15arspc Submission 91

8/8/2019 15arspc Submission 91

http://slidepdf.com/reader/full/15arspc-submission-91 1/11

1

EXAMINING THE APPLICATION OF A PHYSICS-BASEDSHALLOW WATER BATHYMETRY MAPPING METHODOLOGY

TO ALOS AVNIR-2 MULTI-SPECTRAL DATA IN THE GREATBARRIER REEF

Sagar, S and Wettle, M

Geoscience AustraliaGPO Box 378 Canberra ACT 2601

02 [email protected]

Abstract

Mapping the variable bathymetry of a large aquatic area such as the GreatBarrier Reef (GBR) presents a number of challenges, both in terms of cost and

practicality. Traditional ship-based sonar methods are restricted to depthsexceeding ~15m and the cost of conducting such surveys can be prohibitive.Laser Airborne Depth Sounding (LADS) is often used as a high-resolutionbathymetry mapping tool, particularly in shallow water (0 – 20m). However, theacquisition of this data is costly, and not suited to the widely dispersed shallowwater reef areas that characterise the GBR.

The use of satellite remote sensing data, such as that from the ALOS AdvancedVisible and Near Infrared Radiometer (AVNIR-2) sensor, has the potential to beused as a cost-effective tool to develop broad–scale shallow water bathymetrymapping techniques; it offers a large spatial coverage (70km swath width) and ahigh ground pixel resolution of 10m.

In this work we examine the application of a physics-based bathymetry mappingapproach using multi-spectral AVNIR-2 data over the GBR. The basis of theapproach is an optimisation driven per-pixel estimation of a set of environmentalvariables, including water column depth, from a semi-analytical expression ofsub-surface remote sensing reflectance.

Characteristics of AVNIR-2 data have been found to present a number of issuesthat must be addressed before moving to a broad-scale operational context.These include a limited capacity to apply rigorous sun de-glinting algorithmsand significant quantisation effects on the vertical resolution of the depth

estimates produced, particularly at depths beyond approximately 15m.

Initial results show depths can be derived to approximately 15-20m, with anaccuracy of 10-20% of existing LADS data in the study area. Further work isoutlined to potentially move to a more operational, broad-scale context, and tofurther understand the limitations and possibilities of working with AVNIR-2 forbathymetry applications.



2

I. INTRODUCTION

The Great Barrier Reef (GBR) covers in excess of 340,000 km2 and stretchesfor some 2,000 km along the eastern coast of Queensland, making it the largestcoral reef in the world. The reef provides a significant economic contribution atboth a state and commonwealth level, primarily through tourism, whilst also

being a focus of considerable conservation and environmental managementefforts. One of the important fundamental datasets for such applications is waterdepth, or bathymetry.

Mapping the variable bathymetry of such a large aquatic area presents anumber of challenges, both in terms of cost and practicality. Traditional ship-based methods, such as multi-beam sonar, are restricted to depths exceeding~15m and the cost of conducting such surveys can be prohibitive, especially inremote areas. Laser Airborne Depth Sounding (LADS) is often used as a high-resolution bathymetry mapping tool, particularly in shallow water (0 – 20m) tocomplement ship-based survey data. However, the acquisition of this data is

costly, and not suited to the widely dispersed shallow water reef areas thatcharacterize much of the GBR.

Spaceborne remote sensing can be used for shallow water bathymetryestimation, potentially as a cost-effective tool to complement other surveymethods. However, the majority of techniques developed in this field have beenbased around empirical methodologies, which require known measurementsover a range of depths in the study area in order to calibrate the algorithmsbeing used (Bierwirth et al. 1993; Stumpf et al. 2003). This requirement wouldlikely restrict the use of these methodologies in many remote areas of the GBR,preventing a continuous coverage across the extents of the numerous reefsystems.

We propose the use of a physics-based algorithm, which is able to derivebathymetry from the remote sensing signal without the need for knowncalibration depths. A number of studies have utilized this type of approach,typically using hyperspectral data (Goodman et al. 2008; Brando et al. 2008).However, from an operational perspective of covering extended areas of thereef, utilizing hyperspectral data becomes impractical due to a small swathwidth and high associated costs.

