landsat as an aid in the preparation of …...d. k. warne the australian national university...
TRANSCRIPT
D. K. WARNE
The Australian National UniversityCanberra, A.C.T., 2600, Australia
Landsat as an Aidin the Preparation ofHydrographic Charts
Costs are reduced when Landsat MSS imagery is employed insupport of hydrographic surveys.
INTRODUCTION
MAPPING FROM satellite imagery offersthree main advantages over conven
tional hydrographic survey techniques(Turner and Mitchell, 1977):
(1) Broad coverage is available at comparatively low cost;
(2) Those operations to which the imagery issuited-delineation of reefs and islandsand mapping shallow water-are the most
given to the depth measurement aspect ofhydrographic mapping.
Other studies (Byrne and Honey, 1977;Polcyn, 1976) have demonstrated that, underfavorable conditions, Landsat is able tomeasure water depths up to approximately20 metres within 10 percent (RMS) of themeasured value. Penetration is reduced aswater clarity falls.
The principle of bathymetric mapping
ABSTRACT: Among the many claims made for Landsat MSS imageryis the ability to map water depths in shallow areas. A study has beenconducted at the Australian National University to explore the extent to which this technique could be applied in a production mapping situation. The study, while confirming the technique in principle, demonstrated a number of implementational difficulties. Theimplications of these difficulties are discussed and it is shown thatan integrated approach, involving satellite imagery, computeranalysis, human interpretation, and ground truth collection, canovercome the major problems. Ground truth profiles at a spacing of5 kilometres yield water penetrations to about 20 metres at an effective depth accuracy of 10 percent of measured value. The extent ofsupporting data determines reliability of the final product and canbe varied to provide an appropriate balance between cost andreliability for a particular project.
difficult and frequently dangerous operations for shipbome surveys; and
(3) A single image can bridge expanses offeatureless ocean.
Turner and Mitchell (1977) and Lyons(1977) discuss the problem of updatingcharts of reefs and islands while Fleming(1977) describes the application of Landsatimagery to geometric control across largestretches of water. Here consideration is
with Landsat is straightforward. Neglecting multiple reflections within the body ofthe sea, the optical model is of the form:
R = a + b . exp (-c . z) (1)
where R is the radiance measured at thedetector; z is the water depth; and a, b, andc depend on the prevailing optical characteristics of the sea and overlying atmosphere.
PHOTOGRAMMETRIC ENGINEERING AND REMOTE SENSING,
Vol. 44, No.8, August 1978, pp. 1011-1016.1011
1012 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING, 1978
a represents the signal that would be returned from "infinitely" deep water and isthe sum of the atmospheric path radianceand the radiance reflected hom the sea surface and within the sea water itself. b givesthe effects of attenuation of the light at thesea surface. Both a and b include an atmospheric attenuation term. c is the effectiveattenuation coefficient for light passing toand from the sea bed with appropriate allowance being made for the non-vertical,and possibly indirect, light path.
All coefficients are dependent on thewavelength of the light. Unfortunately, noLandsat detector is optimally placed spech'ally, but the best response is obtained fromthe visible green sensor (band 4) while thered (band 5) may be useful in waters of up toa few metres in depth and for detecting nearsurface anomalies.
In 1975, the Australian National University's Department of Engineering Physics,in liaison with the Division of ationalMapping, set out to explore the extent towhich the principle could be applied in aproduction mapping situation. The TorresStrait was selected because of the largeexpanses of shallow water and the numberof other features of interest in the scene.
Discussions with officers of the Divisionof National Mapping have indicated thatexisting mapping standards need not be considered absolute. They argued that thesehad largely been determined by the capabilities of conventional survey techniques,not necessarily by what users required. Inthe bathymetric mapping case, wherechanges tend to be slow, precision was notas critical as the standards might indicate.In the light of the costs in time and money ofconventional surveys, it was, in some cases,a matter of anything being better thannothing. These arguments apply to bothgeometric and depth accuracy.
