Home | Contact ST  
Follow ST

Feature Article

Hyperspectral Oceanographic Data: Potential Application in Open Oceans
Improving the Ability to Assess Phytoplankton Biodiversity By Processing High Spectral Resolution Observations

Elena Torrecilla
Ph.D. Candidate
Technical University of Catalonia
Barcelona, Spain

Dr. Jaume Piera
Principal Scientist
Marine Technology Unit
Spanish National Research Council
Barcelona, Spain

Anja Bernhardt
Research Assistant
Helmholtz University
Young Investigators Group Phytooptics
Alfred Wegener Institute for Polar and Marine Research
Bremerhaven, Germany

One of the most powerful and fastest growing technologies in the field of ocean optics is hyperspectral instrumentation, which has been recently integrated into a variety of oceanographic observing systems. The capability to obtain high spectral resolution measurements at hundreds of narrow and closely spaced wavelength bands—from ultraviolet to near-infrared and with a resolution of better than 10 nanometers—provides the opportunity for improvements in the extraction of information regarding phytoplankton and other optically significant constituents in seawater.

Hyperspectral measurements potentially enable the extraction of more accurate spectral information when compared with multispectral measurements. For instance, there are spectral features related to characteristic pigment absorption peaks that are essential to identifying the presence of certain phytoplankton assemblages in the ocean. In that sense, the hyperspectral approach offers the potential to map phytoplankton community composition beyond the estimation of the chlorophyll a concentration, common to all taxonomic groups.

Hyperspectral technology, being less invasive and time-consuming than traditional marine biology observation procedures, can be used to characterize phytoplankton distribution from large-scale patterns (i.e., using remote-sensing observing platforms) to small-scale structures (i.e., using in-situ platforms such as vertical profilers or autonomous underwater vehicles).

The increasing availability and advantages of in-situ and remotely sensed hyperspectral measurements of ocean waters leads to a need for ongoing evaluation and optimization of high-resolution processing methods. In order to address this goal, a research initiative is being carried out by the Marine Technology Unit of the Spanish National Research Council and the Helmholtz University Young Investigators Group Phytooptics, a cooperative project between the University of Bremen and the Alfred Wegener Institute for Polar and Marine Research. This research initiative, which began in July, focuses on improving processing strategies of hyperspectral data sets collected recently by the Phytooptics Group on board the RV Polarstern, which are used to better assess variability of phytoplankton in the open ocean.

This article covers some of the processing advances achieved in hyperspectral oceanographic observations. First, it details the optimal application of derivative spectroscopy focused on demonstrating improvements in the identification of different phytoplankton pigment assemblages in the open ocean. Second, it studies the feasibility of using hierarchical cluster techniques to deal with high-dimensional spectral data sets that would allow automatic classification of different oceanic scenarios. These techniques could then be used to better understand phytoplankton biodiversity and dynamics.

Spectral Shape Analysis
Identifying different oceanic phytoplankton assemblages based on hyperspectral resolved observations is possible because algal groups coexisting in the ocean demonstrate different optical signatures. These optical differences are the major determinant of the variation in the spectral shape of underwater optical properties and depend on different factors (e.g., each phytoplankton group's cell size, internal structure and specific pigment composition).

In order to take advantage of higher spectral resolution observations, the research initiative's analysis is focused on the application of spectral shape analysis. In particular, derivative spectroscopy was used to compute changes in the curvature of a given spectrum. Furthermore, closely related spectral features are separated, which better enables the subsequent extraction of information about optically significant phytoplankton groups in the water.

(Top left, top middle) Hyperspectral technology. (Bottom) Above-water remote sensing reflectance measurements, Rrs(l), collected by hyperspectral radiometers at different stations (A through H) on the ANT-XXV/1 cruise on the RV Polarstern in 2008. Click to enlarge.

Successful extraction of spectral details of interest through derivative analysis requires a proper adaptation of two parameters involved in the data processing. First, it should be considered that the identification of spectral details of interest in the derivative spectra strongly depends on the selection of the sampling interval used in the computation of changes in shape over each entire spectrum. Therefore, this factor needs to be defined according to the spectral resolution and characteristics of the collected hyperspectral data sets. The underlying idea is that spectral data features of interest with a smaller scale than the selected sampling interval will not be preserved in the derivative results.

Second, derivative analysis is notoriously sensitive to signal noise; smoothing techniques must be previously applied to derivative computation of hyperspectral data, and a suitable smoothing filter size must be selected. For each analyzed hyperspectral data set, a compromise between the ability to resolve fine spectral details and the reduction of noise effects in the derivative spectra must be achieved.

Cluster Approach
Cluster techniques have been extensively used in many disciplines as a valuable methodology for unsupervised classification of patterns. The team's research aims to automatically classify different phytoplankton pigment assemblages by using hierarchical cluster analysis (HCA) applied to hyperspectral oceanographic data sets.

