Presentation is loading. Please wait.

Presentation is loading. Please wait.

Satellite Images for Flood Monitoring and Damage Assessment Immagini da satellite per il monitoraggio delle alluvioni e la valutazione del danno S.

Similar presentations


Presentation on theme: "Satellite Images for Flood Monitoring and Damage Assessment Immagini da satellite per il monitoraggio delle alluvioni e la valutazione del danno S."— Presentation transcript:

1 Satellite Images for Flood Monitoring and Damage Assessment Immagini da satellite per il monitoraggio delle alluvioni e la valutazione del danno S. B. Serpico¹ G. Moser¹, S. Dellepiane¹, E. Angiati¹ G. Boni², R. Rudari² ¹ University of Genoa, Italy ² CIMA Research Foundation, Savona, Italy

2 Outline Introduction Remote sensing for flood response and management Joining hydrology/hydraulics and remote sensing Key ideas and overall architecture Land cover, 3D buildings, changed/flooded areas from remote sensing Vulnerability, elements at risk, and damage Examples of advanced methods for land cover and change mapping Markov random fields, region-based analysis, multisource fusion Experimental examples and case studies Tanaro River (Italy), 2009 Shkodër (Albania), 2010 Conclusion

3 Outline Introduction Remote sensing for flood response and management Joining hydrology/hydraulics and remote sensing Key ideas and overall architecture Land cover, 3D buildings, changed/flooded areas from remote sensing Vulnerability, elements at risk, and damage Examples of advanced methods for land cover and change mapping Markov random fields, region-based analysis, multisource fusion Experimental examples and case studies Tanaro River (Italy), 2009 Shkodër (Albania), 2010 Conclusion

4 Introduction Satellite remote sensing and Earth observation (EO) for disaster monitoring and damage assessment Spatially distributed and temporally repetitive observations Multispectral and synthetic aperture radar (SAR) images All-weather and day-and-night through SAR Very high resolution (VHR) (up to ~30 cm) and short revisit times (12 to 24 h) with current EO missions (e.g., COSMO-SkyMed, Pléiades) Critical need for a multidisciplinary approach: remote sensing and geophysical sciences Boulder County (CO), September 2013

5 Introduction Focus of the talk
Potential of remote sensing and its combination with hydraulic modeling for flood monitoring and damage assessment Special focus on advanced approaches to land cover classification and change detection from remote sensing in flood applications Approaches developed mainly within applied research projects: OPERA, Protezione civile dalle alluvioni, Italian Space Agency and Italian Department of Civil Protection, COSMO-SkyMed AO ID-2181(1), Italian Space Agency, PRIN(2), Italian Minister of Education, University, and Research, (1) Development and validation of multitemporal image analysis methodologies for multirisk monitoring of critical structures and infrastructures (2) Very high spatial and spectral resolution remote sensing: a novel integrated data analysis system

6 Remote Sensing vs. Flood Risk
Prevention and prediction phases Improved predictions of hydro-meteorological processes by mapping land cover and water bodies from EO Mitigation and risk assessment phase Often out-of-date cartography; time consuming in situ surveys. Improved elements-at-risk and vulnerability assessment by mapping land cover and buildings through remote sensing image analysis Monitoring and management phases Multitemporal remote sensing image analysis (before vs. after event) allows assessing the flood impact (flooded areas, ground changes) Need for a methodology that incorporates hydraulic modeling to estimate damage from these EO-derived thematic products. Here, focus on these two aspects

7 Outline Introduction Remote sensing for flood response and management Joining hydrology/hydraulics and remote sensing Key ideas and overall architecture Land cover, 3D buildings, changed/flooded areas from remote sensing Vulnerability, elements at risk, and damage Examples of advanced methods for land cover and change mapping Markov random fields, region-based analysis, multisource fusion Experimental examples and case studies Tanaro River (Italy), 2009 Shkodër (Albania), 2010 Conclusion

8 Key Ideas of the Proposed Approach
 Land cover ► Elements at risk EO-based land-cover classes are reorganized into classes of elements at risk that show similar behaviors with respect to flooding.  Elements at risk ► Vulnerability Classes of elements at risk are assigned proper loss functions so that damage is known as a function of hydraulic forcing.  Flooded/changed areas ► Flood exposure indicators Through physically-based hydraulic models constrained to EO-based flooded or changed areas, water-depth maps and other hydraulic parameters can be computed as best estimates of the real condition.  All derived information is merged to map the actual damage.

