ADVANCED REMOTE SENSING TECHNIQUES FOR FORESTRY APPLICATIONS:A CASE STUDY IN SARAWAK (MALAYSIA)

1Edmond NEZRY, 1Francis YAKAM-SIMEN, 2Paul ROMEIJN,1Iwan SUPIT, 3Louis DEMARGNE

1 PRIVATEERS N.V., Private Experts in Remote SensingGreat Bay Marina, P.O. Box 190, Philipsburg, Netherlands Antilles

E-mail: 113247.1054@compuserve.com, Internet: http://www.treemail.nl/privateers/

2 TREEMAIL, International Forestry AdvisorsPrins Bernhardlaan 37, BW 6866, Heelsum, The Netherlands. Internet: http://www.treemail.nl

3 SPOT IMAGE5 rue des Satellites, BP 4359, 31030 Toulouse Cedex 4, France. Internet: http://www.spotimage.fr

ABSTRACT

This paper reports the operational implementationof new techniques for the exploitation of remotesensing data (SAR and optical) in the framework offorestry applications.In particular, we present a new technique forstanding timber volume estimation. This techniqueis based on remote sensing knowledge (SAR andoptical synergy) and forestry knowledge (foreststructure models), proved fairly accurate.To illustrate the application of these techniques, anoperational commercial case study regarding forestconcessions in Sarawak is presented. Validation ofthis technique by comparison of the remote sensingresults and the database of the customer has shownthat this technique is fairly accurate.

I. INTRODUCTION

An important issue in today's operational remotesensing is to assess the potentialities of the joint useof multi-sensor satellite imagery to improve thediscrimination between forest and non-forest,between forest types, between crops and clearcuts,and between crop types.In addition, several complementary issues, relativeto terrain topography or to the detection andmonitoring of man-made infrastructures and naturalstructures must also be addressed.For a complete and comprehensive exploitation ofmulti-sensor synergy and complementarity, new

strategies to retrieve information relevant to theseapplications have been developed and evaluated inGeographical Information System (GIS)environment.Such strategies have been developed according torequirements widely expressed by the community ofusers. In order to achieve the general above-mentioned objectives, existing processing andinformation extraction methods specific to SyntheticAperture Radar (SAR) and to Optical imagery havebeen revisited and improved. New additionaltechniques based on control systems principles havebeen developed.Results of a commercial project carried out inSarawak, Malaysia, are shown, to illustrate therange of techniques that can, and very often must,be used in a variety of forestry remote sensingapplications.In this project, the data at disposition were: a SPOT-XS 3-channels optical image, two RADARSATSAR (C-HH band) images taken at differentincidence angles, and an ERS SAR (C-VV band)image.

II. GEOREFERENCING IN GISENVIRONMENT

All data have been georeferenced in UniversalTransverse Mercator (UTM) projection, using Earthellipsoid WGS 1984. The georeferenced remotesensing images have been integrated in a GIS as lowlevel layers.

III. PRE-PROCESSING OF THE REMOTESENSING DATA

Pre-processing of the remote sensing images hasbeen particularly stressed. Indeed, the quality of thispre-processing contributes substantially to theaccuracy of the final thematic products.

a) Optical (SPOT-XS) images:

Atmospheric corrections are carried out using astatistic-empirical method based on both the resultsof a Tasseled Cap analysis, and an atmosphericmodel. The Tasseled Cap transformation belongs tothe family of Principal Component Analysis (PCA)transformations. The difference to the usual PCA isthat the principal axes are not obligatorilyperpendicular.

In the course of this atmospheric correctionprocedure, haze detection is performed, and a hazeimage is produced in order to estimate the opticaldepth of atmospheric aerosols.This haze image, where atmospheric effects areenhanced could also be used for other applications.

b) SAR images (RADARSAT and ERS):

Complete radiometric calibration of the SARimages is carried out, according to the specificationsof the various data providers.

Then speckle noise filtering is performed. To thisend, Bayesian Maximum A Posteriori (MAP)single-channel and vector statistically adaptivespeckle filters recently developed by PRIVATEERSN.V. are used.

Detectors based on the local spatial autocorrelationfunctions of both the speckle and the scene in SARimages are incorporated to these filters, to refine theevaluation of the non-stationary first order localstatistics. This improves both restoration of thescene textural properties and the preservation ofscene structural elements, to produce specklefiltered images without loss in spatial resolution [1].Two filters have been developed, for single SARimages, or for multi-channel SAR images. It can bedemonstrated that these new Bayesian speckle filterspresent the structure of control systems [2].

Thus, their application is regarded as the firstprocessing step of application-oriented controlsystems designed to exploit the synergy of differentSAR sensors, or of SAR and Optical sensors.

IV. TOPOGRAPHY

a) Generation of Digital Elevation Model:

First of all, a SAR stereogram is produced, using thetwo RADARSAT multi-incidence SAR images.From this stereogram, a Digital Elevation Model(DEM) is produced by radargrammetry.In the study shown, the horizontal DEM accuracy isbetter than 15 meters; the vertical DEM accuracy isbetter than 20 meters.

