Home Random Page


CATEGORIES:

BiologyChemistryConstructionCultureEcologyEconomyElectronicsFinanceGeographyHistoryInformaticsLawMathematicsMechanicsMedicineOtherPedagogyPhilosophyPhysicsPolicyPsychologySociologySportTourism






Satellite oceanography: New results 1 page

 

 

This chapter presents recent advances in remote sensing of the White Sea from satellite Earth observation (EO) sensors. Section 6.1 presents the methods for quan- titative remote sensing of water quality in marine environments such as the White Sea. Section 6.2 describes the patterns of thermo-hydrodynamic processes and fields derived from the NOAA Advanced Very-High Resolution Radiometer (AVHRR) data. Section 6.3 discusses remote sensing of the White Sea using active (radar) and passive microwave sensors. Finally, Section 6.4 presents studies of the White Sea ice cover using passive microwave sensors, which are used to characterize the spatial variability of sea ice over the past two decades.

 

6.1 OPTICAL REMOTE SENSING

 

6.1.1 Background

Satellite remote sensing has acquired an ever-increasing importance within the framework of both the problem of global climate change and the continuously accentuating problem of aquatic ecology, which in turn is intrinsically related to the world population s need for drinkable water, fish resources, and other seafood. The optical properties of natural waters are such that the electromagnetic radiation in the visible part of the spectrum is only capable of penetrating into the water column (e.g., Bukata et al., 1995). It naturally explains the exceptional role of satellite sensors operating in the visible spectral range in studying the processes occurring under the air-water interface. Penetrating through the water column, the incident light interacts with the aquatic medium, leading to pronounced changes in its spectral composition. The presence of scattering matter (including the water molecules per se), results in a certain fraction of the downwelling sunlight being scattered back into the atmosphere. This backscattered light carries


 

information about the optical properties of the water column back to the remote sensor. In that, the color of the targeted water body is a convolution of all photon interactions that occurred therein.

Both the depth of the sunlight penetration into the water column and the color perceived by a remote observer are functions of the content of absorbing and scattering substances coexisting in the given aquatic medium (e.g., Pozdnyakov and Grassl, 2003). In the clearest waters of open oceans the light can propagate down through hundreds of meters (e.g., the Sargasso Sea) and then partially return to the water-air interface, whereas in turbid waters this depth might be reduced to several tens of centimeters (Jerlov, 1976; Kirk, 1981).

In addition to the above-mentioned mechanisms of absorption and scattering, the sunlight photons propagating in natural waters can be involved in the processes of fluorescence by waterborne agents as well as in intricate interactions with the bottom. This complicates enormously the interpretation of both the spectral compo- sition of emerging light and the perceived water color (Pozdnyakov and Grassl, 2003).

Morel and Prieur (1977) coined the term case I waters for clear offshore oceanic waters. Case I waters are those in which phytoplankton, together with their accompanying and strictly co-varying deterial material of biological origin are the principal agents responsible for changes in the optical properties of the water body and, hence, its color. Thus, the remotely recorded variations in the spectral envelope of the water-leaving light signal can be unequivocally related to the concentration of phytoplankton/chlorophyll. Based on this property of case I waters, the so-called band ratio empirical and semi-analytical algorithms (such as the so-called OC4) (O Reilly et al., 1998, 2000) have been developed and used for processing images taken over open oceans. This and many similar band ratio algo- rithms are based on the ability of phytoplankton chlorophyll to efficiently absorb light in the blue part of the visible spectrum, leaving the emergent light intensity at ca 510-520 nm nearly intact. Thus the ratio of remotely recorded signals in these two wavelength bands can be related to the concentration of chlorophyll.



The NASA bio-optical band ratio algorithms for inferring the desired informa- tion from both the Sea-viewing Wide Field-of-View Sensor (SeaWiFS) and the Moderate-resolution Imaging Spectroradiometer (MODIS) (e.g., Ackleson, 2001) were designed to meet the main challenge of the world s oceans: to determine, in open oceanic waters, the temporal and spatial distributions of the concentration of phytoplankton (via its proxy chlorophyll) as a principal player in carbon cycle variations in relation to global change. The SeaWiFS Data Analysis System (SeaDAS) has become a comprehensive tool for this type of image processing (http://seadas.gsfc.nasa.gov).

