526 p. Leberl, F., and M. Gruber, 2003. Economical large format aerial digital camera, GIM International, 17 (6):1–5. Metternicht, G., 1999. Change detection assessment using fuzzy sets and remotely sensed data: An application of topographic map revision, ISPRS Journal of Photogrammetry and Remote Sensing, 54 (4):221–233. Möller, M., 2003. Urbanes Umweltmonitoring Mit Digitalen Flugzeugscannerdaten, Wichmann Verlag, Heidelberg. 126 p. Neukum, G., and HRSC-Team, 2001. The airborne HRSC-AX cameras: Evaluation of the technical concept and presentation of application results after one year of operation, Proceedings of Photogrammetric Week 2001, (Fritsch D. and R. Spiller, editors), Wichmann, Heidelberg, pp. 117–130. Petit, C.C., and E.F. Lambin, 2001. Integration of multi-source remote sensing data for land cover change detection, International Journal of Geographical Information Science, 15(8):785–803. Sandau, R., B. Braunecker, H. Driescher, A. Eckardt, S. Hilbert, J. Hutton, W. Kirchhofer, E. Lithopoulos, R. Reulke, and S. Wicki, 2000. Design principles of the LH Systems ADS40 airborne digital sensor, International Archives of Photogrammetry and Remote Sensing (IAPRS), Amsterdam, 33/B1, pp. 258–265. Schiewe, J., 2003. Integration of multi-sensor data for landscape modeling using region-based approach, ISPRS Journal of Photogrammetry and Remote Sensing, 57(5–6):371–379. Wewel, F., and F. Scholten, 1999. High-resolution stereo camera (HRSC): Multispectral data acquisition and photogrammetric data processing, Proceedings of the 4th International Airborne Remote Sensing Conference and Exhibition, Ottawa, Volume 1, pp. 263–271.
(Received 21 June 2004; accepted 01 February 2005; revised 31 May 2005)
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
