INFERNO - A PROJECT FOR THE INTEGRATION OF REMOTE SENSING INFORMATION IN OPERATIONAL WATER BALANCE MODELLING AND FLOOD FORECASTING
F. Appel a, H. Bach a, A. Löw b, B. Waske b, R. Ludwig b, W. Mauser b, W. Schulz c, U. Merkel c, N. Demuth d
VISTA Geowissenschaftliche Fernerkundung GmbH, Gabelsbergerstrasse 51, 80333 München Tel. +49 89 5238 9802 ; Fax. + 49 89 5238 9804 ; appel@vista-geo.de b Department für Geo- und Umweltwissenschaften, Sektion Geographie, Universität München c Landesanstalt für Umweltschutz Baden-Württemberg, Hochwasservorhersagezentrale, Karlsruhe d Landesamt für Umwelt, Wasserwirtschaft und Gewerbeaufsicht, Rheinland-Pfalz, Mainz
a
ABSTRACT
Methods to accurately assess and forecast flood discharge are a fundamental requirement in practical hydrology. However, existing rainfall-runoff models, seldom consider the spatial characterisation of the land surface, which is essential for an accurate description of processes relevant for runoff formation. Especially land surface conditions of high temporal variability, like soil moisture and snow properties, influence the extent of a flood event and are hence prerequisite boundary conditions and state variables prior to and during the event. The project InFerno (Integration of remote sensing data in operational water balance and flood prediction modelling), funded by the DLR (No. 50EE0053), has been established in 2001 to demonstrate the potential and applicability of remote sensing data (from optical and microwave imagery) for improving operational flood forecasting by retrieving information for model parameterization and data assimilation. New operational remote sensing sensors provide data in a temporal resolution of 3 days (and better), which is a premise for operational applications in mesoscale catchments. Existing algorithms for the processing of remote sensing data, based on former satellite data are currently adapted and extended for the use of ENVISAT. The focus has been given to the quantitative retrieval of near surface soil moisture, spatial snow distribution and snow properties, all being major flood supporting factors in the test areas of the Neckar (~14.000km²) and the Mosel (~28.000km²) river basins. The project consortio comprises a collaboration between science, a value–adding company and end users to provide sustainable and transferable techniques settled close to the actual needs of flood warning services. The common effort of the project partners is to investigate, how these products can be suitably assimilated in the Large Area Runoff Simulation Model (LARSIM), which operates on a 1 km² resolution. Algorithms and methods for a multiscale retrieval of snow covered area and its liquid water content from optical (NOAA-AVHRR) and microwave (ERS, RADARSAT, ENVISAT-ASAR) image interpretation and advanced rationing techniques are shown. The water content of the top-soil layer, being a crucial indicator for flood risk, is determined from a semi-empirical model, which compensates topographical and vegetation effects (biomass, incidence angle) from the radar backscatter signal. The model is applied on ERS and ASAR-WSM imagery. Results of the concept realization, the imagery processing chain, the parameter retrieval, the scaling techniques and the integration of products are demonstrated for the Neckar and Mosel basins in Baden-Württemberg and Rheinland-Pfalz. Keywords: water balance modelling, data assimilation, soil moisture, snow cover, ENVISAT ASAR, NOAA-AVHRR
1. INTRODUCTION
Methods to accurately assess and forecast flood discharge are mandatory to reduce the impact of hydrological hazards. However, existing rainfall-runoff models rarely consider the spatial characteristics of the watershed accurately, which is essential for a suitable and physically-based description of processes relevant for runoff formation. Land surface parameters of high temporal variability, like soil moisture and snow properties are as yet hardly available and used in operational forecasts. In case of the Neckar catchment, rainfall in combination with snow melt is the most critical situation for flood generation. The snow pack in the low mountain ranges of the Neckar is rather thin and inconsistent. The snow water equivalent (SWE) is highly variable and can, as well as the extension of the snow cover, change rapidly. Due to frequent temperature alteration, the snow pack accumulates and melts off several times during a winter season. With regard to possible flood events, small snow packs with a depth from about 20 cm and an SWE up to 50mm are of special interest, because they can melt off within a storm and the SWE will be added to the amount of effective precipitation and hence intensifies the flood. Due to enduring difficulties in the spatial and temporal determination of the snow water equivalent, flood forecasts remain a difficult task under snowmelt conditions. How remote sensing observations allow the reduction of these difficulties is analysed and described here. The concept is that remote sensing methods improve flood forecasting by providing information on the actual extent and properties of the snow cover in the watershed. Remote sensing shall facilitate the regionalisation of hydrological models due to synoptic coverage and homogeneous data provision.
