**3. Case study in SAV mapping using hyperspectral data**

#### **3.1 Hyperspectral algorithm to correct overlying water effects**

A new water-depth correction algorithm was developed to improve detection of underwater vegetation spectra signals. The algorithm was developed conceptually, calibrated, and validated using experimental and field data. The conceptual model was based on the idea that the upwelling signals measured from a water surface is the sum of the energy reflected from the water surface, the water column, and also from the water bottom surface. The energy reflected from the water surface and the water column (the volumetric reflectance) was combined as a single term because the surface reflectance is a constant and does not change with water depth changes (Lu & Cho, 2011).

The effects of the overlying water column on upwelling hyperspectral signals were modeled by empirically separating the energy absorbed and scattered by the water using data collected through a series of controlled experiments using hypothetical bottom surfaces that either 100% absorbs or 100% reflects (Cho & Lu, 2010). Later, the white surface (the 100% reflecting surface) was replaced with a gray surface with a known reflectance to reduce problems associated with enhanced multi-path scattering (Lu & Cho, 2011). The experimental setting allowed the calculation of water absorption and scattering values for up to a water depth of 60 cm, and the light remaining at water depths that were beyond the experimentally measured points were estimated by establishing the mathematical relationships between water depth and the vertical attenuation coefficient (*Kd*) derived from the experimental data (Washington et al., 2011). The depth- and wavelength-dependent water absorption and volumetric scattering factors (0-100 cm; 400 – 900 nm) were calculated and applied to independently-measured underwater vegetation signals and airborne hyperspectral data taken over shallow seagrass beds, to remove the effects of the overlying water.

#### **3.2 Successful water correction in the infrared region**

The empirically driven correction algorithm significantly restored the vegetation signals, especially in the NIR region, when applied to independently measured reflectance of underwater plants taken over indoor and outdoor tanks (Cho & Lu, 2010; Washington et al. 2011). The algorithm was also successful in restoring the NIR signals originating from seagrass-dominated sea floors when applied to airborne hyperspectral data of Mississippi and Texas coastal waters (Cho et al., 2009; Lu & Cho, 2011; Fig. 2). As stated earlier, NIR reflectance serves as the primary cue for discriminating benthic vegetation from other substrates. Due to the restored NIR reflectance, the correction algorithm increased the NDVI values for the seagrass pixels (Lu & Cho 2011).

#### **3.3 Seagrass classification using water corrected image**

#### **3.3.1 Ground truth data collection**

Several hundred ground data points were collected over seagrass beds in Redfish Bay, Texas, in the summer of 2008 (June – July). Seagrass species makeup, water depth, vegetation percent coverage, and bottom substrate type were recorded at each site. Site

materials within the water column, the surface of the water itself, and physical properties such as the absorption of energy in the NIR and beyond can all affect the ability to discriminate the relatively small differences in ratios of accessory pigments and chlorophyll (Kutser 2004).

A new water-depth correction algorithm was developed to improve detection of underwater vegetation spectra signals. The algorithm was developed conceptually, calibrated, and validated using experimental and field data. The conceptual model was based on the idea that the upwelling signals measured from a water surface is the sum of the energy reflected from the water surface, the water column, and also from the water bottom surface. The energy reflected from the water surface and the water column (the volumetric reflectance) was combined as a single term because the surface reflectance is a constant and does not

The effects of the overlying water column on upwelling hyperspectral signals were modeled by empirically separating the energy absorbed and scattered by the water using data collected through a series of controlled experiments using hypothetical bottom surfaces that either 100% absorbs or 100% reflects (Cho & Lu, 2010). Later, the white surface (the 100% reflecting surface) was replaced with a gray surface with a known reflectance to reduce problems associated with enhanced multi-path scattering (Lu & Cho, 2011). The experimental setting allowed the calculation of water absorption and scattering values for up to a water depth of 60 cm, and the light remaining at water depths that were beyond the experimentally measured points were estimated by establishing the mathematical relationships between water depth and the vertical attenuation coefficient (*Kd*) derived from the experimental data (Washington et al., 2011). The depth- and wavelength-dependent water absorption and volumetric scattering factors (0-100 cm; 400 – 900 nm) were calculated and applied to independently-measured underwater vegetation signals and airborne hyperspectral

data taken over shallow seagrass beds, to remove the effects of the overlying water.