Looking forward to broad-scale mapping of the GBR, it is necessary to identifydata options which are cost-effective: a large spatial coverage in combination

with a level of horizontal spatial resolution sufficient to represent fine-scale reefbathymetry features.

We present the initial findings of a pilot study currently being undertaken toassess the suitability of multi-spectral data from the AVNIR-2 (Advanced Visibleand Near Infrared Radiometer Type 2) sensor on board the ALOS (AdvancedLand Observing Satellite) operated by the Japan Aerospace Exploration Agency(JAXA).

The implications of using this data source as a potential foundation for GBR-wide mapping are then discussed, in view of moving to a more operationalimplementation across the region.



3

II. DATA SOURCE

A. Satellite Data

Data from the AVNIR-2 sensor is 8-bit, 4 band multi-spectral (Blue - 400nm,

Green - 560nm, Red- 650nm and NIR - 825nm) with a ground pixel horizontalresolution of 10m. With a swath width of 70km, each scene covers 4,900 km 2 ata nadir acquisition. The scene used for this pilot study was acquired on the 2ndApril 2009 and illustrates the potential this sensor has to cover large tracts ofthe GBR with a single processed scene (Fig. 1); comparative scene footprintsare shown for the commercial Quickbird sensor and the hyperspectral Hyperionsensor.

To enable its use in the physics-based methodology outlined in Section IV, thedata are converted from raw data to reflectance values using a MODTRAN(MODerate resolution atmospheric TRANsmission) based atmospheric

correction procedure with the c-WOMBAT-c interface (Brando & Dekker, 2003).B. Model input parameters

The optimization process in this project, outlined further in Section IV, uses alibrary of benthic substrate reflectance spectra to determine, on a pixel-by-pixelbasis, the ratio composition of any two substrates (e.g. 70% sand 30% coral).The spectral library used in this project consisted of 3 common benthicsubstrates; Sand, Massive Coral and Algae. These reflectance spectra weresourced from the Centre for Remote Sensing and Spatial Information Science ofthe University of Queensland.

In addition, a range of parameters associated with the optical properties of the

water column constituents are required as an input to the model, describing theabsorptive and backscattering components of the water. The optical waterparameters used in this work were provided by the Environmental EarthObservation Group of CSIRO Land and Water.

In an optimal modelling environment, these spectra and optical water propertieswould be determined through site-specific fieldwork. This scenario is notrealistic for broad-scale areas across the entire GBR. Therefore, a componentof this work is to assess the suitability and success of using generic field dataas model inputs.

III. STUDY SITE

The selected study area is detailed in Fig. 1, and is located approximately 90kmnorth of Bowen off the Queensland coast. The rationale for the selection of thissite was based on three key criteria; 1) the availability of a high quality AVNIR-2scene 2) high density coverage of LADS data, and 3) an area deemedrepresentative of a typical reef environment in the GBR.



4

Figure 1. Location of the study site and ALOS AVNIR-2 scene. The extent of the GreatBarrier Reef can be seen offshore from the Queensland Coast. Indicative scene

footprints from the Quickbird and Hyperion sensors are shown as insets forcomparison.

Whilst known depth values are neither required for the modelling process norlikely to be available for many areas of the GBR, for the purposes of this studyLADS data was deemed necessary to assess the accuracy of the derivedbathymetry model. It is also required to underpin further work in quantifyingquality indicators produced by the modelling algorithm, discussed in Section VI.

IV. PHYSICS-BASED ALGORITHM

This study applies a physics-based bathymetry mapping approach, SAMBUCA(Semi-Analytical Model for Bathymetry, Un-mixing, and ConcentrationAssessment), which was developed in the ENVI/IDL environment as anobjective and repeatable algorithm for extracting aquatic environmentalinformation on a pixel-by-pixel basis from remote sensing data (Brando et al.2008; Wettle & Brando, 2006). It is based on the approach conceptualized byLee et al. (1998, 1999) for retrieving environmental variables from remotesensing data using an analytical model and optimization routine.

The application of a physics-based, radiative transfer method has manyadvantages, including the ability to apply one algorithm to time series data, anddata from a variety of sensors. In the context of the work in this paper, it is theability of this type of method to extract depth, water column constituents andbenthic substrate composition from a single remote sensing measurementwhich is essential. This is due to the anticipated lack of field data (known



5

depths, water properties and benthic substrates) for a significant proportion ofareas of the GBR.

