OPTIMIZING CLASSIFICATION ACCURACY OF ESTUARINE MACROPHYTES: BY COMBINING SPATIAL AND PHYSICS-BASED IMAGE ANALYSIS JM Anstee1, EJ Botha1, RJ Williams2, AG Dekker1 and VE Brando1 1 CSIRO Land and Water, Canberra, ACT 2 Industry & Investment NSW, Cronulla, NSW
[email protected]
ABSTRACT Accurate baseline data of macrophyte extent is vital in estuarine monitoring. Previous techniques have often been laborious and subjective, while a purely empirical methodology often precludes transferring the method to other systems. The development of objective physics-based inversions models allows for the retrieval of; water depth, substratum composition and concentration of the water constituents from hyperspectral imagery. This paper describes approaches required to apply this method to QuickBird multispectral data from 2003 and 2008 over an estuarine lake. The addition of the inversion models quality control, improved the classification accuracy.
INTRODUCTION For a number of reasons, such as climate change and urban planning, environmental managers face challenges in decision making particularly in coastal areas [1]. Combined anthropogenic pressure and rising sea level may significantly alter the ecology of estuarine systems. Of particular importance is the fate of estuarine macrophytes, including seagrass, mangrove and saltmarsh as they provide fish habitat, contribute to food chains, enhance water clarity and reduce erosion. As estuarine ecologies are modified so too will coastal economies [1]. There is consequently an increasing need to obtain regular, accurate and up-to-date information on the extent of many estuarine resources including the macrophytes. Baseline data on the cover of estuarine macrophyte cover was produced for a portion of the southeast Australian coast (New South Wales (NSW)), in the early 1980s [2]. Maps were produced from aerial photographic interpretation (API) using the camera lucida technique, an analogue exercise that took at least 10 person-years to complete. While the methodology was a standard approach for its time, it had systemic operational errors that reduced its accuracy [3]. A hybrid analytical inversion approach coupled to image classification was conducted several years later in the form of a retrospective change analysis study using historical multispectral Landsat satellite imagery. The change in the subtidal habitat in one large estuary, Wallis Lake, was examined over the 14 year period 1988 - 2002 [4]. Key recommendations from [4] were to investigate the use of higher resolution imagery and alternative image interpretation techniques, for improved
classification accuracy. Traditionally, interpretation of remote sensing data has been image-based, and thus empirical. However, empirical approaches are not easily transferrable across study areas or data types [5]. Object-oriented analysis (eg, image segmentation) can increase classification accuracy by integrating additional (non spectral) information, such as texture and pattern, where the potential for spectral confusion exists [6]. Recently, physics-based inversion models have evolved to utilize semianalytical (bio-optical) methods that can incorporate water column properties and benthic substratum composition [7, 8]. These models allow us to accurately simulate the response of airborne or spaceborne sensors to various benthic substrata (forward modelling). The non-linear optimization of these models (inverse modelling) can be implemented to simultaneously derive water depth, water column properties and substratum composition from hyperspectral satellite imagery [7, 8]. The application of these methods for multispectral satellite imagery has several limitations [9]. Multispectral satellite sensors have wider spectral bands to ensure sufficient signal to noise. The increased width of the spectral bands often diminishes the identifying characteristic spectral features from the target being measured [9]. The method when applied on multispectral data may result in an overestimation of the inverted very low absorption coefficient in the blue and green band when the water is clear (and low in organic matter). In this part of the spectrum, the pure water absorption feature is 4-5 times greater than the constituent absorption and therefore dominates the absorption part of the spectrum. However, in turbid water (with higher concentrations of coloured dissolved organic materials (CDOM) and non-algal particulates (NAP), both of which absorb broadly in the lower wavelengths of the visible spectrum) the pure water absorption is relatively insignificant. [9] found the multispectral blue, green and red bands to have relatively small uncertainties in the modelled values (0.3m-1) and therefore these are suitable for use in semi-analytical models for the retrieval of coastal bio-optical properties. The challenge is to retrieve as much information as possible from the multispectral data within these limitations and therefore the processing pathways developed for hyperspectral imagery were modified for their application to multispectral data. This project attempted the application of the physics-based inversion model to high spatial resolution wide spectral band satellite imagery.
