STORM FOREST DAMAGE MAPPING BASED ON VHR InSAR DATA C. Maire* and H. Yésou

Service Régional de Traitement d'Image et de Télédétection (SERTIT), bv Sébastien Brant, 67 400 Illkirch France Ph; +33 (0) 3 90 24 46 43, Fax +33 (0) 3 90 24 46 46, herve@sertit.u-strasbg.fr , http://sertit.u-strasbg.fr

*Presently at DLR, German Aerospace Center, Oberpfaffenhofen, 82234 Wessling , Germany,

Ph +49(0)8153283302, fax, +49 (0)81 53 281 444, cyrille.maire@dlr.de ABSTRACT Remote sensing can be a good, effective and economic method to realise damage mapping consecutive to an extreme storm such as the one in December 1999 which severely hit Western Europe forests. Many recent studies tend to highlight this geomatic solution through high resolution optical and radar data. Here, the question asked is if higher resolution data give more efficient results. The French National Space Agency, CNES, has contracted SERTIT to evaluate both simulated data of the SPOT 5 satellite and VHR InSAR data, which are considered as a prototype for the future SAR sensor generation (2006-2010). Simulated data have been taken during airborne imaging campaigns in 2000 winter and spring periods, over the area of study covering the Haguenau Forest test site. This paper focuses on the assessment of Very High Resolution images (VHR), INSAR data. The results obtained confirm the interest of such data for damage detection in forestry areas. KEYWORDS: Storm, forest, damage mapping, InSAR, VHR, DEM, results dissemination 1. INTRODUCTION The devastating storms which struck continental Western Europe in December 1999 destroyed the equivalent of several years of normal forest exploitation in France. In this context it was of primordial importance to map the damage for short-term actions: the location of the damage, the assessment of risk of fires in the damaged areas, as well as for long term reforestation. 1.1. State of the art

Several ways were used to perform damage inventory. The French National Forest Office (ONF) picked up damage extent and intensity in checking forest parcels. It constituted a fastidious and expensive work, done only for the most intensively exploited areas and many localised damages where small extensions were dressed. A national photographic campaign was in the same time elaborated by the Forestry National Inventory Office (IFN) at the scale of 1:30000. It was still a very expensive solution, hard to realise, because of a manual interaction to delimit the affected areas. Another limit of this solution was the time required to produce the cartography at the scale of the country. Therefore, Remote sensing data appear to be useful for these types of applications, thanks to SAR sensors with their “all-time-weather” capability and (semi-) automatic methods to extract the damaged areas (higher probability to detect changes within forested areas). If applications based on the EO data exploitation are more and more common, there are still few references concerning damage mapping prior to the 99’ events, based on Landsat TM data [1] or more precisely, in our case, on SIR-C/X-SAR data [2]. Recent studies/applications had emphasised the SAR data capabilities for change detection in forest area, analysing the dynamic coherence between two tandem pair acquired before and after the storm [3, 4]. The results carried were promising; nevertheless, with the loss of ERS 1 in March 2000, this approach is no longer to be considered. It will be possible again on a medium-term with very high resolution systems such as Radarsat 2/3 (2006) or the Cosmo Skymed constellation (2005). In this context, an evaluation of VHR data for such applications was planed by CNES to evaluate both simulated data of the SPOT 5 satellite and VHR InSAR airborne data [5, 6, 7, 8].

____________________________________________________________

Proc. of FRINGE 2003 Workshop, Frascati, Italy,1 – 5 December 2003 (ESA SP-550, June 2004) 92_maire

1.2. Test site

The study has been carried out over the Alsace region in NE France, and more precisely over the Bas-Rhin department which has been one of the areas most affected by storm. Test and validation phases have been carried out mainly over the Haguenau. forest. Located 30 km North of Strasbourg, it constitutes the second largest plain forest in France and is covered by 20 000 ha of deciduous and coniferous stands. Over this site, a rich database has been made available containing classical forester’s information such as stand species, exploitation information, soil, etc [9, 10]. A damage map per sub-parcel realised by the French Forest National Office (ONF) shortly after the event was also available. This ONF forestry database has been used for EO data exploitation as well as for results analysis and validation. 1.3. Approaches tested

Two approaches were elaborated to extract damaged areas from the InSAR datas: Manual extraction from the intensity InSAR image, Diachronic DEMs differences.