In HCA, each hierarchical cluster tree is created by partitioning a set of input data (i.e., a set of hyperspectral observations) into groups of objects. The clustering is performed using a linkage algorithm based on a previously calculated pairwise similarity distance between all objects included in the input data set. In that sense, the similarity between objects is computed, taking into account the complete spectral behavior of hyperspectral data.

HCA is particularly useful in the case of this research initiative because it is hard to define a priori a relevant set of spectral features in the original optical data. As previously mentioned, there might be a great variability in hyperspectral oceanographic information over the entire spectral range due to phytoplankton variability. Moreover, an angular distance has been found to be the most appropriate similarity measure for this type of data and application. Differences in the spectral shape of hyperspectral data are more evident than differences in magnitude.

The performance of the cluster-based approach has been improved when using an angular distance between hyperspectral oceanic measurements and in combination with a previous application of derivative spectroscopy to the analyzed data sets. The advantages offered by this cluster methodology have been exploited to conduct several sensitivity tests. These tests illustrate the influence of different parameters in the performance of HCA. Among these parameters are the smoothing filter size and derivative sampling interval and also the spectral range considered when the optical data are analyzed.

Diagram illustrating the general approach to radiative transfer modeling of optical underwater scenarios and cluster analysis with the aim of characterizing phytoplankton's distribution through the water column.

The potential usefulness of hyperspectral oceanographic data for discriminating different phytoplankton pigment assemblages in open-ocean environments has been demonstrated by applying an unsupervised HCA to hyperspectral remote-sensing reflectance measurements, Rrs(λ), collected at several stations within the Atlantic Ocean. However, after considering the second derivative of Rrs(λ), spectra corresponding to stations with the same phytoplankton pigment composition were grouped more closely and separately in the cluster analysis. It is also worth noting that the ability to address automatic discrimination has been optimized by selecting the suitable parameters for the derivative calculations (i.e., values of sampling intervals and smoothing filter window size of nine nanometers), cluster analysis (i.e., an angular distance as a similarity metric) and spectral range of analysis (i.e., from 435 to 495 nanometers). Results indicate that an effort must be made to test the influence of these parameters if similar analyses are carried out using these processing tools.

Due to the increasing use of hyperspectral radiometers on different oceanic observing platforms, larger hyperspectral data sets from various oceanic environments are becoming more available. Therefore, it is worthwhile to continue evaluating the proposed processing methodologies or to research new analysis strategies. In order to pursue these challenging research tasks, radiative transfer models play a key role. They provide the possibility of performing numerical simulations in controlled underwater light environments and improve exploration of the feasibility of any processing technique.

Future work will focus on exploring some bio-optical provinces detected by analyzing larger hyperspectral oceanographic data sets with the processing methods described. Assessing hyperspectral oceanographic data's potential relation to some ecological provinces defined in the literature is essential to improve understanding of marine phytoplankton's role in the global marine ecosystem and biogeochemical cycles.

The authors acknowledge Dr. Astrid Bracher, head of the Phytooptics Group, for her support within the framework of this research cooperation. They also thank the crew and technical staff of the RV Polarstern for assistance throughout data collection.

This study was supported by the Spanish National Research Council (projects HYDRA-PIE06/301102 and ANERIS-PIF08/015) and the Spanish Ministry of Education (Ph.D. European Mention Program 2010).

For a full list of references, contact Elena Torrecilla at torrecilla@utm.csic.es.

Elena Torrecilla performs research under a joint program between the Technical University of Catalonia (UPC) and the Marine Technology Unit in Barcelona, Spain. She holds an M.S. in telematics engineering from UPC. Her research interests include hyperspectral sensors, signal processing and bio-optical-based pattern recognition techniques.

Dr. Jaume Piera's research interests are related to information technologies applied to marine biology, particularly in bio-optics. He holds a B.Sc. in biology from the University of Barcelona, a B.S. in electronic engineering from the Technical University of Catalonia and a Ph.D. in environmental sciences from the University of Girona, Spain.

Anja Bernhardt is a research assistant in the Helmholtz University Young Investigators Group Phyto-optics at the Alfred Wegener Institute for Polar and Marine Research in Bremerhaven, Germany. She has been collecting in-situ data on several cruises, and her current research analyzes hyperspectral, bio-optical and satellite data. She holds a diploma in physics from the University of Bremen, Germany.

-back to top-

-back to to Features Index-

Sea Technology is read worldwide in more than 110 countries by management, engineers, scientists and technical personnel working in industry, government and educational research institutions. Readers are involved with oceanographic research, fisheries management, offshore oil and gas exploration and production, undersea defense including antisubmarine warfare, ocean mining and commercial diving.