However, the applicability of satellite optical sensors changes significantly when the targeted waters are within oceanic/marine coastal zones, closed seas, or when they constitute the so called surface freshwater (i.e., lakes, rivers, etc.). Their use here consists first and foremost in surveying temporal and spatial variations in the trophic status of such waters. The proximity of marine coastal and freshwater to the neigh- boring/encapsulating lands results in massive admixtures of suspended minerals


6.1

(brought in either with river discharge or land runoff) and dissolved organic content (doc) of both allochthonic and autochthonic nature (for details see e.g., Kondratyev et al., 1999), respectively, coming from soil humus and originating from the phyto- plankton decomposition by bacterioplankton. Due to the multiple mechanisms of generation, suspended minerals (sm) and doc generally do not correlate with the coexisting phytoplankton in the water masses. It means that the optical properties of such waters (called case II waters) are not related exclusively to phytoplankton but also by other substances that can vary mutually independently (e.g., sm and doc). These substances are not only undesirable noise for inferring the concentration of phytoplankton (or its proxy chlorophyll (chl )), but together with chl also constitute a set of very important indicators of the trophic status of case II waters and, hence their ecology.

In addition, the amount and spatial-temporal distribution of sm is presently acknowledged as an indicator of dispersion in aquatic bodies of radionuclide debris, as well as other environment impairing substances. Spatial and temporal variations in the sm pathways within closed seas, and marine coastal and surface waters is closely related to abrasion/ablation of the coastline and appearance/dis- appearance of sand banks, etc., which is highly consequential for both navigation and civil engineering.

Thus, the determination of concentrations of sm and doc (along with chl ) in marine coastal and freshwater bodies (i.e., case II waters) is the major incentive for satellite oceanography using optical sensors. At the same time, elaboration of tech- niques for simultaneously inferring the desired information on the content of chl, sm, and doc in case II waters presents a serious scientific challenge, mostly due to the concerted impact of the above constituents on the light signal being measured.

Impacting the water-leaving light to the measure of their inherent optical prop- erties and actual concentrations of chl, sm, and doc, conjointly determine both the spectral distribution of the immerging light and its intensity. Owing to this specific property of case II waters, it is impossible to infer the concentration of only one component (e.g., chl ), such that all of the constituents should be retrieved. This feature determines the unavoidable prerequisite to the search for an approach for processing case II waters: the retrieval algorithm should be able to determine the concentrations of all three concentrations simultaneously.

In the section that follows, a detailed description of the concept and properties of such an advanced retrieval algorithm will be thoroughly discussed and analyzed in terms of its applicability to: (a) case II waters with different compositions of color-producing agents (CPAs) (which, in this context, are primarily chl, sm, and doc); (b) levels and nature of noise of the input signal; and (c) operational efficiency and the attainable accuracies of CPA retrievals.

 

 

6.1.2 A new water quality retrieval algorithm for case II waters

When viewing a water surface from an altitude h at a zenith angle e to the surface normal and at an azimuth angle <p to the principal plane of the sun, a remote sensor


 

captures the radiance At, which is comprised:

At(e , <p , z, ,\)= [Aw(+0, e , <p , ,\)+ Ar(e , <p , ,\)]Ta + A (e , <p , z, ,\) (6.1)

where Awand Ar are the components arising from the light coming out from beneath the water surface and reflected by the water surface in direction (e , <p ), respectively; Ta is that portion of the surface-leaving radiance, which is neither absorbed nor scattered out of the field of view, and known as the diffuse transmittance of the atmosphere; A is the path radiance resulting from that solar radiation crossing through the field of view of the remote sensor which is scattered towards it by the atmosphere.