Pre-processing QuickBird (2.6m resolution) satellite imagery of Wallis Lake was acquired in January 2003 and September 2008. Image pre-processing (co-registration, atmospheric correction, segmentation) allowed for a standardized repeatable processing pathway required for comparisons between satellite data. A standardised atmospheric and air/water interface correction [11] was applied to the images to retrieve subsurface irradiance reflectance R(0-). In general, closure of atmospheric correction is performed by comparison between concurrent in situ and satellite water reflectance spectra over the same targets. However, collecting such in situ data concurrent with the overpass is often not feasible [12], as was the case with this project. Therefore, simulated spectra were compared to the satellitederived water reflectance. The simulated spectra were computed via the water column properties and benthic spectral library collected in the field. Figure 1 shows a reasonable closure between the QuickBird atmospherically and air–water interface corrected data to the simulated spectra. The spectra matches to within the calculated environmental noise equivalent reflectance difference [NE∆R(0-)E] in an image. The NE∆R(0-)E includes noise in the image data such as atmospheric variability, the air– water interface with swell, wave and wavelet induced reflections and refraction of diffuse skylight and direct sunlight [13, 14]. Validation data for satellite image based thematic maps require a spatially dense sampling design. By utilising image segmentation, point observations can be converted to polygons by taking advantage of the homogeneity of benthos in the vicinity of the geo-located point observations. This method partially addresses the problem of GPS accuracy, as non-differentially corrected observations can lead to errors of up to 30m and
Subsurface Irradiance Reflectance R(0-)
0.040
0.035
0.030
0.025
0.020
0.015
0.010
0.005
0.000 400
450
500
550
600
650
700
750
800
Wavelength (nm) satellite derived R(0-)
model der ived R(0-)
Figure 1: Forward model simulation over a green algae substratum for Pipers Bay, Wallis Lake. The spectra matches to within the NE∆R(0-)E represented by the vertical error bars. QuickBird geometric accuracy can be up to 10 pixels [15, 16]. Image segmentation reduces data complexity by introducing additional information such as pattern, texture and depth zonation [17]. This generally improves the fine scale separability of some seagrass species which may be spectrally indistinct from each other at depth. The QuickBird images were segmented into optically deep (where substrate does not influence R(0-)) and optically shallow portions for individual processing. Image processing A satellite imagery’s suitability for coastal mapping is determined by the sensors characteristics and mapping aims. Coastal and inland waters require high spatial resolution data due to the often large variation of water quality in relatively small areas [9]. Inversion of water quality information and benthic substrata is constrained by a satellite imagery’s spectral and spatial characteristics (Figure 2). Bathymetry can be resolved from reasonably low spatial and spectral resolution data, whereas chlorophyll-a requires a higher number of spectral bands (in the visible) to invert successfully. Benthic substrate-type mapping requires high spatial and spectral resolution data for optimal separation although good results can be achieved with moderate spectral resolutions. In the water column chlorophyll-a requires a higher number of pigment specific spectral bands (in the visible) to invert successfully. CDOM and NAP can be retrieved from medium spatial and low spectral resolution data, such as Landsat. As in optically deep parts of the coastal lake there is no need for QuickBird’s 2.6m resolution, the optically deep image segments were re-sampled to spatial resolution of 26m resulting in a higher signal to noise ratio. Figure 2: Inversion of water quality information and benthic substrate is 5 constrained by satellite imagery’s spectral and spatial characteristics. Bathymetry can be resolved from reasonably low spatial and spectral resolution data, whereas chlorophyll requires a higher number of spectral spectral resolution bands (in the visible) to invert bathymetry CHL CDOM NAP benthic type successfully. spatial resolution
METHODOLOGY Wallis Lake is an immature barrier estuary with extensive areas of estuarine macrophytes [2, 10], located on the central coast of NSW. Due to the lake’s variable water quality and substratum complexity, it poses a challenging target for developing an objective remote sensing monitoring method. A field campaign was conducted at representative sites around the lake in May and September 2008. In situ data were collected for model parameterisation, calibration and validation of the results. The aim was to characterise the optical properties of the water column, variations in optical properties of the benthos and collect geo-located validation data of benthos type, distribution and bathymetry. In situ optical properties of benthos and water constituents were used to create a representative library of the spectral responses within the estuary. In order to visualise the relationship between water constituents (concentrations and optical properties), benthic reflectance and the remotely observable reflectance, a forward simulation approach was implemented. This approach allowed a sensitivity analysis of the benthic spectral library and validation of the atmospheric correction. To select the appropriate satellite spectral bands for input into the physics-based processing pathway, the spectral separability of the field spectra was determined (converted to these multispectral bands). This required simulating the behaviour of each benthic spectrum at increasing water depth in the water type defined for this study. Benthic constituents that were spectrally distinct at depth were selected for the spectral library.
