Automated Damage Assessment for Rapid Geospatial ...

0 downloads 0 Views 3MB Size Report
Daniel HÖLBLING and Stefan LANG. 1 Introduction. The FP7 project G-MOSAIC (GMES services for Management of Operations, Situation. Awareness and ...
Automated Damage Assessment for Rapid Geospatial Reporting – First Experiences from the Haiti Earthquake 2010 Dirk TIEDE, Christian HOFFMANN, Petra FÜREDER, Daniel HÖLBLING and Stefan LANG

1

Introduction

The FP7 project G-MOSAIC (GMES services for Management of Operations, Situation Awareness and Intelligence for regional Crises, http://www.gmes-gmosaic.eu/) aims at identifying and developing products, methodologies and pilot services for the provision of geo-spatial information in support of EU external relations policies and at contributing to define and demonstrate the sustainability of GMES global security services. Following the Haiti earthquake on January 12th, 2010, the G-MOSAIC Rapid Geospatial Reporting Service has been activated. It was initially requested by the UN cartographic section and the Spanish Red Cross in order to produce geo-spatial products in rush mode to assist relief efforts in Haiti. Satellite imagery immediately acquired after the disaster were processed by G-MOSAIC partners and first geo-spatial information were delivered to the users on January 16th. In this paper we share some experiences resulting from automated damage analysis in this context.

2

Study Area and Data

The Haitian towns Carrefour and Léogâne are located within the Ouest Department of Haiti. The coastal town Carrefour is located about 6 km west of the capital Port-au-Prince (see figure 1).

Fig. 1: Overview of the study area west of Port-au-Prince

D. Tiede, C. Hoffmann, P. Füreder, D. Hölbling and S. Lang

208

Léogâne lies about 30 km westwards of Port-au-Prince, only 12 km northwest from the epicenter of the earthquake (USGS, 2010). The damage analyses for Carrefour were conducted on GeoEye-1 pre- and post-disaster satellite imagery. For Léogâne GeoEye-1 (pre-images) and WorldView-2 (post-images) data was available. Pre-images for Carrefour were acquired on July 27th 2009 and for Léogâne on October 1st 2009. Post-images for Carrefour were taken on January 13th 2010 and for Léogâne on January 15th 2010. The spatial resolution of the GeoEye-1 and WorldView-2 sensors is 0.5 m in the panchromatic band for commercial use and 1.65 m (1.84 m for WorldView-2) in the multispectral bands. The pre- and post-images for Carrefour consisted of a multispectral image with three optical bands, a NIR band and a panchromatic image. The images for Léogâne were delivered in pan-sharpened format, but only RGB bands were made available.

3

Methods

3.1 Automated Damage Assessment Rulesets were developed in an object-based image analysis (OBIA) environment to automatically extract relevant information as indications for damaged buildings. Rulesets for information extraction and change detection analysis were written using Cognition Network Language (CNL), a modular programming language implemented in the Definiens eCognition 8 software environment. Rule-based classifiers are used for knowledge representation, making explicit the required spectral and geometrical properties as well as spatial relationships for advanced class modeling (TIEDE et al. 2010). Because of the tight time frame in this emergency situation, additional challenges, especially concerning the quality of the imagery, had to be tackled:    

Different recording conditions of the timely available images (recording angle, seasons) Time constraints in pre-processing of the imagery resulting in limited geometric accuracy and a mismatch between pre- and post-images (geometric shift) and also radiometric differences Available imagery for Léogâne was missing the fourth (NIR) band GeoEye-1’s 0.41 m and WorldView-2’s 0.46 m spatial resolution in the panchromatic band was down-sampled by the provider to 0.5 m for commercial use (and also in this humanitarian relief effort)

In the first step the images were orthorectified using the freely available ASTER Global Digital Elevation Map (GDEM), which was especially important in the hilly area in the southern part of Carrefour. The impact from the other limiting factors mentioned above was minimized in the ruleset development process. The damage assessment itself was based on indicators; in this case the shadows casted by buildings before and after the earthquake. To avoid deriving false positives from vegetation shadows a vegetation mask has been created for both images (see also VU et al. 2004). Missing NIR information for vegetation extraction in the town of Léogâne could partly be compensated by using visible greenness instead of the NDVI (Normalized Differenced Vegetation Index).