Regarding the previous method, the hypothesis was the following. We have two accurate DEM, which contain the surface of the canopy (X band), acquired before and after the storm. The difference analysis should be corresponding to a change detection method, since the damages are directly linked to a difference of elevation between the two acquisitions. 2. DATABASE ELABORATION 2.1. Used InSAR VHR data

An advantage of interferometric systems is their capacity to generate topographic or altitude information with a metric relative accuracy. Up to now, airborne data are used to simulate (at least partially) future very high resolution space systems. The short revisit period of these acquisition systems is a condition to preserve an important temporal coherence, and consequently very accurate DEM, even over forest areas [11, 12,13]. When derived from airborne InSAR data, the related accuracy of generated DEM’s is much better being infra metric than centimetric [14, 15]. These airborne sensors allow a simulation of the future space systems containing X band with metric to multi-metric resolution such as Cosmo Skymed, or Terra SAR. . Another source of VHR DEM with vertical metric accuracy could be the interferometric cartwheel [16]. Within the project framework, it has been possible to test the operational potential of such systems on forest damage mapping and forest exploitation monitoring. VHR InSAR data, X band, with sub-metric resolution, had been acquired in May 2000 over the Haguenau forest using the airborne AeS-1 system. Aerosensing GmbH [Intermap] generated derived products such as coherence, intensity images and a VHR DEM. 2.2. Geometric corrections

Those products were in slant range geometry. Their exploitation for our purpose requires making the conversion in ground range geometry. Nevertheless, since the acquisition parameters were unknown, only a very accurate network of terrain references allow us to perform the transformation from the slant range to the ground range geometry. Therefore, an important GPS acquisition had been done using a differential static method. The principle of this method is to use two GPS receptors simultaneously. While one receptor is positioned on a known reference point, the other receptor is used to generate a network of reference points among the studied area. Of course, as the data were already acquired, the reference points had to be clearly visible in the SAR intensity image and well distributed to reduce the co-registration errors. Indeed, instead of measuring an absolute positioning from the GPS satellite constellation, a relative positioning, between the two receptors is realised, eliminating atmospheric perturbations The accuracy of this system of two receptors increases both planimetric and altimetric accuracy from 5-10 meters (with one receptor) to several centimeters. This method was the only one providing good corrections of the data according to their very high resolution. Moreover, it was determinant to retrieve a homogenous accuracy of the elevation of the DEM, since our approach exploits directly the canopy elevation to quantify the damages.

2.3. Pre-Storm DEM simulation

The approach tested requires two elevation information in order to extract the affected area; one before the storm and the other one after the storm. Since we only had access to the InSAR derived DEM (Fig. 2), which represents the forest after the storm, we simulated the forest elevation before the storm. Ground elevation and canopy height are required to build the elevation of the canopy. For the first component, a stereo DTM generated from a SPOT 4 stereoscopic pair was used. The canopy height had been simulated based on grown plant models of the Haguenau forest and the ONF database, which delimits the forest in mono-cultural parcels. The combination of the parcel position, its tree population, the age of the tree and the model related allowed us to generate an average elevation for each parcel. Then the superposition of the parcel elevation and the DTM constitutes a simulated DEM and represents the forest before the storm (Fig. 2). 2.4. Stereo DTM enhancement

Despite the low relief variation of the study site, the stereo DTM had to be enhanced in order to limit the errors of elevation induced by the terrain approximation, regarding the resolution and the accuracy of the InSAR DEM.

Indeed, in our method, only the variations of the canopy height are determinant, since fortunately the elevation of the terrain remained the same before and after the storm. As a result, both DTM enhancement and InSAR DEM corrections (presented in the previous part) were done in one step, in order to avoid as much as possible the errors induced by the terrain elevation approximation. Thus; the aim is to extract directly the canopy height of the InSAR DEM and compare it directly to the one simulated. Fig 1 presents a summary of the geometry correction of the InSAR DEM as well as simultaneously the principle of damage extraction method.