9 Overall Architecture Ground truth data for training
Possible further land cover map Multispectral (VHR) image Multispectral VHR image Supervised classification Land cover map Segmentation, edge detection, 3D modeling Extracted 3D buildings Elements-at-risk mapping Elements-at-risk map System of loss functions Vulnerability map Estimation of aereal or building-specific damage percentages Damage map Image pair (before-after the flood) Unsupervised change detection Flooded-area color display Flooded-area mapping Change map Color display of flooded areas Map of flooded areas Hydraulic modeling Flood exposure indicators Digital terrain model S. B. Serpico, S. Dellepiane, G. Boni, G. Moser, E. Angiati, R. Rudari,, Proceedings of the IEEE, 100: , 2012

10 Land Cover from Multispectral Images
Ground truth data for training Possible further land cover map Land cover map Input image with training pixels Multispectral (VHR) image Supervised classification Land cover map Elements-at-risk mapping Elements-at-risk map System of loss functions Vulnerability map Multispectral VHR image Segmentation, edge detection, 3D modeling Extracted 3D buildings Estimation of aereal or building-specific damage percentages Primary role of optical multi/hyperspectral images: Medium resolution (~30 m) to VHR (0.3 to 5 m), depending on the size of the considered basin and investigated area. Main methodological approach: supervised classification, making use of Ground truth for training (examples of the land cover classes). Most used techniques: Bayesian, kernel, neural, fuzzy; region/object- based; texture analysis; probabilistic contextual models. Image pair (before-after the flood) Unsupervised change detection Change map Damage map Flooded-area color display Color display of flooded areas Flooded-area mapping Map of flooded areas Hydraulic modeling Flood exposure indicators Digital terrain model

11 3D Buildings from VHR Images
Ground truth data for training Possible further land cover map Detected roofs and shadows VHR image Multispectral (VHR) image Supervised classification Land cover map Elements-at-risk mapping Elements-at-risk map System of loss functions Vulnerability map Multispectral VHR image Segmentation, edge detection, 3D modeling Extracted 3D buildings Estimation of aereal or building-specific damage percentages Image pair (before-after the flood) Unsupervised change detection Change map Damage map Main methodological approaches: 3D reconstruction from stereo optical pairs, LiDAR, SAR interferometry, segmentation and edge/shadow detection. Flooded-area color display Color display of flooded areas Flooded-area mapping Map of flooded areas Hydraulic modeling Flood exposure indicators Digital terrain model

12 Elements at Risk Ground truth data for training Possible further land cover map Multispectral (VHR) image Supervised classification Land cover map Elements-at-risk mapping Elements-at-risk map System of loss functions Vulnerability map Multispectral VHR image Segmentation, edge detection, 3D modeling Extracted 3D buildings Estimation of aereal or building-specific damage percentages Mapping EO-based land-cover classes and possible ancillary land covers (e.g., Corine) to element-at-risk classes through a lookup table. Interactive many-to-many mapping. Image pair (before-after the flood) Unsupervised change detection Change map Damage map Flooded-area color display Color display of flooded areas Flooded-area mapping Map of flooded areas Hydraulic modeling Flood exposure indicators Digital terrain model

13 Vulnerability Ground truth data for training Possible further land cover map Multispectral (VHR) image Supervised classification Land cover map Elements-at-risk mapping Elements-at-risk map System of loss functions Vulnerability map Flood damage [%] (0 ≤ D ≤ 1) Flood exposure indicators (e.g., water depth) ith element-at-risk class or building type Multispectral VHR image Segmentation, edge detection, 3D modeling Extracted 3D buildings Estimation of aereal or building-specific damage percentages Degree of loss of an element at risk, due to an event of given magnitude. Main methodological approach: loss functions Image pair (before-after the flood) Unsupervised change detection Change map Damage map Flooded-area color display Color display of flooded areas Flooded-area mapping Map of flooded areas Hydraulic modeling Flood exposure indicators Digital terrain model