2. OPERATIONAL FLOOD FORECAST AND THE INFERNO – PROJECT
The flood forecast centre HVZ of the LFU (Environmental Protection Agency of the State of BadenWuerttemberg, Germany) is responsible for flood monitoring along most rivers in Baden-Württemberg, among them the Neckar, which covers about half of the State (approx. 14.000 km²). During a flood event, runoff forecasts are calculated and provided every hour for up to 35 gauges in Baden-Württemberg. The results are presented to the public via different media like phone, fax, internet etc. [1]. Forecast times range from 8 to 48 hours [2]. In 2001, the HVZ started the project “InFerno” (“Integration of remote sensing data in operational water balance and flood forecast modelling”, funded by the German Aerospace Centre DLR) [3]. It has been established to test the potential and applicability of remote sensing data for improving operational flood forecasting and water balance modelling by retrieving information from remote sensing for model parameterisation and data assimilation [4]. The focus has been given to the quantitative retrieval of soil moisture, spatial snow distribution and snow properties to improve the simulation of the snow cover and snowmelt (step 1) and the discharge forecast by using soil moisture data (step 2). The project consortium consists of four partners. The flood forecasting centre HVZ and the LUWG define the required remote sensing products from an operational hydrologist perspective. The scientific research, which is needed to develop algorithms for parameter retrieval, is conducted by the Department of Earth and Environmental Sciences at the University of Munich. The value adding company VISTA serves as the integrating link between research and hydrologists through the transcription and extension of the research results into operational products and tools. The HVZ and the LUWG is further analysing the modifications that have to be
Fig. 1: Provided remote sensing products and assimilation concept for the hydrological model LARSIM.
performed to suitably consider this new spatial information source from remote sensing in the hydrological model LARSIM, described in section 5. An overview of the remote sensing products and derived information is given in Fig. 1. The operational snow cover maps are derived from NOAA-AVHRR data, the extent of wet snow and the soil moisture products are based on ASAR-WSM data.
3. REMOTE SENSING OF SNOW
The following part of this paper describes the retrieval and assimilation of snow properties from remote sensing data with a focus on ENVISAT data. Besides snow properties also soil moisture maps of the soil surface are provided to the HVZ. These are calculated in combination with the snow products. Details on the soil moisture retrieval are described shortly in section 4. 3.1 DETECTING SNOW COVERED AREAS (SCA) Snow has typical spectral characteristics that can be used to automatically derive snow cover maps from optical remote sensing data, if spectral information from the visible to the shortwave infrared is available. Automatic procedures use the fact that in contrast to clouds, snow covered surfaces show a low reflectivity in the shortwave infrared section of the electromagnetic spectrum, while both surfaces have exceptionally high reflectances in the visible. NOAA-AVHRR data is used as an optical data source for the SCA classification. The classification of each pixel is conducted by applying empirically determined thresholds to a spectral index image, resulting in a classification distinguishing snow, clouds and snow free areas. For the use in the hydrological model, a further information layer was derived from this classification, the snowline. The snowline is defined as the first snow-free pixel adjacent to the snow cover. In 2002, an operational processing chain was completed and implemented to automatically process snow cover maps from NOAA-AVHRR data for Southern Germany. It processes the satellite imagery (e.g. calibration, navigation, geocoding) and performs snow mapping with less then one hour after data reception [4]. Figure 2 shows an example of a SCA map of the Neckar watershed.
Fig. 2: NOAA-AVHRR image (band 1 and 2) from February 19th 2003 (left); Classified Snow Covered Area using the index method of Derrien (right) with green=snow free, white=snow covered, yellow=snow line, grey=clouds. The boundary of the Neckar catchment and selected snow stations where snow water equivalent is measured 3 times a week are illustrated.