Figure 1: Radargrammetric DEM, built usingRADARSAT data.

b) Ortho-rectification:

In remote sensing images acquired by side-lookingsensors, relief is systematically displaced withregard to its true geographical location, due to theviewing geometry. This effect is particularlyimportant in SAR images. It also affects Opticalimages acquired in off-nadir looking direction.

For both the optical and the SAR images, thecorresponding local geometrical corrections

(geocoding, or in other words ortho-rectification)are performed, using the DEM.

c) Correction of terrain induced effects:

For both the optical and the SAR images, theradiometric corrections for slope angle and aspectangle effects are carried out, using the DEM.These corrections account for the variation ofeffective scattering area with local slope andorientation, as well as for the variations inillumination conditions.Effects of the variations in illumination conditionson the image radiometry are evaluated using aphysical backscattering model (SAR images), or asemi-empirical radiance model (Optical images).

Figure 2: Terrain corrected (geocoded), and geo-referenced SAR images. Radarsat-S7 in red,

Radarsat-S2 in green, ERS in blue).

V. DETECTION ISSUES

Forest managers and timber industry often needaccurate knowledge of the hydrological network(rivers and water bodies) and of the forestcommunication network (roads, and forest and/or

logging tracks). Often, the hydrological network isconsidered as part of the communication network.

Automatic extraction of the hydrological andcommunication network is performed in the remotesensing images by pattern recognition techniquesbased on the analysis of local first order [3], and/orsecond order local image statistics [2].Since, due to the geometrical viewing conditions,complete detection by using either Optical or SARdata alone is often impossible, detection isperformed simultaneously in at least two remotesensing images.Within cloud free areas of the SPOT image,detection is carried out using the SPOT and aRADARSAT image. This way, complementaritybetween Optical and SAR information enables todetect reciprocally missing features and to achievecomplete detection.Within the cloud-covered areas of the SPOT image,the linear network is extracted using the twoRADARSAT images acquired at different incidenceangles to detect reciprocally missing features and toachieve complete detection.In a further step, the extracted network isautomatically vectorized within the GIS. This vectorlayer can then be superimposed to any raster layer(DEM, timber volume map, classification, etc.) for abetter synoptic understanding of areas of interest, orfor further data analysis.

VI. FOREST AND LAND-USE INVENTORY

Previous remote sensing forestry studies inIndonesia had assessed the thematiccomplementarity between SPOT-XS and SAR data[4]. This complementarity is also exploited in thestudy shown.Indeed, within the cloud-free part of the SPOTimage, classification is performed on a datasetformed of mixed SPOT/RADARSAT derivedindices. These indices, obtained through TasseledCap analysis and SPOT/RADARSAT data fusionare designed to select from the dataset the part ofinformation content that is relevant to the currentapplication [cf. 5].ISODATA clustering is applied to these indices.The rationale of clustering algorithms is to identifynatural groupings (clusters) in the selected featurespace through an iterative statistical learningprocess.Within the cloudy part of the SPOT scene,classification is performed using the twoRADARSAT images acquired at different incidenceangles, and the ERS image. The classificationmethod used is the extension to the case of multi-

channel SAR images of the supervised classificationmethod developed by PRIVATEERS N.V. in [6] forthe case of mono-channel SAR images. Thismethod, that uses both the radiometric and thetextural properties of the thematic classes provideshighly accurate results when the thematic classesare appropriately defined.In a final step, these two partial classifications havesimply been merged to produce the finalclassification. As shown, the merging border isalmost invisible and both partial classifications fitsatisfactorily to each other.In this final forest pre-inventory, the followingclasses are identified: very high yield forest, highand medium yield cutover, low yield cutover and thesecondary forest, clearcuts and very low yield areas,bare soils and agricultural surfaces, water bodies.Ground and GIS-based validation shown that resultsare fully satisfactory.

Figure 3: Forest pre-inventory. Classes are: waterbodies (blue), bare soil and agriculture (orange),

clearcuts and low regrowth (gray), low yieldcutover (light green), medium to high yield (medium

green), primary forest and very high yield (darkgreen).

VII. ESTIMATION OF STANDING TIMBERVOLUME

Timber volume (or equivalently woody standingbiomass) estimation is made, using a new methoddeveloped [cf. 7] and validated by PRIVATEERSN.V. and TREEMAIL.This method uses both Optical/SAR data fusiontechnique, and forest structure models [8].Within the cloud-covered areas of the SPOT image,forest pre-inventory classification results andautomatic training in the SAR images based onresults obtained within the cloud free areas of theSPOT image compensate for the lack of Opticaldata.

Figure 4: Standing timber volume estimation(from black to white: 0 to 460 m3/ha).

VIII. VALIDATION OF STANDING TIMBERVOLUME ESTIMATION

Validation of the results has been carried out, bycomparing these remote sensing estimations of thestanding timber volume, with the database of theprofessional foresters of the customer (1287 plots oftimber volume measurements within the study area),

an industrial timber concession established inSarawak. Clearly, the two kinds of estimations hadbeen obtained independently, and by totally different

techniques. The resulting comparison is presentedin the Figure 5 below.