The basis of this approach is the expression of the subsurface remote sensingreflectance as a set of variables, and a parameterization of this model(RSmodelled) to define it as a function of a set of environmental parameters. Anoptimization routine is used to minimize the difference or error between themeasured subsurface remote sensing reflectance (RSmeasured) andRSmodelled by varying the variables and retrieving the environmentalparameters, including water depth, that correspond to the lowest error betweenRSmeasured and RSmodelled (Wettle & Brando, 2006).

V. RESULTS

A. AVNIR-2 Data Assessment

An impeding factor in the use of optical imagery for aquatic applications is thepresence of sun-glint: radiance from the sun reflected directly off the water

surface to the sensor, induced by wave effects and/or the geometry of thesensor and sun location.

The sunglint-contaminated signal contains little information about waterconstituents or benthic features (Kutser et al. 2009). The bathymetry estimationprocess is therefore compromised in each pixel containing sun-glint.

The nadir acquisition nature of the majority of AVNIR-2 data means it issusceptible to sun-glint. This is based on the sun elevation angle at the time ofacquisition, and to a lesser extent, glint produced from wave and wind inducedeffects. Commonly used glint-correction techniques for optical imagery assumea linear relationship across glinted and non-glinted pixels between each of the

visible bands and the near-infrared (NIR), in optically deep water (Hochberg etal. 2003; Hedley et al. 2005). In our initial assessment of AVNIR-2 data, wehave observed this linear relationship to be less defined, preventing a rigorousde-glinting of the imagery. From further examination, we consider this to be dueto the limited radiometric range of the AVNIR-2 data (8-bit) in comparison toother data sources for which this approach has been used. Figure 2 shows thelinear relationship between the NIR and Red bands over a glinted deep waterarea for both Quickbird and ALOS AVNIR-2 data. The linear relationship is welldefined in the 11 bit (0-2048) Quickbird data, but almost non-existent in thelimited range of the ALOS 8-bit (0-255) data.

Figure 2. Comparison of the linear relationship of the NIR and Red bands over glinteddeep water for Quickbird 11-bit data (left) and ALOS AVNIR-2 8-bit data (right)



6

For this reason, we prefer to look at options for the mitigation of glint whenlooking at image acquisition on a broader-scale, such as detailed in Section VI.

Image masking is required for pixels containing features where the bathymetryalgorithm is unable to derive reliable depths, such as white-caps, breakingwater, exposed reef and cloud cover. A simple NIR threshold approach wasfound to be the most time-efficient and reliable methodology, which areimportant criteria for extending mapping coverage to numerous scenes.

B. Validation and Accuracy Assessment

Extensive LADS coverage of the study area, ranging from 0.1m to greater then40m, was used to assess the accuracy of the derived bathymetry model. Aspreviously mentioned, the use of LADS data for validation is twofold, 1) toassess the accuracy of the derived model, and 2) to provide a tool to allowfuture quantification of the error/quality indicators produced by the SAMBUCA

physics-based algorithm.The derived bathymetry model was tide adjusted to Lowest Astronomical Tide(LAT) to correspond to the datum of the LADS data, using the average of thethree surrounding tide gauge readings in the area at the time of imageacquisition.

Prior to validation, the position of specific bathymetric features (reef crest edgesetc) in the study area were assessed in relation to their location defined by theLADS data. A horizontal offset of some 20-30m was observed between the two,which has implications for broader mapping applications (as discussed inSection VI), as well as for valid accuracy assessment.

To align the ALOS AVNIR-2 more closely with the LADS data, an image toimage registration was completed using an existing Quickbird image of thestudy area as a base reference. A bathymetry transect was then completed,comparing the shifted ALOS depth model to the corresponding surfacegenerated from the LADS data using an inverse distance weighted interpolation.(Figure 3)

Figure 3.B shows the original 10m resolution ALOS bathymetry modelcompared to the LADS profile. Significant ‘noise’ is evident in the ALOS data,particularly in deeper sections of the lagoon. It is considered that the limitedradiometric resolution of the ALOS data may contribute these artefacts; wherein

a quantisation effect becomes more pronounced as the changes in reflectanceat deeper depths approach the noise threshold of the data.