3.2 WET SNOW DISTRIBUTION The use of optical remote sensing data is limited to cloud free areas, which restricts its usability in operational terms. To bridge this gap, InFerno follows a synergetic multisensoral approach, taking advantage of the fact that wet snow can be detected from weather-independent Synthetic Aperture Radar (SAR) data. Dry snow is largely transparent for C-band SAR data (5.3 GHz, as used onboard ERS, RADARSAT and ASAR), with the backscattered signal being significantly dominated by the underlying soil properties. While dry snow has a penetrability of several meters, a liquid water content of 3 Vol. % reduces the penetration depth to a few centimeters. The strong absorption of the signal in the snow layer obviates the signal from the underlying soil, leading to a significant backscatter reduction of several dB. Nagler and Rott [5], [6] have made use of this large backscatter reduction for a methodology to map areas with wet snow cover and hence display regions of increased flood risk. The method was successfully adapted to classify wet snow areas in the Neckar watershed. While most studies on wet snow estimation were conducted in alpine areas, the project area of InFerno lies in the low mountainous area along the Neckar, where the small scale heterogeneity of land use shows explicit backscatter variability due to the respective differences in surface roughness. In order to quantitatively improve classification results, these differences are compensated applying land use dependent thresholds. Since ENVISAT ASAR in wide swath mode can provide images with high temporal frequency of about 3 days, it is in principle possible to derive this information on wet snow distribution twice or three times a week. This temporal observation frequency is comparable to operationally available field measurements of snow water equivalent, that are provided to the HVZ by the German Weather Service 3 times a week from up to 70 stations in the Neckar catchment. However, different user requests of ASAR acquisitions reduce the number of observations in wide swath mode. This conflict of various ASAR modes certainly aggravate the operational use of ENVISAT ASAR for hydrology.
Fig. 3: Example of wet snow and wet soil detection with ENVISAT microwave satellite imagery (lower, left) in comparison with a snow classification 2 days before the SAR acquisition (upper, left). The overlay of the melting snow pixels on the NOAA image (right image) illustrates the multi-sensoral data fusion.
3.3 MULTISENSORAL DATA FUSION The method of multisensoral data fusion is illustrated in Figure 3. The snow classification in the upper left is based on optical imagery from 27th Feb 2003. Large areas in Southern Germany are snow covered. Two days later an ASAR-WSM acquisition allows the determination of regions with melting snow. The resulting classification is shown in the lower left of Figure 3. Wet snow is classified in the Black Forest and the Suebian Alb. In the lower regions along the Danube valley the snow has already melted completely. These regions show up in an orange tone illustrating very wet soil. These soils are saturated with water from the snow melt and can not store any more rainfall. Thus, these regions are potentially critical for flood formation. The comparison or superimposing of the SAR classification with the – two days older – optical snow classification in Figure 3 demonstrates the reliability of the wet snow detection: all distinguished pixels are located inside the snow covered area mapped by AVHRR imagery.
4. SOIL MOISTURE
To derive surface soil moisture information from ENVISAT ASAR data, a method formerly developed for ERS data is used. It is based on an semi-empirical compensation of different terms affecting the radar (C-Band) backscatter. The soil moisture can be calculated, with help of additional soil texture and land use information, by the relationship between measured backscatter coefficient and the dielectric constant. This method is also applicable for WideSwath (WSM) datasets, if the incidence angle is corrected [7]. The retrieved soil moisture maps have a high level of accuracy in comparison to field measurements [9]. Soil moisture measurements achieved from ASAR Data showed an rms error of less then 6 Vol. %. Experiences of field campaigns proved an inner field variation of up to 5%. For an operational derivation of spatial soil moisture observations, an automatic processing of ASAR WSM was developed [9]. With available NearReal-Time datasets, soil moisture maps can be provided within 3-6 hours after satellite overflight. An example of near-real-time derived soil moisture mapping is shown in figure 4.
Fig. 4: Example of a soil moisture map of Southern Germany (24. April 2004), including Neckar, Upper Rhine and larger Parts of the Upper Danube catchment. Soil Moisture retrieval is not possible for forested areas, settlements and water bodies. (assigned as ‘no data’)
5. THE LARGE AREA RUNOFF SIMULATION MODEL LARSIM
Among other models, the HVZ applies the water balance model LARSIM (Large Area Runoff Simulation Model) to calculate flood forecastings for several gauges in Baden-Württemberg. LARSIM enables continuous spatially distributed process simulations of the water balance terms for mesoscale catchments [10]. LARSIM requires detailed spatial information about the simulated catchment area like land use (classified from satellite imagery), topography (DEM), vectorized river network and field capacities of soils. A method for the integration of satellite imagery and ground station measurements was developed and incorporated in the snow module of LARSIM. The method uses the measured snow water equivalent SWE at the ground stations and the snow line from the snow-cloud-classification to calibrate the parameters of the snow module – primarily the threshold temperature Tcrit, which is assigned to distinguish rain from snowfall. The method operates automatically.