Figure 5: Standing timber volume: Customer ground estimations versus remote sensing estimations.

Figure 5 shows that the objectives fixed by thecustomer in terms of estimation accuracy (less than15% error) are achieved in 87% of the cases.In an additional 6.4% of the cases, the mismatchbetween ground and remote sensing estimationsremains inferior to 30%, which is still better thanwhat is achieved by concurrent techniques presentedin the literature in the case of dense tropical forests.

Figure 5 enables to evaluate the domain of validityof the present technique:- From 0 m3/ha to 200 m3/ha, this technique givesvery satisfactory results. No other techniqueprovides such accuracy in this range of timbervolume.- From 200 m3/ha to 260 m3/ha, this technique stillprovides useful results (error of the order of 15%).- Above 260 m3/ha, estimations become erratic andare systematically under-estimated. Therefore,further improvements (forest model, use of differentfrequencies or remote sensing data) are needed toextend the validity domain of this technique, forstanding timber volume over 260 m3/ha.

XI. CONCLUSION AND PERSPECTIVES FORTHE PREVENTION OF FOREST FIRES

The techniques presented above result fromPRIVATEERS/TREEMAIL research anddevelopment investment initiative.They have already been successfully validated, firstwithin scientific pilot projects granted by the majorSpace Agencies, then in the framework ofcommercial application projects.It is noticeable that these techniques oftenoutperform dramatically previous techniques. Theyopen new perspectives for a wide range of forestryapplications:

With regard to forest fires in El Niño climaticconditions, these (and any other) remote sensingtechniques are useless if applied too late, when giantwild fires already rage.However, among the techniques and applicationsshown, several are of particular relevance, and can

represent a major asset in the framework of a forestfire prevention project:

· Estimation of forest biomass is important toassess the amount of potential vegetal fuel andevaluate fire risks.

· Identification and monitoring of (regrowing)clearcuts and agriculture areas are an importantissue in the assessment of fire risks linked:

- to the change in drainage conditions and patterns, and the formation of both peat soils (e.g. by drying out peat

swamps for agriculture),- to the formation of dense but dryer vegetation cover (clearcut, regrowth).

· Knowledge of terrain topography is mandatoryto assess the difficulties of access to potentiallythreatened areas.

· Up to date identification and mapping of theground communication network (roads andforest tracks) as vector GIS layers ontop of thetopographical maps is needed, in order to enablerapid intervention of fire fighters brigadeswherever a wild fire breaks out.

· Georeferencing and geocoding of thesecartographic products, integrated in a GIS, ismandatory to:

- locate accurately and without any ambiguities the potentially threats in the prevention phase,- assess the difficulties of intervention to burning areas and to manage efficiently the use of fire fighters in critical situations, when a wild fire breaks out.

X. REFERENCES

[1] E. Nezry, M. Leysen and G. De Grandi, 1995:"Speckle and scene spatial statistical estimators forSAR image filtering and texture analysis: Some

applications to agriculture, forestry and point targetsdetection", Proceedings of SPIE, Vol.2584, pp.110-120, September 1995.[2] E. Nezry, F. Zagolski, A. Lopes and F. Yakam-Simen, 1996: "Bayesian filtering of multi-channelSAR images for detection of thin details and SARdata fusion", Proceedings of SPIE, Vol.2958,pp.130-139, September 1996.[3] A. Lopes, E. Nezry, R. Touzi and H. Laur, 1993:"Structure detection and statistical adaptive specklefiltering in SAR images", International Journal ofRemote Sensing, Vol.14, nr.9, pp.1735-1758, June1993.[4] E. Nezry, E. Mougin, A. Lopes and J.P.Gastellu-Etchegorry, 1993: "Tropical vegetationmapping with combined visible and SARspaceborne data", International Journal of RemoteSensing, Vol.14, nr.11, pp.2165-2184, July 1993.[5] E. Nezry, S. Rémondière, G. Solaas and G.Genovese, 1995: "Mapping of next season's cropsduring the winter using ERS SAR", ESA EarthObservation Quarterly (ESA publication), nr.50,pp.1-5, December 1995.[6] E. Nezry, A. Lopes, D. Ducros-Gambart, C.Nezry and J.S. Lee, 1996: "Supervised classificationof K-distributed SAR images of natural targets andprobability of error estimation", IEEE Transactionson Geoscience and Remote Sensing, Vol.34, nr.5,pp.1233-1242, September 1996.[7] G. Kattenborn and E. Nezry, 1996: "Analysis ofan ERS-1 SAR time series and optical satellite datafor forestry applications in temperate zones",International Archives of Photogrammetry andRemote Sensing, Vol.31, Part B7, pp.331-339, July1996.[8] R.A.A. Oldeman, 1990: "Forests: Elements ofSilvology", Chapter 6 "Silvatic mosaics", pp.388-558, 624 p., Springer Verlag, 1990. See also thereferences cited in this chapter.