After glint and atmospheric corrections had been applied, subsets of the Wallis Lake imagery were processed. Similar to the approach outlined by [8], SAMBUCA, an enhanced implementation of the inversion/optimization by [7] was applied to the corrected satellite imagery. The SAMBUCA model was parameterised with the field derived measurements. RESULTS From the optically deep portions of the imagery, estimates were produced of the non algal particulate matter (NAP) while chlorophyll-a and coloured dissolved organic matter (CDOM) concentrations were fixed (Figure 3) as their variability was deemed to be below the NE∆R(0-)E . For the optically shallow portions of the imagery estimates of bathymetry and benthic composition were retrieved (Figure 4).
Output from SAMBUCA also estimates the accuracy between the modelled spectra and the image spectra. This estimation can therefore be an indication of the reliability, or confidence in, the SAMBUCA bathymetry and substratum estimates. A more detailed description of the data retrieval and error assessment procedure used in the SAMBUCA model can be found in [8]. To validate model accuracy, SAMBUCA bathymetry estimates were compared with in situ depth measurements collected during the field campaign (Figure 5). Substratum type for each image pixel is output from the SAMBUCA model as a percentage of two substratum types. Classification accuracy of substratum maps was assessed with standard post-classification statistical analysis tools using a validation dataset of geo-referenced in situ observations collected during the field campaign. As GPS accuracy can be variable (as great as 30m), point observations are often difficult to use to assess satellite classification accuracy. Polygons defined by the segmentation of the imagery were selected that co-located with a field observation and were assigned that value for depth or substratum-type. When more than one observation was recorded in each polygon an average value was calculated.
The segmentation was undertaken to delineate homogeneous feature and depth gradients which could then be applied to the field observations. Quality control assessment improved the R2 for both the 2003 and 2008 derived bathymetry estimates (Figure 5). 4.0
3.5
3.0
measured depth (m)
Figure 3: The retrieved NAP concentration from the optically deep regions of the QuickBird images from (a) 2003 and (b) 2008 after quality control correction. The optically shallow portion is grey. The measured range in May 2008 was 2.43 – 10.73 mL-1.
Figure 4: The quality-controlled retrieved dominant substratetype estimate of (a) 2003 and (b) 2008; and the qualitycontrolled retrieved bathymetry estimates from (c) 2003 and (d) 2008 from the QuickBird images.
2.5
2.0
1.5
1.0
0.5
0.0 0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
retrieved depth (m) QB08
1:1 line
QB03
QB08 QA removed
QB03 QA removed
Figure 5: The comparison between the SAMBUCA derived bathymetry estimates (x-axis) and the field observations (y-axis) from the optically shallow regions of the QuickBird images from 2003(▲) and 2008 (♦ ) after quality control correction. The before/after quality control R2=0.21/0.44 (2003) and R2=0.20/0.76 (2008).
[3] A.J. Meehan, R.J. Williams and F.A Watford, “Detecting trends in seagrass abundance using aerial photograph interpretation: problems arising with the evolution of mapping methods. Estuaries 28: 462-472, 2005. [4] A.G. Dekker, V.E. Brando and J.M. Anstee, “Retrospective seagrass change detection in a shallow coastal tidal Australian lake”, Remote Sensing of Environment. 97: 415-433, 2005. [5] T. Kutser, A.G. Dekker, and W. Skirving, “Modeling spectral discrimination of Great Barrier Reef benthic communities by remote sensing instruments”, Limnology and Oceanography 48: 497-510, 2003. [6] K. Johansen, N.C. Coops, S.E. Gergel, and Y. Stange, “Application of high spatial resolution satellite imagery for riparian and forest ecosystem classification”, Remote Sensing of Environment. 110: 29-44, 2007.