In an attempt to deal with these artefacts a 3x3 median filter was applied to theALOS bathymetry model, the results of which are shown in Figures 3.C and3.D. A median filter was chosen to avoid introducing new depth values to themodel as would be done by a smoothing or low-pass filter. Whilst the resultsshow a better model fit to the LADS data, a trade-off between fine featuredefinition and overall model accuracy is made using this approach.



7

Figure 3: GBR Validation Area - bathymetry transect. Original ALOS AVNIR-2 derivedbathymetry (A) and profile comparison to LADS data (B) vs AVNIR-2 median filtered

data (C) and profile comparison to LADS data (D)



8

Figure 4 shows the resulting scatterplot of the AVNIR-2 derived median filtermodel depths against those of the LADS data, with a R2 value of 0.82. Thedistribution of points has been visualised in a colour density slice, and showsgood agreement with depths up to approx 14-15m. Some quantisation is visibleat depth, with clusters of depths derived at increasing intervals beyond approx

10m, most clearly with the four discrete depths values shown at the 15-17mregion.

Statistical analysis shows a RMSE value of 1.17, with 36% of the deriveddepths being within 10% of the LADS validation value, and 63% being within20%.

Figure 4: Scatterplot of AVNIR-2 derived bathymetry model (median filtered) versusLADS data

Error and quality indicators produced as part of the SAMBUCA methodology are

an indication of the success of the modelling and inversion result, giving a pixel-by-pixel confidence estimation of the derived depth. Additional indicators areproduced to highlight areas where an optical depth limit has been reached (i.e.where the sensor cannot ‘see’ the bottom with enough confidence to derive areliable solution). The specifics of these indicators are beyond the scope of thispaper, and we refer to Brando et al. (2008) for further details.

These indicators can be used to mask pixels in the bathymetry model in order toremove low confidence solutions from the final model. Fig. 5 shows thebathymetry model produced for the study area, with pixels masked in whiteusing the indicators described above. The extent of the areas successfully

modelled by the algorithm support the potential for using generic substrate andwater parameters on a broader scale.



9

Figure 5. Shallow water bathymetry model derived from ALOS AVNIR-2 multi-spectraldata over the GBR study area.

VI. IMPLICATIONS FOR FUTURE WORKThe combination of a large spatial coverage per-scene and a high pixelresolution gives ALOS AVNIR-2 data a strong potential as a tool for broad-scaleoperational bathymetry mapping of the GBR. However, there are a number ofdata issues to consider which this pilot project has highlighted.

Given the spatial footprint of an AVNIR-2 scene, coverage of the GBR could beachieved with approximately 25-30 scenes. However, due to issues such asglint, cloud cover and atmospheric effects, a level of data redundancy isrequired to ensure contiguous coverage.

Methods to correct for glint in AVNIR-2 data are yet to be rigorously developed

or tested. We have developed a screening tool which assesses each scene forglint potential using scene metadata and the resulting sun-sensor geometry.Further work will involve possible bulk-processing of scene metadata toevaluate glint potential across the available scenes in the GBR. Examination ofthe JAXA ALOS catalogue (https://auig.eoc.jaxa.jp/auigs/en/top/index.html)shows a good coverage of AVNIR-2 across the GBR, with a good level ofredundancy in many areas of the reef. Assessment of how many of thesescenes may be suitable for bathymetry mapping will prove a valuable step indetermining the potential of moving to a more operational context.

The ability to integrate satellite derived bathymetry with other bathymetry

datasets is essential to ensure the relevance of the satellite bathymetry tostakeholders. Accurate geo-referencing of the AVNIR-2 data, and therefore the



10

bathymetry model, is important not only for the offset and validation issuesdescribed in Section V, but also to ensure seamless integration with data-setssuch as multi-beam sonar and LADS. The lack of control points over aquaticareas, such as the GBR, with which to perform rigorous geo-referencing,remains an issue that must be addressed to coordinate a broad scale approach.

Results of the pilot project have indicated that more work is needed tounderstand the effects of the 8-bit radiometric resolution and the signal to noisecharacteristics of AVNIR-2 data.

As discussed in Section V, we consider that the inability to establish linearrelationships in the de-glinting process may be due to the limited range anddiscrete values defined in both the visible and NIR bands for deep water.Similarly, the clusters of depths derived by AVNIR-2 at depths greater than 15mmay be a function of the limited spectral discrimination given by the 8-bit datarange. Further work will focus on evaluating and testing these possiblequantization effects.