Soundings currently used in the preparation of hydrographic charts in Australia aremeasured in hlthoms tor depths greater than11 fathoms, and fathoms and feet for depthsless than 11 fathoms; or metres for depthsgreater than 31 metres, and metres anddecimetres for depths less than 31 metres;depending on whether the survey wascarried out before or after metrication. Itwas realized that Landsat could not matchthis accuracy, and it was agreed that a depthresolution of 10 percent of nominal depth or1 metre, whichever was the greater, wouldserve as a guideline.
THE TORRES STRAIT SCENE
Scene 1026-00035, scanned at 10:03 am on
" 'flest il.ee
c .... Jt>'I1'11t.\.o
ct
'Il1'>\>~S<: of • Hammond Rockc!j d d
,.C"~ woo...,??- Island. .aIMllon 0 0
~~~ sland Thursday
OISland
'),~.,.
0cJ 0 C...
• j}York
()
" Ii.,~
&~
¢.'v'V'to..:!O
AustralianMainland
d .a,
FIG. 1. Torres Strait test areas.
18th August 1972, was used for the study.Sun elevation was 46 degrees while azimuthwas 58 degrees. All detectors were set at lowgain. The scene stretches from Cape York inthe south to Papua-New Guinea in the north,but all test areas were drawn from the southwest corner of the image (Figure 1) as thiscontained the most significant features.
Tidal height was approximately 1.4 mwith a variation of less than 0.3 m across atest area (Admiralty Tide Tables, Vol 3,1972). The tidal stream, including prevailingcurrent, measured at Hammond Rock inPrince of Wales Channel was approximately1.5 knots in a westward direction. At theentrances of the channel this rate wouldhave been reduced to approximately 0.5knots. Streams in Endeavour Strait lag thoseat Hammond Rock by about 40 minutes and,except in the more restricted sections of thestrait, would not have exceeded 0.5 knots.
Ground truth soundings used in the studywere generally over 30 years out of date andso it was frequently impossible, withinproject constraints, to determine with certainty whether some discrepancies reflecteda change in water depth over time or wereanomalous. In most cases the physical characteristics of the locality provided goodreason to believe that the latter was true.
EXPERIMENTAL METHOD
Intensity and depth values were collected
AID IN THE PREPARATION OF HYDROGRAPHIC CHARTS 1013
Substituting Equation 2 into Equation 1gives
R= a + b . exp [-c' (D + y)]= a + [b . exp (-c . y)] . exp (-c . D) (3)
Equation 3 still has the same form asEquation 1 and, since the numerical fitapproach is used to calculate optical modelcoefficients, all depths will be measuredrelative to the same datum used for groundtruth soundings. This can result in a negative depth which is not detected as land byMSS band 7. Such a situation signifies thatthe sea bed is above the zero depth datumbut below the surface at the time of thesatellite overpass. This generally impliesthat under some conditions it will be exposed.
along profiles specified by reference tosuitable land features. This method of datacollection was selected on the grounds thatit was easiest to implement and simulatedthe likely situation in a real survey. Thegeographical arrangement of the groundtruth data long a profile, in itself, provedan aid to interpreting anomalies as is described below. A numerical fitting technique was used to calculate the coefficientsin Equation 1. The study centered on agroup of islands and reefs under the influence of a moderate tidal stream ratherthan on a more open section of sea so thatit could reasonably be expected to uncovercomplications which affect the practicalimplementation of the technique.
Results were evaluated in four ways:
(1) Degree of fit that could be obtained alongthe ground truth profiles;
(2) Qualitative assessment over a broad areausing a pseudo-color display of calculateddepths;
(3) Location of depth contours; and(4) Degree of fit that results along test
profiles.
EFFECT OF TIDES
Tides will affect the analysis in two mainways. Firstly, tidal streams may pick up sediments and mix waters of different types.There will be consequent changes in opticalparameters which must be allowed for. Inaddition, tides result in actual depthchanges which can vary considerably over aLandsat scene.
Let D be the depth referred to the datumand y the difference between the depth atthe time of the satellite overpass, z, and D.That is:
* "Geometric" features denote those related tothe spatial characteristics of the two-dimensionalimage as distinct from depth or radiometricfeatures.