Fig.1. Flowchart : inSAR DEM correction / Damage extraction

3. DIACHRONIC DIFFERENCES OF DEMs The altimetric accuracy of the two DEMs was within 3 meters, which is very satisfactory in a forestry environment. So, larger changes observed between the two DEMs would represent logging due to normal exploitation as the last update of the forestry database was done in 98, and windfall due to the December 99 storm.

Fig.2. Perspective view of the Elevation data ( simulated DEM (left)) / (InSAR DEM (right)) overlapped with SPOT 5 VHR panchromatic data

Two approaches of change computation have been tested [17]. The first one corresponds to a simple difference between the two DEMs followed by a threshold (Fig. 3). In this case the chosen threshold and the young stand height are very similar inducing some confusion. These restrictions over the small parcel elevations are mainly due to the model employed to get the canopy surface. To better take into account these limitations, a second approach has been done. It consists in a normalization based on the initial trees height. A threshold is then applied on the ratio according to the tree stand (deciduous/coniferous). The distinction between coniferous and deciduous has been done since the resulting canopy surface modelled depended of the density of the population and therefore of its regularity.

Fig.3. Simple DEM Difference overlapped with InSAR intensity image (difference in meters).

4. DAMAGE MAPPING FROM THE INTENSITY IMAGE The main objective was to demonstrate that a classical photo-interpretation was possible directly from the InSAR intensity image (Fig. 4). In fact, a similar study based on optical VHR images acquired in winter has highlighted the difficulties to extract the damage located in deciduous areas. The response measured over such areas is closed to the one of soil since deciduous trees have lost their leaves. Moreover; change detection methods, i.e. diachronic analyses as the previous approach, need to dispose of data acquired before the studied event. It highlights the archive exploitation and valorization topics according to the interest area (acquisition possibilities/availability storage). Such VHR SAR data appear to be a solution for forestry purpose. The damages are clearly identified and the interpretation of the data is not as difficult as in urban areas for instance. Nevertheless, a minimum knowledge of SAR image is required to fully extract relevant information.

Fig.4. Damage mapping performed by photo-interpretation based on the intensity image

5. ANALYSIS Damage maps produced using the presented methods have been compared with the reference data of French Forestry Service. Quantitative and qualitative criterions were computed at two levels; internal and externals validation. The first level consisted in a relative evaluation of the methods elaborated whereas the second one is an absolute validation using the ONF damage map at the parcel level.

GD: Good Detection, FD: False Detection, ND: No Detection

Qlty rate = GD FD rate = GD

FD TxND =

GD

ND

Distinction deciduous/coniferous Approach Qlty rate FD rate ND rate Simple DEM difference 0.79 0.12 0.15

Normalised DEM difference 0.84 0.10 0.08 PIAO InSAR Intensity image 0.68 0.09 0.49

Table. 1. Quantitative estimators obtained

The calculated quality parameters of this mapping are very good, with a detection rate of 84% associated with low rates of “false detection” and of “no detection”, respectively 10% and 8% (Table. 1.). Results carried by the photo-interpretation were limited by not detecting small extension damages spread around the area.

6. CONCLUSION

A complete processing chain of both radar and optical data for forest damage assessment was elaborated. It could be used as input in a fire risk decision making system as well as for the design and implementation of a tool for information dissemination through the internet. At present, due to financial considerations, this application has remained at the demonstration level despite the great interest of the user communities. The obtained results based on the INSAR VHR data are also very interesting. Concerning future sensors it is shown that forest monitoring at a local or regional scale should be a possible application for future INSAR systems such as Skymed Cosmo or the Interferometric Cartwheel. These results emphasise the potential of Earth observation data, optical images and SAR coherence products for operational forest damage mapping as well as the potential of future sensors. All the communities involved in forest management from State services to the private owner associations and the Civil Defence have been very interested by the potential shown here. 7. ACKNOWLEDGEMENTS