re Figure 6: The point field observations on the 2008 image (a) on the validation map (b) for comparing with the modelled substratum-type. The SAMBUCA derived estimate of substratum-type for the 2008 QuickBird image were compared with the point measurements in a subset of the lake ( Figure 6). The quality control assessment removed pixels that were optically deep and could not achieve a high closure between the image spectra and the modelled output spectra leading to a more accurate substrate-type thematic map of bright sand, a mixed class of shelly substratum and sparse seagrass and dense seagrass. The results in Figure 6 show a reasonable agreement. Our next step is to process the entire lake using this method and to compare more rigorously with in situ data. DISCUSSION The benefits of implementing this processing pathway include: • the in-built SAMBUCA quality control measures that lead to the exclusion of suboptimal data from the comparison, increasing confidence in the output • the use of segmentation to assign field observations to polygons not points. • the potential development of a more consistent trend assessment processing pathway. • The non-empirical approach can be transferrable across study areas or data types. Despite some limitations of the low spectral resolution of multispectral imagery, it has the ability to support management decisions by improving the effectiveness of coastal monitoring programs. Furthermore, when this approach is used to detect change it becomes a cost-effective source of information because it can be applied on historical, current and future images. Increased satellite spectral resolution will potentially improve the retrieval of coastal bio-optical properties using this physics-based approach with the benefit of being able to utilise the pre-existing spectral libraries. REFERENCES [1] NSW Government, “Draft Coastal Planning Guideline: Adapting to Sea Level Rise”, State of NSW Department of Planning, DOP 09_050, October 2009. [2] R.J. West, C.A. Thorogood, T.R. Walford and R.J. Williams, “An estuarine inventory for New South Wales, Australia”, Fisheries Bulletin 2. Department of Agriculture, New South Wales. 140 pp, 1985.
[7] Z. Lee, K.L. Carder, C.D. Mobley, R.G. Steward and J.F. Patch, “Hyperspectral remote sensing for shallow waters: 2. deriving bottom depths and water properties by optimization”, Applied Optics, 38, 38313843, 1999. [8] V.E. Brando, J.M. Anstee, M. Wettle, A.G. Dekker, S.R. Phinn and C. Roelfsema, “A physics based retrieval and quality assessment of bathymetry from suboptimal hyperspectral data”, Remote Sensing of Environment, 113, 755-770, 2009. [9] Z.P. Lee, “Applying narrowband remote-sensing reflectance models to wideband data”, Applied Optics, Vol. 48, No. 17, 3177-3183, 2009. [10] R.J. Williams, G. West, D. Morrison and R.G. Creese, “Estuarine Resources of New South Wales, CCA 04”. In (NSW Department of Planning, ed.): The NSW Comprehensive Coastal Assessment Toolkit, DVD 1, 64 pp, 2007. [11] V.E. Brando and A.G. Dekker, “Satellite hyperspectral remote sensing for estimating estuarine and coastal water quality”, IEEE Transactions on Geoscience and Remote Sensing, 41, 1378-1387, 2003. [12] A.G. Dekker, V.E. Brando, J.M. Anstee, N. Pinnel, T.,Kuster, H.J. Hoogenboom,, et al. “Chapter 11: Imaging spectrometry of water”, In: F. D. van derMeer, & S. M. de Jong (Eds.), Imaging Spectrometry: Basic Principles and Prospective Applications: Remote Sensing and Digital Image Processing (pp. 307−359). Boston: Kluwer Academic Publishers Dordrecht, 2001. [13] A.G. Dekker and S.W.M. Peters, “The use of the thematic mapper for the analysis of eutrophic lakes: A case study in The Netherlands”, International Journal of Remote Sensing, 14, 799–822, 1993. [14] M. Wettle, V.E. Brando and A.G. Dekker, “A methodology for retrieval of environmental noise equivalent spectra applied to four Hyperion scenes of the same tropical coral reef”, Remote Sensing of Environment, 93, 188−197, 2004. [15] C. Roelfsema and S. Phinn, “Integrating field data with high spatial resolution multispectral satellite imagery for calibration and validation of coral reef community maps” Journal of Applied Remote Sensing, Vol. 4, 043527 (2010). [16] Quickbird Imagery Products: “Product Guide, I”. DigitalGlobe, , DigitalGlobe, Inc. . Longmont, CO: p78, 2008. [17] S.L. Benfield, H.M. Guzman, J.M. Mair, and J.A.T. Young. “Mapping the distribution of coral reefs and associated sublittoral habitats in Pacific Panama: a comparison of optical satellite sensors and classification methodologies”, International Journal of Remote Sensing, 28(22): p50475070, 2007.
