Presentation on theme: "Jeremy M. Kerr National Coral Reef Institute"— Presentation transcript:
1 Worldview-02 offers new capabilities for the monitoring of threatened coral reefs Jeremy M. KerrNational Coral Reef InstituteNova Southeastern UniversityKhaled bin Sultan Living Oceans Foundation
2 Mapping coral reefs ecosystems Economically and socially importantDeclining healthMarine spatial planning - integrating spatial information into policyGeospatial informationBenthic habitat mapsBathymetric mapsEcological and oceanographic modelsIntegrating remotely-sensed data:Bridges gap between regional- and field-scale observationsConstrained by spectral and spatial resolution of the dataWorldview-02 increases the available spatial and spectralinformationElkhorn coral (Acropora palmata)Boulder star coral (Montastraea spp.)QuickBird imagery (True color)Optically-derived bathymetryBenthic habitat map
3 Deriving depth from a multispectral image Important optical properties of waterBeer’s law: light attenuates exponentially with depthLight attenuation varies across the visible spectrumStumpf et al. (2003) estimated depth based on these two properties with the model:Where= estimated depth= reflectance of band i= constant set to ensure that the logarithm is positive, = constants estimated by fitting a linear regression of the band-ratio to known depthsObservation:Increasing the polynomial from first-order to second- or third-order can lead to better depth estimates.Can the Stumpf model be improved?
4 Data Worldview-02 (WV2) image 10-km by 10-km (100 km2) area in the Florida KeysPixel resolution of 2-m by 2-m (4 m2)Coincides with a pre-existing bathymetric mapProvided the spectral inputs for the expanded depth derivation modelExperimental Advanced Airborne Research LiDAR (EAARL) Digital Elevation Model (DEM)Created through a combined effort by the USGS, NASA, and NPS (USGS 2008)Depths values for approximately 129 km2Depth estimates have error of ±15 cmPixel resolution of 1-m by 1-m (1 m2)Provided the ‘ground-truth’ data for fitting the models and assessing accuracy of predicted depths
5 Data WorldView-2 image (True color): ReefVegetationSandBoatsOptically-deep water*WorldView-2 image (True color):Subset 1 (red box) was used for model comparison.Subset 2 (orange box) and Transects A, B, C, and D (black lines) were used for accuracy assessments.Each Subset included 4 million pixelsTransects included between 2000 and 3000 pixelsEAARL DEM:Subsets and Transects are drawn for reference* - Optically-deep waters are those in which the sea-floor depth is greater than the detection limit (here: depths > 20 m, approximately).
6 Pre-processing WV2 image: Sea-surface state between the ‘traditional’ and ‘new’ bands variedPhysical sensor offsetLimited the possible band-ratiosSun-glint removal (SGR)Based on the model of Hedley et al. (2003)Applied to irradiance valuesDeep-water correction (DWC)Applied to apparent reflectance valuesEAARL DEM:Resampled using nearest-neighbor kernal to match the spatial resolution of the Worldview-02 imageClose-up from the a) coastal and b) blue bands demonstrating the observed difference in pixel values between the ‘traditional’ and ‘new’ bands. The top area in the yellow box (rows 1 and 2) is similar in both bands, but the bottom area (rows 3 to 7) has a different pattern in each band.
7 Expanding the Stumpf model Can the Stumpf model be improved?Hypothesis: Increasing the spectral information input into the Stumpf model will improve depth estimates.Increasing the number of band-ratios input into the model results in an expanded model:Where the n-th band-ratio, , is now:The constant e is added to ensure that the value inside the logarithm is greater than 0.What is the optimal set of band-ratios for depth-estimation?