The empirically driven correction algorithm significantly restored the vegetation signals, especially in the NIR region, when applied to independently measured reflectance of underwater plants taken over indoor and outdoor tanks (Cho & Lu, 2010; Washington et al. 2011). The algorithm was also successful in restoring the NIR signals originating from seagrass-dominated sea floors when applied to airborne hyperspectral data of Mississippi and Texas coastal waters (Cho et al., 2009; Lu & Cho, 2011; Fig. 2). As stated earlier, NIR reflectance serves as the primary cue for discriminating benthic vegetation from other substrates. Due to the restored NIR reflectance, the correction algorithm increased the NDVI

Several hundred ground data points were collected over seagrass beds in Redfish Bay, Texas, in the summer of 2008 (June – July). Seagrass species makeup, water depth, vegetation percent coverage, and bottom substrate type were recorded at each site. Site

**3. Case study in SAV mapping using hyperspectral data 3.1 Hyperspectral algorithm to correct overlying water effects** 

change with water depth changes (Lu & Cho, 2011).

**3.2 Successful water correction in the infrared region** 

**3.3 Seagrass classification using water corrected image** 

values for the seagrass pixels (Lu & Cho 2011).

**3.3.1 Ground truth data collection** 

Fig. 2. The original (left) and water-corrected (right) airborne AISA Eagle hyperspectral image at 741 nm obtained over seagrass beds in Redfish Bay, TX in 2008.

location was recorded to accuracies within 1 m using a Real Time Kinetic (RTK) GPS. When necessary, the preselected random sites were shifted to avoid dry or unreachable locations.

The field collected data were entered into a spatial database along with descriptive attributes to help determine which class each sampling site would be assigned to. Data points were then randomly divided into training or accuracy assessment points.

#### **3.3.2 Image processing and vector classification**

Image data were obtained in 63 bands of the AISA Eagle Hyperspectral sensor over the seagrass beds in October 2008 and corrected for atmospheric effects. Since selection of the proper bands for analysis helps reduce noise introduction and processing burden (Borges et al. 2007), several selection techniques were used within this project, including Principal Component Analysis and regression analysis. Ultimately, seven bands recommended by Fyfe (2003) were used. To reduce noise, these were again reduced to 5 bands, as recommended by Cho et al. (2009).

Image segmentation is performed prior to image classification. Segmentation groups like pixels into homogenous areas. Initially, an unsupervised classification using ISODATA (Iterative Self Organized Data Analysis Technique A) (De Alwis et al., 2007) was performed. After the initial image classification, each segmented vector in the output was assigned to a seagrass/substrate class (i.e. *Thalassia testudinum*, *Halodule wrightii*, *Ruppia maritima*, Mixed Beds, Bare, or Unclassified). Those which contain only one type of point (*'Halodule'*) were considered to be finally classified. Those classified as mixed or unknown were removed from the classified vector set, a mask created of their spatial footprint, and the entire

Remote Sensing of Submerged Aquatic Vegetation 305

water absorption of the near infrared energy. The fluctuating water depths, high amounts of suspended particles and colored dissolved organic matter in shallow littoral zones make it even more challenging to map benthic vegetation using remotely sensed data. A new waterdepth correction algorithm was developed conceptually, and calibrated and validated using experimental and field data. The effects of the overlying water column on upwelling hyperspectral signals were modeled by empirically separating the energy absorbed and scattered by the water using data collected through a series of controlled experiments. The empirically driven algorithm significantly restored the vegetation signals, especially in the NIR region. Due to the restored NIR reflectance, which serves as the primary cue for discriminating SAV from other substrates, use of the water corrected airborne data increased the NDVI values for the SAV pixels and also improved the seagrass classification accuracy. Our continuing efforts to incorporate turbidity and CDOM into the algorithm, in developing a graphical user interface, and in implementing the algorithm into a module that can be called from commercially available image processing software promise a user-

The work was supported by grants from the National Geospatial-Intelligence Agency (Grant No. HM1582-07-1-2005 and HM1582-08-1-0049) and National Oceanic and Atmospheric Administration-Environmental Cooperative Sciences Center (Grant No.NA17AE1626).