The desired information about the water quality parameters (viz. concentrations of phytoplankton chlorophyll, suspended minerals and dissolved organics, and some others) is solely contained in the component Aw, which could be related to the water inherent optical properties (IOPs) (the coefficients of absorption a, scattering b, attenuation c, and {3(e , <p , et, <pt, z, ,\) - the volume scattering function of the angle between the unscattered ray traveling in direction (et, <pt) and the scattered ray traveling in direction (e , <p )) through the coefficient of volume reflectance just below the water surface:

R(-0, ,\) = Eu(-0, ,\)/Ed(-0, ,\) (6.2)

where Eu(-0, ,\) and Ed (-0, ,\) are the upwelling irradiance and downwelling irra- diance just below the water surface.

The IOPs are known (e.g., Jerlov, 1976) to be additive in nature:

I I I I

a = ai b = bi bb = bbi {3 = {3i i = 1, 2, 3, ... , I (6.3)

i i i i

where i is the number of each CPA (including the water per se) composing the natural aquatic medium.

b
The additive equations (6.3) could be further modified via introducing specific absorption a and specific scattering b (backscattering b ) coefficients (also referred to as absorption and scattering (backscattering) cross sections):

 

I I I


a = Cia b = C b b


= C b {3 = C {3


i = 1, 2, 3, ... , I (6.4)


i i i b

i i


i bi i i i

i i


where Ciare concentrations of each CPA.

Therefore, given the specific IOPs, the bulk water optical properties could be assessed through in situ determinations of the major CPAs.

Thus, for remote sensing of natural waters, satellite images should be processed in such a way, that it will eventually yield accurate spectral values of upwelling radiance Aw, which can further be related either to R or remote sensing reflectance Rrs, which is the emerging upwelling radiance normalized to the downwelling irra- diance impinging the water surface (see also below, this section).

The removal of all other components in eq. (6.1) is called atmospheric correction. Understandably, the success of water quality parameter retrieval largely depends on


6.1

the accuracy of the removal of the atmospheric impact on the signal captured by a satellite sensor, as well as the resilience/robustness of the employed retrieval algorithm to the inevitable contamination of the legitimate signal with the light originating due to the optical impact of the intervening atmosphere.

As emphasized above in this section, for pristine/offshore oceanic waters qualify- ing as case I waters (recall, these are waters where phytoplankton together with the accompanying and co-varying products of their life cycles as well as some micro- scopic organisms such as flagellates, bacteria, and viruses are the principal agents determining the variations in optical properties of the water), it was believed for a long time that considerable simplifications are possible. Through a combination of statistical and semi-analytical, universally applicable, relationships between either the emerging upwelling radiance Aw(,\, +0)tor remote sensing reflectance Rrs(,\,

+0)tratios at two or more wavelengths, and the desired parameter, normally just

the concentration of chlorophyll-a, a simple and robust retrieval scheme has been developed.

The preferred choice was the ratio between two spectral bands, often the water- leaving radiance Aw(,\i, +0) in the blue, close to the wavelength of maximum absorp- tion by chlorophyll (,\i ), normalized by Aw(,\i, +0) at a wavelength ,\i , for which Aw(,\, +0) could be assumed quasi-independent of chlorophyll-a (leaving alone the water per se). This leads to very simple algorithms of the form:

Cchl = A[Aw(,\i, +0)/Aw(,\i , +0)]-B (6.5)

where A and B are regression coefficients.

When doc and sm are present at low concentrations the results of retrieval of chlorophyll content improves when ,\i and ,\i are shifted to longer wavelengths. Therefore, various ,\i /,\i pairs (443/520 nm; 443/550 nm; 520/550 nm; 520/670 nm) have been tried for a given combination of Cchl , Cdoc, and Csmto best fit eq. (6.5) (NASA, 1993; Siegel et al., 1999; for other references, see Kondratyev et al., 1999). Naturally, the regression coefficients A and B in eq. (6.5) are area and season- specific, since the optical properties of CPAs constituting the concentration vector C(Cchl , Csm, Cdoc) vary with space and time.