Automated Damage Assessment for Rapid Geospatial Reporting

209

The geometric inaccuracy of the pre- and post-images and the resulting positional differences of the extracted shadow objects were tackled using object-linking that overcomes strict object hierarchies. Thereby, despite the geometrical shift and divergent overlapping areas between the image objects, we achieved a spatial comparison (size, shape etc.) of the shadow objects resulting in a damage indication class.

3.2

Damage Assessment Maps

The extracted damage indicators for the two areas were analyzed and conditioned applying kernel density methods (SILVERMAN 1986). The kernel density was calculated using point features – in this case the centroids of the extracted damage indicator objects on an output raster cell size of 20 m × 20 m. The resulting maps (damage indicator maps) were supposed to give a fast and easy to grasp overview about the spatial distribution and intensity of damages in the area (see figure 2).

4

Results and Discussion

Figure 2 shows the resulting damage assessment map for the area of Carrefour as it was delivered to the requesting users. An additional map was produced for the area of Léogâne. Distributed computing enabled the rapid processing of the data sets. A multi-core blade server installed at Definiens (Definiens AG, Munich) consisting of ~ 20 – 60 cores

Fig. 2:

Damage assessment map based on the automated approach for Carrefour. Darker tones indicate higher damage density.

210

D. Tiede, C. Hoffmann, P. Füreder, D. Hölbling and S. Lang

(dependent on load balancing) could be used for the analysis and reduced the processing time for the area of Carrefour (17 km2 and 0.5 m GSD) to approximately 4 minutes. Both maps did undergo a project-internal quality check by the service chain leader EUSC (European Union Satellite Centre). The map of Carrefour passed the quality check and was published and delivered to the users. The damage assessment for Léogâne showed good results in the outskirts of the town, but an underestimation of damages in the city centre compared to manual assessments. Two reasons were indentified, which mainly caused problems: (1) The WorldView-2 images for Léogâne were only provided with three spectral bands (RGB), the missing NIR band hampered the differentiation between vegetation / shadow of vegetation and some of the buildings; (2) a lot of construction work in the area led to undesired results (also in the case of visual interpretation). Therefore the map was not delivered to the user but was used internally to support the manual assessment in this area (for cross-checking). Additional validation of the results was conducted by comparing them with several manually digitized damage assessment maps provided by other institutions. The results confirmed the findings of the project-internal quality check. Additionally, an accuracy assessment for 100 randomly selected damage indicators for each site revealed a user’s accuracy of more than 72 % for the Carrefour area and more than 80 % for the area of Léogâne. For the central area of Carrefour some false detected damages were found which resulted from higher buildings and the shadows casted by these causing differences between the pre- and the post-imagery (due to slightly different viewing angles). This could be tackled in the future through threshold modifications in the ruleset. A fairly high user’s accuracy in this automated analysis shows the potential of supporting the more time consuming manual analysis. It has to be stated, that the aim of such an automated approach is not to replace the manual interpretation. It rather helps users and manual interpreters to get a faster impression of the spatial distribution of damages in emergency situations. By purpose the provided maps show no absolute values of detected damages but only tendencies. Absolute figures are very much depending on the resolution of the available imagery and even on aerial images not all damages are visible from the bird’s eye view.

References SILVERMAN, B. W. (1986), Density Estimation for Statistics and Data Analysis. Chapman and Hall, New York. TIEDE, D., LANG, S., ALBRECHT, F. & HÖLBLING, D. (2010), Object-based class modeling for cadastre-constrained delineation of geo-objects. Photogrammetric Engineering & Remote Sensing, 76 (2), pp. 193-202. USGS – U.S. GEOLOGICAL SURVEY (2010), Magnitude 7.0 – Haiti Region. – http://earthquake.usgs.gov/earthquakes/eqinthenews/2010/us2010rja6/ (accessed: 22 Apr. 2010). VU, T. T., MATSUOKA, M. & YAMAZAKI, F. (2004), Shadow analysis in assisting damage detection due to earthquakes from QuickBird imagery. Proceedings of the 10th international society for photogrammetry and remote sensing congress, pp. 607-611.