Ackleson, S. G., & Klemas, V. (1987). Remote Sensing of Submerged Aquatic Vegetation in

Beget, M. & Di Bella, C. (2007). Flooding: The Effects of Water Depth on Spectral Response of Grass Canopies. *Journal of Hydrology*, Vol.335, No.3-4, pp. 285-294. ISSN 0022-1694 Borges, J.; Marçal, A. & Dias, J. (2007). Evaluation of feature extraction and reduction

Bukata, R.; Jerome, J.; Kondratyev, K. & Pozndyakov, D. (1995). Optical Properties and Remote

Cho, H.J.; Lu, D. & Washington, M. (2009). Water Correction Algorithm Application for

Cho, H.J.; Kirui, P. & Natarajan, H. (2008). Test of Multi-spectral Vegetation Index for

Spectra. *Remote Sensing Letters,* Vol.1, No.1, pp. 29-35, ISSN 2150-7058 Cho, H.J., & Lu, D. (2010). A Water-depth Correction Algorithm for Submerged Vegetation

Spectra. Remote Sensing Letters, Vol.1, No.1, pp. 29-35

*Resource Conference,* pp. 154-157

*Sensing of Environment*, Vol.22, No.2, pp. 235−248, ISSN 0034-4257

Lower Chesapeake Bay: A Comparison of Landsat MSS to TM imagery. *Remote* 

methods for hyperspectral images. In: *New Developments and Challenges in Remote Sensing*, Bochenek, Z., pp. 265-284, Millpress, ISBN 978-90-5966-053-3, Rotterdam Brando, V.E.; Anstee, J.M.; Wettle, M.; Dekker, A.G.; Phinn, S.R. & Roelfsema, C. (2009). A

Physics Based Retrieval and Quality Assessment of Bathymetry from Suboptimal Hyperspectral Data. *Remote Sensing of Environment,* Vol.113, No.4, pp. 755-770, ISSN

Sensing of Inland Coastal Waters. New York: CRC Press.ISBN 978-0849347542 Cho, H.J. & Lu, D. (2010). A Water-depth Correction Algorithm for Submerged Vegetation

Underwater Vegetation Signal Improvement. *Proceedings of the 2009 MS Water* 

Floating and Canopy-forming Submerged Plants, *International Journal of Envirnmental research and Public Health,* Vol.5, No.5, pp. 447-483, ISSN 1661-7827

friendly application and wide use of the algorithm in the near future.

**6. Acknowledgment** 

0034-4257

**7. References** 

classification procedure re-run on the image for that footprint area only. When it became impossible to further classify the image by this method, a supervised classification was performed, using a selection of training points as training data, which produced monospecific vectors. The same mixed-method classification procedure was performed to the image data after the water correction algorithm was applied.

#### **3.3.3 Improved accuracy assessment**

The water correction algorithm improved the classification accuracy results in an image subset from an overall accuracy of 28% to approximately 36%. Identification of the species *Halodule wrightii* improved from 33% user's accuracy to almost 78%, and *Thalassia testudimun* from 0% users accuarcy to almost 17%. Although these numbers appear to be somewhat low, several factors must be considered: this analysis only used a subset of the imagery, which allowed the area analysed to have less variation, but there was also less training and accuracy assessment data to work with.

Using the full set of imagery and the combination of supervised and unsupervised classification, results have improved considerably; we have achieved overall accuracies of over 60%. In addition, we have calculated seagrass presence/absence using the complete corrected imagery, achieving overall accuracies of over 95%. With the future addition of *in situ* turbidity and bathymetric data, these accuracies should continue to improve.

#### **4. Current efforts in developing water correction module and graphical user interface**

We have continued improving the algorithm by including turbidity (measured in NTU, Nephelometric Turbidity Unit) as an additional function. In addition, the water correction algorithm is currently being implemented as a module that can be called from the ENVI (ENvironment for Visualizing Images, ITT Visual Information Solutions)'s programming environment using a Graphical User Interface (GUI) (Gaye et al. 2011). Under the current module and the GUI, users are able to select water depth (0 – 2.0 m) and turbidity (0-20 NTU) using slide bars, for which (a) given hyperspectral band(s) can be corrected. The corrected reflectance can be generated and compared to the original one at either a given pixel, within a small subset; and the original and corrected images can be displayed.