While tidal heights are compensated atanyone location, tidal height differencescause the coefficient b in Equation 1 to varyfrom one place to another. Consequently, anew numerical fit must be calculated wherever there is a signficant variation in tidalheight. Test areas in the current study aresmall enough for this effect to be insignificant, but it must be considered under operational conditions.
RESULTS
Attempts to fit Equation 1 to raw Landsatdata yielded maximum penetrations of approximately 15 m with an RMS error of 25percent and with considerably higher maximum errors. There is a clear need for radiometric correction and enhancement beforebathymetric mapping can be carried out.The limited useful range of signal availablefor oceanographic studies and the sophistication of the necessary correction procedures makes it mandatory that digital ratherthan photographic data be used. Data preprocessing was applied such that there wasa three-times contrast stretch with an effective reduction of noise, including striping,so that it did not rise above 0.5 grey levelsduring the stretching process. The correction is more complex than this descriptionmight indicate and a more detailed accountof the correction process is in preparation(Warne, 1977). Both penetration and accuracy were improved by the enhancementprocess.
Even with radiometric correction it wasnot always possible to get a good numericalfit in relating radiance to depth profiles.Discrepancies appeared to arise for tworeasons. The first is the geometric· limitations of the data, especially after a loss ofresolution resulting from the data correctionprocess. Relatively large errors occurred inthe vicinity of steep gradients or small features. Anomalous data points of this typewere edited out of the profiles so as not todisturb the numerical fitting process. Secondly, the optical characteristics of theocean and atmosphere may vary along theprofile. Figure 2 shows a profile where thereis a clearly defined anomaly at one end anda change in model coefficients towards themiddle. In such cases it is necessary to segment the profile into lengths correspondingto homogeneous areas of ocean. Accuracy ofnumerical fit depends on the extent to which
(2)z = D + Y
1014 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING, 1978
DEPTH
+
(4)z = In (b)/c
In some areas the calculated penetrationsfell below 10 metres; however, this waslimited to situations where all availableground truth depths were significantly shallower than the penetration limit. Consequently, little confidence can be placed inthe numerical prediction. The only confirmed penetrations less than 20 metreswere in areas with the greatest tidal streamactivity. This indicates that under morefavorable tidal conditions the penetrationwould be improved. In the more favorableareas of the image, penetration was confirmed to be in excess of 20 metres. Calculated penetrations were as high as 40metres but, once again, it was not possibleto verify this figure.
The exponential form of the relationshipcauses resolution to fall off rapidly as depthapproaches the penetration limit. Overallresolution cannot be better than that foundfor the control profile, which is largelydetermined by the level of noise. The degree to which it falls short of this figurewill depend on the variation of opticalcharacteristics that is permitted before newground truth data are taken and new parameters calculated. Analysis of the resultsshows that lines of soundings 5 kilometresapart could detect the broad scale variationsin Endeavour Strait sufficiently to keep theRMS errors to about 10 percent. In someinstances water separated by many timesthis distance had the same characteristicswhile in others the 5 kilometre soundingpattern, while it detected the variations,was too coarse to give a good indication ofwhere to define the boundary between thetwo sets of parameters. The slight tidalstream through the strait at the time of theoverpass would not provide the most favorable conditions and so under some circumstances a coarser sounding pattern may befeasible. This is supported by the fact thatvariation was greatest where the strait wasnarrowest and the stream consequently at itsfastest.
More thorough ground truth collection isrequired to clarify localized anomalies oralternatively the area must be left foranother overpass of the satellite. Theseanomalies are of two types:
(1) localized disturbances in optical characteristics of the water or atmosphere; and
depth, z, such that the radiance, R, exceededthe deep water signal, a, by one grey level.That is
... +
O~t. frolll re91on. ofdifferent optical parallletera
...
+
...+
++
+
FIG. 2. Radiance versus depth diagram.