14 Change Detection from Multitemporal Images
Ground truth data for training Primary role of SAR for change detection in flood applications. Main methodological approach: unsupervised detection, since Ground truthing is often incompatible with emergency applications. Most used techniques: Bayesian decision, finite mixtures, Markov models, expectation-maximization, wavelets, segmentation. Possible further land cover map Multispectral (VHR) image Supervised classification Land cover map Elements-at-risk mapping Elements-at-risk map System of loss functions Vulnerability map Multispectral VHR image Segmentation, edge detection, 3D modeling Extracted 3D buildings Estimation of aereal or building-specific damage percentages Before After Image pair (before-after the flood) Unsupervised change detection Change map Damage map Flooded-area color display Color display of flooded areas Flooded-area mapping Map of flooded areas Hydraulic modeling Flood exposure indicators Digital terrain model

15 Flooded Area Mapping from Multitemporal Images
Multispectral (VHR) image Multispectral VHR image Before After Ground truth data for training Possible further land cover map Primary role of SAR for flooded area mapping. Display of flooded areas: cross-normalization of multitemporal SAR images, fusion into color composites. Detection of flooded areas from SAR: hard or fuzzy segmentation (especially seed-growing) of multitemporal SAR images. Supervised classification Land cover map Elements-at-risk mapping Elements-at-risk map System of loss functions Vulnerability map Segmentation, edge detection, 3D modeling Extracted 3D buildings Estimation of aereal or building-specific damage percentages Image pair (before-after the flood) Unsupervised change detection Change map Damage map Flooded-area color display Color display of flooded areas Flooded-area mapping Map of flooded areas Hydraulic modeling Flood exposure indicators Digital terrain model

16 Water Depth Ground truth data for training Possible further land cover map Multispectral (VHR) image Supervised classification Land cover map Elements-at-risk mapping Elements-at-risk map System of loss functions Vulnerability map Main methodological approach: 2D hydraulic modeling. Shallow water equations for real-time application. Boundary and initial conditions based on flooded/changed areas from remote sensing. Ensemble of model runs. The ensemble member that best matches flood extension is selected. Multispectral VHR image Segmentation, edge detection, 3D modeling Extracted 3D buildings Estimation of aereal or building-specific damage percentages Image pair (before-after the flood) Unsupervised change detection Change map Damage map Flooded-area color display Color display of flooded areas Flooded-area mapping Map of flooded areas Hydraulic modeling Flood exposure indicators Digital terrain model

17 Damage Mapping Ground truth data for training Possible further land cover map Multispectral (VHR) image Supervised classification Land cover map Elements-at-risk mapping Elements-at-risk map System of loss functions Vulnerability map Multispectral VHR image Segmentation, edge detection, 3D modeling Extracted 3D buildings Estimation of aereal or building-specific damage percentages Image pair (before-after the flood) Unsupervised change detection Change map Damage map Flooded-area color display Color display of flooded areas Given water depth and loss functions, damage is derived. Flooded-area mapping Map of flooded areas Hydraulic modeling Flood exposure indicators Digital terrain model

18 Outline Introduction Remote sensing for flood response and management Joining hydrology/hydraulics and remote sensing Key ideas and overall architecture Land cover, 3D buildings, changed/flooded areas from remote sensing Vulnerability, elements at risk, and damage Examples of advanced methods for land cover and change mapping Markov random fields, region-based analysis, multisource fusion Experimental examples and case studies Tanaro River (Italy), 2009 Shkodër (Albania), 2010 Conclusion

19 Supervised Land Cover Classification
Pattern recognition formalization The input image X and the output land cover map Y are (realizations of) 2D stochastic processes (random fields). Goal: estimate Y, given X and the training pixels Need to take benefit from spatial-geometrical information, especially at VHR Region-based methods: combine segmentation and classification, and are especially effective for classes with geometrical structures (e.g., urban). Markov random field (MRF) models for the joint statistics of distinct pixels IKONOS, 4-m resolution, Alessandria (Italy), RGB color composite with superimposed training pixels

20 Markov Random Fields MRF models MRF models for classification
Representation of the statistical interactions among the class labels of the pixels, by using only local relationships (neighborhood): MRF models for classification Bayesian decision becomes equivalent to minimizing a locally defined energy function U(Y| X). Efficient minimization methods (e.g., graph cuts). By suitably defining the energy, both spatial information and multiple information sources (e.g., multisensor, multiscale) can be fused for land-cover mapping purposes. Pixel j is a neighbor of pixel i (j ~ i)