With a lack of validation data such as LADS likely in many areas of the GBR,accuracy of the derived models will only be able to be assessed through theerror/quality indicators output of the physics-based algorithm. As these are notdirectly related to the accuracy of the derived depth, further work will focus onusing the current study area and available LADS data to investigate the validityof adding an element of accuracy quantification to these indicators.

VII. CONCLUDING REMARKS

With the broad spatial coverage and high resolution that data from the ALOS

AVNIR-2 allows, we consider it a viable and cost-effective tool to enableshallow-water bathymetry modelling to complement more traditional techniquesin the extensive areas of the GBR.

Initial results have shown accuracies of 10-20% of the modelled depth to beachievable, dependant on data quality. Refinements to the process currently inprogress to improve these results and quantify the quality indicators producedas part of the physics-based methodology.

We have identified issues that need to be addressed to move towards a moreoperational context, in aiming for a contiguous coverage of the entire GBR. Thisoutput would serve as a fundamental data-set to a variety of stakeholdersinvolved with research, management and industry in this unique environment.

ACKNOWLEDGMENTS

The authors thank the University of Queensland – Centre for Remote Sensingand Spatial Information Science for providing substrate spectra, and the CSIROLand & Water group for providing optical water property parameters and theSAMBUCA source code.



11

REFERENCES

Bierwirth, P.N., Lee, T.J. and Burne, R.V. (1993) Shallow Sea-floor Reflectanceand Water Depth derived by unmixing multispectral imagery. PhotogrammetricEngineering and Remote Sensing, 59(3) 331-338.

Brando, V.E., Anstee, J.M., Wettle, M., Dekker, A.G., M., Phinn, S.R. andRoelfsema, C A (2009) Physics Based Retrieval and Quality Assessment ofBathymetry from Suboptimal Hyperspectral Data, Remote Sensing ofEnvironment 113, pp. 755-770, 10.1016/j.rse.2008.12.003

Brando, V. E. and Dekker, A. G. (2003) Satellite hyperspectral remote sensingfor estimating estuarine and coastal water quality. IEEE Trans. Geosci. RemoteSens., 41:1378–1387

Goodman, J.A, Lee, Z.P. and Ustin, S.L. (2008) Influence of atmospheric andsea-surface corrections on retrieval of bottom depth and reflectance using asemi-analytical model: a case study in Kaneohe Bay, Hawaii, Appl. Opt., 47, F1-

F11Hedley, J. D. , Harborne , A. R. and Mumby, P. J (2005) Simple and robustremoval of sun glint for mapping shallow-water benthos , International Journalof Remote Sensing, 26(10):2107-2120.

Hochberg, E.J., Andrefout, S. and Tyler, M.R. (2003) Sea Surface Correction ofHigh Spatial Resolution Images to Improve Bottom Mapping in Near-ShoreEnvironments. IEEE Trans. Geosci. Remote Sens., 41:1724–1729.

Kutser, T., Vahtmae, E., Praks, J. A (2009) sun glint correction method forhyperspectral imagery containing areas with non-negligible water leaving NIRsignal, Remote Sensing of Environment 113(10), pp. 2267-2274

Lee, Z. P., Carder, K. L., Mobley, C. D., Steward, R. G., and Patch, J. S. (1998)Hyperspectral remote sensing for shallow waters:1. a semianalytical model.Applied Optics, 37(27):6329–6338

Lee, Z., Carder, K. L., Mobley, C. D., Steward, R. G., and Patch, J. F. (1999)Hyperspectral remote sensing for shallow waters:2. deriving bottom depths andwater properties by optimization. Applied Optics, 38(18):3831–3843

Stumpf, R.P., Holderied, K. and Sinclair, M. (2003) Determination of waterdepth with high-resolution satellite imagery over variable bottom depths.Limnology and Oceanography, 48, pp. 547–556

Wettle, M. and Brando, V.E. (2006) SAMBUCA – Semi-Analytical Model forBathymetry, Un-mixing and Concentration Assessment. CSIRO Land and WaterScience Report, 22/06

15arspc Submission 91

Documents

Transcript of 15arspc Submission 91