8 The optimal combination of band-ratios Model selection: Based on Akaike Information Criteria (AIC; Akaike 1973)Sample size: 1 million pixelsRandomly selected from the 4 million pixels included in Subset 1Best-fit model:Full model (containing all 6 band-ratios)Same model selected for all 4 pre-processing levelsWhat is the optimal level of pre-processing?Band iBand jBand ratioBlueGreenRedCoastalYellowRed-edgeLimited to the within the ‘traditional’ and ‘new’ sets due to the observed offset63 candidate models4 levels of pre-processing
9 The optimal level of pre-processing Aggregated and compared all 252 models using AICOverall Best-fit model:Full-model with both SGR and DWCOther results:Top 3 models were full models with differing levels of pre-processing.8 of the top 10 models had both SGR and DWC applied.The top 10 models all contained 4 or more band-ratios.How does sample size affect the model selection results?The top 10 models based on AIC values.AICc is the AIC value after correction for small samples sizes.ΔAIC is the difference in AIC values between the model and the best-fit model.AICw is the weight of evidence within the data-set supporting the model.RankInput variablesPre-preprocessingAICcΔAICAICw1SGR, DWC1.002None0.003SGR45678910SGR - Sun-glint removalDWC - Deep-water correction
10 The effect of sample size on the model selection Method:Compared top 10 modelsSample sizes: 100, 1000, , and100 re-samples from Subset 1 selected for each sample sizeResults:Rank 1 model pre-dominantly selected forall 4 sample sizesRank 2 and 3 models occasionally selectedRank 4 through 10 models rarely or neverselectedHow well does this new model perform compared to Stumpf’s original model?Model (Rank)Sample size100100010 000184 %100 %94 %89 %24 %0 %2 %6 %39 %5 %41 %5678910
11 DataNote:Subset 1 (red box) was used for determining optimal model, optimal pre-processing level,and the effects of sample-size.Subset 2 (orange box) and Transects A through D (black lines) were used for the accuracy assessment.
12 Assessing the expanded model Reef topDepth estimates from the full and blue/green models compared with depths form the EAARL DEMOverall increase in precision of depth estimatesBlue/green model:RMSE: 1.03 mNRMSE: 7.32%Full model:RMSE: 0.77 mNRMSE: 5.49%Similar increase in precision along the 4 TransectsIncreased precision over areas with high albedosSandReef topDepth values from the EAARL DEM (black), Full model (blue), and blue/green model (red) along Transects A through D. The full model better estimated depths over reef-top and sand areas.
13 Summary of results Expanded model: Altered an established model from a linear regression to a multiple linearregression.Allowed for increased spectral information to be used for depth estimation.Full model (with 6 band-ratios) selected as the best-fit model.Pre-processing:Sun-glint removal and atmospheric correction improve depth estimates.Sample size:Full model feasible even when ground-truth data are limited.Accuracy:Full model had more precise depth estimates than the blue/green model.It also had more precise estimates over seafloor features with high albedos.
14 Future applicationDepth is a fundamental environmental parameter in marine systems.Ecologically importantResources and disturbances vary with depth.Benthic community composition varies with depth.Benthic habitat mapsDepth can improve classification accuracySupport modelling and analysesSpecies distribution modelsHydrodynamic modelsGeologic historyMarxan analysesFine-scale, detailed bathymetric maps covering large spatial extents improve our knowledge of the distribution of marine resources and the environmental processes shaping these distributions.
15 Future applicationGlobal Reef Expedition: Science Without Borders (www.globalreefexpedition.org)In partnership, the Living Oceans Foundation and host countries will explore, map, and characterize coral reefs in remote locations enabling comparisons with coral reefs located near populated coastlines.Research activities:Acquire multispectral satellite imageryCollect detailed ground-truth andbathymetric dataConduct surveys of benthic and fishcommunitiesEvaluate processes controlling reefcommunity structureTargeted sampling and assessment ofresilience indicatorsEvaluate connectivity within andbetween reef ecosystemsStudy sites in the Bahamas (right):Cay Sal Bank (green box),Great Inagua (blue box),Little Inagua (orange box)Hogsty Reef (red box)
16 Acknowledgements Jeremy M. Kerr National Coral Reef Institute National Coral Reef Institute at Nova Southeastern University’s Oceanographic CenterDr. Sam PurkisGwilym RowlandsAlexandra DempseyKhaled bin Sultan Living Oceans FoundationCapt. Phillip RenaudDr. Andrew BrucknerJudges for DigitalGlobe’s “8-band Research Challenge”Dr. Ing. Anko BörnerDr. Stuart R. PhinnDr. Hrishikesh SamantDr. Walter ScottDigitalGlobeIan GilbertBrett ThomassieJeremy M. KerrNational Coral Reef InstituteNova Southeastern UniversityKhaled bin Sultan Living Oceans Foundation