21 Multiscale Region-Based Markovian Classifier for VHR
Incorporating both neighborhood and region information The multiscale approach is especially appealing for VHR. Computing from X segmentation maps (S) at multiple (K) scales Finer scales: precise spatial details, sensitivity to noise Coarser scales: poor details, stronger immunity to noise MRF to fuse class (Y) and segment (S) labels Multiscale region-based energy One energy contribution for each scale, i.e., each segmentation map Class-conditional distribution of the region label at each scale Neighborhood energy Favors spatial smoothness of output map. G. Moser, S. B. Serpico, J. A. Benediktsson, Proceedings of the IEEE, 101: , 2013

22 Unsupervised Change Detection
A “dissimilarity” image X, which emphasizes the discrimination between changed and unchanged areas, is usually extracted. X and the change map Y are random fields. X may be single-channel (e.g., ratio of two SAR amplitude images) or multichannel (e.g., in the case of multispectral or multisensor images) Goal: estimate Y, given X Need to model data statistics without prior information (no training pixels) and to ensure robustness to speckle (SAR data) and noise Time t₀ Time t₁ Comparison operator (difference, ratio, log-ratio, information-theoretic distance, etc.) X Binary unsupervised classification (“change” vs. “no-change”) Y

23 Data-Fusion Markovian Change Detection
Fusion of spatial and multisensor/multichannel information Individual channels (X) modeled as separate information sources MRF to fuse change labels (Y) and multiple sources (X) Multichannel/multisource energy One energy contribution per channel Probability density function of each channel, given “change” or “no-change” Neighborhood energy Favors robustness to noise and speckle Unsupervised density estimation Parametric density modeling by integrating EM, method of log-cumulants, iterated conditional expectation G. Moser, S. B. Serpico, IEEE TGRS, 47: , 2009 L. Gomez-Chova, D. Tuia, G. Moser, G. Camps.Valls, Proceedings of the IEEE, 103: , 2015

24 Outline Introduction Remote sensing for flood response and management Joining hydrology/hydraulics and remote sensing Key ideas and overall architecture Land cover, 3D buildings, changed/flooded areas from remote sensing Vulnerability, elements at risk, and damage Examples of advanced methods for land cover and change mapping Markov random fields, region-based analysis, multisource fusion Experimental examples and case studies Tanaro River (Italy), 2009 Shkodër (Albania), 2010 Conclusion

25 Experimental Examples and Case Studies
Tanaro River, Italy, 2009 Flood near Alessandria, April 28, 2009 Heavy widespread rainfall in alpine and prealpine areas Tanaro mainly flooded the riverbanks, up to of ~2 km extension, with water depth up to 2-3 m 40 flooded buildings, 6000 people temporarily evacuated Flooded area on April 28, 2009

26 Tanaro: Land Cover Map IKONOS, 4-m res., with test samples Non-contextual Bayesian Classical MRF-based High accuracy on test samples Noisy behavior of non- contextual classification Improved border regularity of the multiscale region-based MRF over classical MRF Importance of multiscale and spatial information for VHR Multiscale region-based Markovian

27 Tanaro: from Land Cover to Vulnerability
IKONOS, 4-m res. Land cover map Elements-at-risk map Vulnerability map

28 Tanaro: Flooded Area and Change Maps
CSK, 1 day after the flood CSK, 2 days after the flood Flooded area map (seed-growing segmentation) Change map

29 Tanaro: 3D Buildings, Water Depth, Damage
Very short time scale of the flooding: hard to capture the max flood extension through remote sensing. The hydraulic model, initialized with EO- detected flooded areas, reconstructs water passage while maintaining hydraulic connectivity. Individual identified buildings could be marked with damage values through the hydraulic model although they were outside the EO-detected flooded area. Damage evaluation without hydrological modeling Damage evaluation with hydraulic modeling Extracted buildings

30 Experimental Examples and Case Studies
Shkodër, Albania, 2010 Huge flood on Jan 11, 2010 Heavy rainfall and reduced snow accumulation (high temperatures) Authorities were forced to release water from three hydroelectric power lakes. 10500 ha inundated, 2500 flooded houses, 6000 people evacuated Flood lasted until the end of Jan 2010. Shkodër (Albania), flooded area, 2010

31 Shkodër: from Land Cover to Vulnerability
Landsat-5 TM, 30-m res. Land cover map Elements-at-risk map Vulnerability map (Classification method in G. Moser, S. B. Serpico, IEEE TGRS, 51: , 2013)

32 Shkodër: Flooded Area Maps
CSK, just after the flood CSK, 21 days after the flood Flooded area color display Flooded-area map

33 Shkodër: from Vulnerability to Damage
Flooded area ► Flood depth ► Flood velocity ► Damage [%]

34 Shkodër: Water Depth and Damage
Water depth without EO Water depth with EO Damage map Only a low-resolution (90 m) DTM was available. Major break lines for the flood are evident in the EO result, but are fragmented in the DTM due to sampling issues. Without remote sensing and with only the DTM, the hydraulic model would erroneously expand much more to the south as compared with what can be seen from the satellite images.