DEPTH
FIG. 3. Radiance versus depth with no indicationof geographical relationship.
profile segmentation is carried. For reasonably macroscopic segmentation, RMS errorsof about 5 to 8 percent were typical, withmaximum errors a few percent higher. Figure 2 illustrates the desirability of maintaining information on the geographicalrelationship of the soundings. The sametrends are not nearly as clear in Figure 3where there is no clue to the order of thesoundings.
All result evaluation methods confirmedthe existence of broad scale and localizeddisturbances of the relationship betweendepth and radiance. In general, the verylocalized anomalies can be clearly seen in orinferred from the image by an interpreterwith the help of ancillary data such as tidaland meteorological reports. Some broadscale trends are also detectable in this way,but others can be located only with the helpof ground truth depth measurement.
Because of the variability in opticalparameters, model coefficients were calculated for each test area separately. Penetration was calculated by determining the
AID IN THE PREPARATION OF HYDROGRAPHIC CHARTS 1015
(2) small features and steep gradients beyondthe resolving power of the MSS systemand subsequent data correction process.
Likely disturbances in optical characteristics, as mentioned above, can generally bedetected by an interpreter through directobservation or inference.
DISCUSSION
The lack of homogeneity in opticalcharacteristics seriously complicates theimplementation of Landsat as an aid in thepreparation of hydrographic charts. Therelationship between parameter variationand tidal streams indicates that many fluctuations are temporary. Consequently, anyapproach based on shipborne measurementof optical parameters and subsequent calculation of coefficients for Equation 1(Polcyn, 1976) will involve considerable difficulties in survey management. Coefficientcalculation based on depth measures is moreviable, but once again the coefficients cannot unreselvedly be applied to a broad scalemapping.
Ground truth data will generally be in theform of depth profiles. It can be seen fromFigure 2 that profiles which cross a significant ridge or trough are more helpful in detecting parameter variations than are monotonic profiles. In addition, ground truth datamay be required to resolve localized anomalies. Accordingly, the Landsat image(s)should be used in determining the pattern ofsounding collection. It may be that analysisbased on ground truth reveals further anomalies which must be resolved by further datacollection. Consequently, some degree ofcoordination between image interpretationand ground truth survey is required. This isfar less restrictive than coordination between survey and satellite overpass.
Precise geographic positioning of groundtruth data is difficult at sea due to the absence of defined landmarks. Fortunately,positioning becomes critical only in areas ofsignificant spatial detail. Ground truth profiles may first be located approximatelyusing standard geographic control procedures and then, if there is signiflcant detail,the interpreter can use this detail to do thefinal positioning interactively. Any residualimprecision will give rise to errors in relating depth to radiance. Such errors mayalso occur due to a lack of sufficient spatialresolution. A slight blurring of a steep gradient may be acceptable in the final product,but the same errors in the ground truth datacan seriously affect the numerical fitting ofEquation 1. The interpreter must be able to
edit the ground truth data to eliminate suchdoubtful areas.
In the above discussion there are severalreferences to the use of an interpreter to resolve various difficulties. There is also ademonstrated need to use the imagery in itsdigital form. An interactive approach is,therefore, indicated. The interpreter has anumber of functions, but the main one is touse his perception of the image, knowledgeof marine characteristics, and ground truthdata in order to segment the image intoregions of sufficiently constant optical characteristics and to identify areas where suchconstant characteristics cannot be assumed.Once this is done the production of thedepth data is straightforward. Mixing of twowater bodies may complicate the procedureand such areas may need to be omitted inthe particular satellite overpass under consideration or, alternatively, an interpolationapproach may be used. Current interactiveimage analysis systems are slanted towardsland cover classification applications and areunsuitable in the present context. An appropriate system is approaching completion atthe Australian National University and willbe described in a later paper.
Changeability in the pattern of opticalparameter values allows the use of imageryfrom multiple overflights as a check that noparameter variation has gone undetected.Use of multiple images also permits thepenetration and resolution for each area ofthe image to be optimized and allows theinterpreter to discard doubtful results froma particular overpass.