This work has been carried out within the framework of the Adema-CNES project 762/2/00/0057 on the behalf of the QTIS department of the. CNES. The authors would like to thank the Alsace forestry community particularly P. GELDRIECH of the Haguenau ONF Station for their support, interest and discussion all along the project. 8. REFERENCES

1. Mukai Y. & Hasegawa I., 1999. Extraction of damaged areas of windfall trees by typhoons using LANDSAT TM data, Int. J. Remote Sensing, Vol. 21: 647-654. 2. Winter R. & Keil M., 1995. Information extraction from shuttle Radar image for forest and agricultural applications. JPL technical report. 3. Herrmann A., Fellah K., de Fraipont P. & Yésou H., 2000. Forest damage mapping using SAR coherence product in North Eastern part of France, Earth Observation Quaterly 65, 30-31 4. Yésou H., Weber C., Herrmann A., Fellah K., Bally Ph., Béquignon J., & de Fraipont P., 2000 - Mapping storm forest damage using SAR cohrence data. The case of the Hauguenau forest Looking at our Earth for the New Millennium. Gothenburg, 16_20 October 2000.ESA SP 421 5. Meyer C., Geldreich P., Yésou H., 2001. Apport des données simulées SPOT 5 pour l’évaluation des dégâts de tempête dans la forêt. Cas de la forêt de Haguenau (Alsace, France). Bull Société Française Photogrammétrie Télédétection. Vol 164 -165: 162-173 6. Yésou H., Clandillon S., Meyer C., Allenbach B., Maire C., de Fraipont P., Favard J.C. & Béquignon J., 2002 : Operational mapping of storm forest damage using ERS coherence product and SPOT optical with a result dissemination through an internet gateway and assessment of future VHR capabilities. Proceedings of "Goeinformation for European wide integration”, 22 Earsel Symposium, Prague, 4-6 June 2002 7. Clandillon S., Yésou H., Meyer C., de Boissezon H. & Favard J.C., 2003: Benefits of SPOT 5 HR and VHR data for forest management and windfall damage mapping. IGARSS Proceedings, 2003 Toulouse 8. Proisy C., Sarti F., Mougin E., Lopes A., Dufrêne E., Le Dantec V., Ruiz C., Borderies P. & Chênerie I., 1999. Temporal variations of interferometric coherence over deciduous forest. Proceedings of CEOS SAR Workshop, 26-29 October 1999, Toulouse, France. 9. Geldreich P., 1994. Le nouvel aménagement de la forêt indivise, Haguenau infos, N°21, 3, 7-18. 10. Geldreich P., 1999. La forêt indivise de Haguenau. Arborescences, N°78, 1999, 12-34. 11. Askne J. and Hagberg J., 1993. Potential of Interferometric SAR for classification of land surfaces, Proc. IGARSS’93, vol. 93CH3294-6 Tokyo, Japan, pp. 985-987 12. Damnert P.B.G., 1996. Accuracy of InSAR measurements in forested areas. Proceedings of Symposium Fringe. ESA SP 406: 37-49. 13. Dwyer E., Pasquali P., Holecz F. & Arino O., 2000. Mapping damage cause by the 1999 Lothar storm in Jura (France) using SAR interferometry. Earth Observation Quarterly, 65, 28-29. 14. Holecz F., Pasquali P., Moreira J., Meier E. & Nuesch D., 1998. Automatic Generation and Quality Assessment of Digital Surface Models generated from AeS-1 InSAR data, Procedings of European Conference on Synthetic Aperture Radar , Freidrischshafen, Germany. 15. Wimmer Ch., Siegmund R., Schwabisch M. & Moreira J., 2000. Generation of High precision DEMs of the Wadden Sea with Airborne interferomtric SAR, IEEE, trans. Geosc. Remote Sensing, 38. 16. Massonet D., 2001. The interferometric cartwheel: a constellation of passive satellite to produce radar images to be coherently combined. Int; J. Remote Sensing, Vol. 22, N°10:30. 17. Maire C., 2002. Evaluation de l’apport de données optique et radar de Très Haute Résolution (THR) pour la cartographie de dégâts de tempête en milieu forestier, Bull XYZ n°93, 28-35.