35 Outline Introduction Remote sensing for flood response and management Joining hydrology/hydraulics and remote sensing Key ideas and overall architecture Land cover, 3D buildings, changed/flooded areas from remote sensing Vulnerability, elements at risk, and damage Examples of advanced methods for land cover and change mapping Markov random fields, region-based analysis, multisource fusion Experimental examples and case studies Tanaro River (Italy), 2009 Shkodër (Albania), 2010 Conclusion

36 Conclusion Synergy between remote sensing and hydro-meteorology is crucial for flood monitoring and damage assessment. Case studies pointed out relationships and complementarities. Using only one of these two components separately would lead to erroneous or more limited results. Complementary properties also wrt in situ surveys: higher local accuracy vs. spatially distributed and repetitive mapping. Accurate mapping of flood-related thematic products through image processing and pattern recognition techniques Current maturity of image and pattern recognition supports not only laboratory experiments but also operational applications.

37 Related and Future Work
Keeping up to date with new missions, sensors, processing capabilities and integrating them in the operational chains for flood management and damage assessment Cloud or GPU processing Need for multitemporal analysis methods that are robust to heterogeneous acquisitions (e.g., different polarizations, different acquisition geometries) Extension to other environmental risks (e.g. earthquakes, forest fires) Reconfiguring currently consolidated operational chains for flood-risk management to exploit the information offered by remote sensing.

38 References J. Richards and X. Jia, Remote sensing digital image analysis, Springer, 2005 G. Boni and F. Siccardi, “Scenes and scenarios,” Public Service Review: European Science and Technology, vol. 10, pp , 2011 A. Leopardi, E. Oliveri, and M. Greco, “Two-dimensional modeling of floods to map risk prone areas,” J. Water Res. Planning Management, vol. 128, pp , 2002 S. B. Serpico, S. Dellepiane, G. Boni, G. Moser, E. Angiati, R. Rudari, “Information extraction from remote sensing images for flood monitoring and damage evaluation”, Proceedings of the IEEE, vol. 100, no. 10, pp , 2012 G. Moser, S. B. Serpico, and J. A. Benediktsson, “Land-cover mapping by Markov modeling of spatial-contextual information in very-high-resolution remote sensing images”, Proceedings of the IEEE, vol. 101, no. 3, pp , 2013 L. Gomez-Chova, D. Tuia, G. Moser, G. Camps-Valls, “Multimodal classification of remote sensing images: a review and future directions,” Proceedings of the IEEE, 103(9): , 2015 G. Moser and S. B. Serpico, “Unsupervised change detection from multichannel SAR data by Markovian data fusion,” IEEE Trans. Geosci. Remote Sensing, vol. 47, no. 7, 2009, pp S. B. Serpico, G. Moser, “Weight parameter optimization by the Ho-Kashyap algorithm in MRF models for supervised image classification”, IEEE Trans. Geosci. Remote Sensing, 44(12): , 2006 G. Moser and S. B. Serpico, “Combining support vector machines and Markov random fields in an integrated framework for contextual image classification”, IEEE Trans. Geosci. Remote Sensing, vol. 51, no. 5, pp , 2013 S. Dellepiane and E. Angiati, “A new method for cross-normalization and multitemporal visualization of SAR images for the detection of flooded areas,” IEEE Trans. Geosci. Remote Sensing, vol. 50, no. 7, pp , 2012 A. De Giorgi, G.Moser, S. B. Serpico, “Parameter optimization for Markov random field models for remote sensing image classification through sequential minimal optimization,” Proc. of IGARSS 2015 , pp , Milan, Italy, 2015 E. Angiati and S. Dellepiane, “Identification of roofs perimeter from aerial and satellite images”, Proc. 17th International Conference on Digital Signal Processing, Corfu, Greece, 2011


Download ppt "Satellite Images for Flood Monitoring and Damage Assessment Immagini da satellite per il monitoraggio delle alluvioni e la valutazione del danno S."

Similar presentations


Ads by Google