Small features that constitute a navigational hazard but are beyond the resolutioncapabilities of Landsat remain a problem.Supporting aerial photography could resolvethis and, at the same time, provide the interpreter with a valuable aid to his analysis.If ground truth collection is done using anairborne profiler, proflling and photographycould be combined to maximize cost savings. Aerial photography with its extra resolution also gives better land shape determination. Byrne and Honey (1977) andFleming (1977) discuss a number of advantages in combining Landsat and aerialphotography.
The ANU study has confirmed that depthpenetration to 20 metres at an accuracy of10 percent RMS is feasible with Landsat, butit has also indicated problem areas that mustbe overcome before production run mapping is a reality. Reliability of the resultswill depend on the optical parameters, thenumber of satellite overpasses used to check
1016 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING, 1978
the analysis, the density of ground truthsoundings, and the amount of supportinginformation such as aerial photography.Consequently, it is closely related to the costexpended on the survey and so the authorityresponsible can achieve a proper balancebetween cost and reliability in any givenproject. For medium quality productsmeeting the standards set out in the introduction, the number of soundings is reducedby a factor of between ten and 50 whichwill result in considerable savings of bothtime and money, especially when comparedto a completely shipborne survey.
REFERENCES
Byrne, P. M., and F. R. Honey, 1977. Air Surveyand Satellite Imagery Tools for ShallowWater Bathymetry, Proc. 20th Australian Survey Conference, Darwin, May 1977, pp. 103119.
Fleming, E. A., 1977. Positioning Off-Shore Fea-
tures with the Aid of Landsat Imagery, Photogrammetric Engineering and RemoteSensing, Vol. 43, No.1, Jan. 1977, pp. 53-59.
Lyons, K. J., 1977. Evaluation of an ExperimentalBathymetric Map Produced from LandsatData, Cartography, Vol. 10, No.2, July 1977,pp.75-82.
Polcyn, F. C., 1976. NASAICousteau OceanBathymetry Experiment: Remote Bathymetry Using High Gain Landsat Data, NASAPublication: CR-ERIM-1l8500-1-F, July1976.
Turner, L. G., and H. L. Mitchell, 1977. SatelliteImagery and its Applications to OffshoreMapping in Australia, Proc. 20th AustralianSurvey Conference, Darwin, May 1977,pp. 91-99.
Warne, D. K., 1977. Analysis and Correction ofRadiometric Defects in Landsat Imagery (inpreparation).
(Received October 26, 1977; accepted March14, 1978)
Notice to Contributors
1. Manuscripts should be typed, doublespaced on 8! x 11 or 8 x 10! whitebond, on one side only. References,footnotes, captions-everything shouldbe douhle-spaced. Margins should beIt inches.
2. Ordinarily t l(;() copies of the manuscript and two sets of illustrationsshould be submitted where the second set of illustrations need not -beprime quality; EXCEPT that fivecopies of papers :>n Remote Sensingand Photointerpretation are needed,all with prime quality illustrations tofacilitate the review process.
3. Each article should include an ab-
stract, which is a digest of the article.An abstract should be 100 to 150 wordsin length.
4. Tables should be designed to fit into awidth no more than five inches.
5. Illustrations should not be more thantwice the final print size: glossy printsof photos should be submitted. Lettering should be neat, and designed forthe reduction anticipated. Please include a separate list of captions.
6. Formulas should be expressed assimply as possible, keeping in mindthe difficulties and limitations encountered in setting type.
Journal Staff
Editor-in-Chief, Dr. James B. CaseNewsletter Editor, M. Charlene Gill
Adveltising Manager, Hugh B. LovingManaging Editor, Clare C. Case
Associate Editor, Primary Data Acquisition Division, Philip N. SlaterAssociate Editor, Digital Processing and Photogrammetric Applications Division, Dr. H. M.
KararaAssociate Editor, Remote Sensing Applications Division, John E. LukensCover Editor, James R. ShepardEngineering RepOIts Editor, Gordon R. HeathChairman of Alticle Review Board, James R. LucasEditorial Consultant, G. C. Tewinkel