6. RESULTS
The results of the calculation considering the snow classification for the calibration of the snow module in LARSIM are shown in Figure 5. The illustration combines temporal courses of modelled versus measured SWE with spatial illustrations of SWE distributions. Combining the ground measurements with the snow classification reduces the amount of snow in some regions compared to the LARSIM model result without data assimilation. In general, data assimilation improved the simulation results of SWE. Just before the beginning of the snowmelt – and this is the most sensitive point in time for flood forecasting – the optimisation, including satellite information, delivered the best results. But during the investigation period, no flood was noticeable affected by snow melt and the verification of the integration of satellite imagery is not completed until now. The presented method for the integration of remote sensing data into the operational model LARSIM was developed using the ground measurements of snow and the location of the snow line, mostly taken from optical imagery. The most striking disadvantage of optical imagery is its weather dependency, whereby no satellite data is available during intense storms. Microwave satellite imagery delivers further information about the actual state of the snow pack, but until now, it can provide reliable information about the SCA or the snow line only under special conditions., e.g. SAR sensors can only see wet/melting snow. This is a disadvantage when the complete snow cover including dry snow shall be retrieved. However, they have the advantage that they can identify the snow cover when it is especially critical for flooding. When rainfall falls on a melting snow cover, snow melt and precipitation sum up and the risk for a flood increases.
7. CONCLUSION
The project InFerno was launched to evaluate the capabilities of remote sensing with respect to operational flood forecasting and management. In this sense, the project setup is breaking new ground in several ways, i.e. by operationally assimilating remotely sensed data and retrieved information into existing information systems and hydrologic models. The major challenge of the project consisted in the implementation of algorithms developed for a multisensoral synergy and the creation of robust, operationally applicable remote sensing products. The assimilation of remote sensing products in an operational hydrological model like LARSIM was initiated and showed good first results. This task is however not solved yet. Especially the incorporation of the information on the melting of the snow cover that SAR can provide is still a matter of research. During the next winter season 2004/2005, when all remote sensing snow products will be provided to the HVZ in an operational way within hours after data acquisition, the added value that can be gained from EO will be studied in more detail.
Fig. 5: Measured SWE (black dots) and calculated SWE at two example stations in the Neckar catchment. The red line is the SWE calculated with LARSIM without optimization. The green line shows the simulated SWE using the snow classification from 19th Feb 2003 and SWE measurements for optimization. The spatial simulation results of SWE for the stations are compared on the left side.
ACKNOWLEDGMENTS
Funding of the project InFerno (project no. 50EE0053) by the German Aerospace Center (DLR) is gratefully acknowledged. The European Space Agency (ESA) supported this work through its PI program (ENVISAT AO 473 & AO 1311).
REFERENCES
1. Schulz, W., Hochwasservorhersagesysteme in Baden-Württemberg für die obere Donau. XX Konferenz der Donauländer, Bratislavia, Slovakia, 4-8. September 2000 (CD-ROM), 2000 2. Homagk, P., Paper, held at workshop of the German Weather Service and the Czech Weather Service in Prague, May, 1998. 3. Schulz, W, Merkel U., Bach H., Appel F., Ludwig R., Löw A., Mauser W., Inferno – Integration of remote sensing data in operational water balance and flood prediction modelling, Proceedings of the International Conference on Flood Estimation. Berne 6-8 March 2002, Switzerland, CHR Report no. II-17, pp.659-668, 2002. 4. Appel, F., Bach, H.,: Near-real-time derivation of snow cover maps for hydrological modelling using operational remote sensing data, IGARSS 03, IEEE 2003 International Geoscience and Remote Sensing Symposium Proceedings, CD-Publication, 2003 5. Nagler, T. et al., SAR based Snow Cover Retrievals for Runoff Modelling. 2nd Workshop on Retrieval of Bioand Geophysical Parameters from SAR Data for Land Applications. CD-ROM, 1998. 6. Rott H. et al., Hydalp - Hydrology of Alpine and High Latitude Basins: Final Report, 2000. 7. Löw, A., Waske, B., Ludwig, R., Mauser, W. (2004): Derivation of hydrological parameters from ENVISAT ASAR wide swath data. Proc. IEEE Int. Geos RS Symposium, September 2004, Anchorage, USA, accepted paper, 2004 8. Mauser, W. , Ludwig, R., Löw, A., “Multiscale soil moisture retrieval to monitor the global change of the water cycle – perspectives form the GLOWA-DANUBE project, Proc. ENVISAT Symposium, 6.-10.Sept.2004, Salzburg, 2004 9. Appel, F., Bach, H., Demuth, N, Loew, A., Ludwig, R., Mauser W., Waske, B., Methodology for the processing of ASAR-wide swath data for the derivation of land surface properties of the Mosel catchment, Proc. ENVISAT Symposium, Salzburg, 2004 10. Bremicker, M., Das Wasserhaushaltsmodell LARSIM - Modellgrundlagen und Anwendungs-beispiele. Freiburger Schriften zur Hydrologie, Institut für Hydrologie, University of Freiburg, Band 11, 2000.