To mitigate this, the above band ratio paradigm was subjected to further improvements: a modified cubic polynomial function that uses the ratio Rrs(490,

+0)/Rrs(555, +0) and simulates well the sigmoidal pattern revealed between log- transformed radiance ratios and chl, was chosen as the pre-launch Sea-viewing Wide Field-of-View Sensor (SeaWiFS) operational chl-a algorithm (for OC2, OC2v2, OC2v3 parameters see, e.g., Smyth et al., 2002). Improved performance was obtained using the ocean chlorophyll algorithm 4 (OC4). This is a four-band (443, 490, 510, 555 nm) maximum band ratio formulation (O Reilly et al., 1998):

 

2 3

Cchl= 10(a0+a1 Rrs+a2 Rrs+a3 Rrs) + a4 (6.6)

where Rrs is the greatest log ratio among the pairs (Rrs(443)/Rrs(555),

 

t+0 indicates that the quantity in question is assessed just above the water surface; ,\ = the wavelength.


 

Table 6.1. Results of the NASA OC4 algorithm application for retrieving the concentrations of chl in case I and II waters (numerical experiments employing the hydro-optical model for mesotrophic freshwater (Kondratyev et al., 1990)). Concentrations of chl, sm, and doc are given in µgl-1, mg l-1, and mgC l-1respectively.

 

Combinations of CPA concentrations used for simulating water volume reflectance R sm = 0.1 doc = 0.1 chl = 0.1 doc = 0.1 chl = 1.0 sm = 0.1

chl = 0.1 chl = 3.0 chl = 5.0 sm = 0.1 sm = 1.0 sm = 5.0 doc = 0.1 doc = 1.0 doc = 5.0

chl by OC4 0.37 3.5 13.5 0.37 3.4 11.0 0.37 3.4 71.8

 

 

Rrs(490)/Rrs(555), and Rrs(510)/Rrs(555)), and a0 = 0.4708, a1 = -3.8469,

a2 = 4.5338, a3 = -2.4434, and a4 = -0.0414.

However, the causal explanation given above in this section, explicitly indicates that neither of the band ratio retrieval algorithms are appropriate for inferring solely chl, leaving alone other CPAs from remotely sensed data when obtained over case II waters. This will be further illustrated through numerical experiments, the results of which are given in Table 6.1. Indeed, as mentioned above, the optical properties of case II waters are such that the retrieval should be simultaneously performed for all maior CPAs.

Bukata et al. (1995) and Pozdnyakov and Grassl (2003) investigated the potential of other approaches to meet this challenge. They have collectively shown that the multivariate optimization technique (more specifically, the Levenberg- Marquardt (L-M) procedure: for references see, e.g., Kondratyev et al., 1990) appears to be a very efficient tool for attaining an accurate solution to the inverse problem when dealing with case II waters.

Briefly, the basic idea of the L-M procedure can be illustrated as follows. If R(,\, C, a, bb) is the water remote sensing reflectance calculated using a known parameterization; C = CPA concentration vector; a, bb = bulk absorption and back- scattering coefficients inherent in the targeted water column; and {Si } is the value of water body remote sensing reflectance obtained from in situ measurements, then the weighted residuals gi can be taken as a measure of concordance between the measured and simulated remote sensing reflectance:

gi = [Si - R(,\, C, a, bb)]/Si (6.7)

where i is the wavelength ,\, at which the hydro-optical measurements have been conducted; and a, bb are the cross sections of water absorption and backscattering, respectively.

Within the framework of the least squares method, the concentration vector

i
C(Cchl, Csm, Cdoc) can be found through minimizing the function of residuals over C:


 
f (C) =

i


g2(C) (6.8)


An iterative approaching the f (C) minimum can be pursued using the L-M technique, which being a variety of the Newton-Gauss method (e.g., Kondratyev


 

et al., 1990), has an enhanced converging capability. To find an optimal concentra- tion vector, corresponding to the desired global minimum of f (C), the following iteration formula is used:


t -1 t (


R(Ck) \


Ck+1= Ck+ ,\k(F kFk+ µkDk)


F k 1 -

S
k


(6.9)


where k is the iteration step; Dk = diag(F t F ) is a diagonal matrix, the main

k k

diagonal of which is composed of the elements F t F ; F (C)= l8R /8C l is a

k k i i


matrix of the (


n x m) order (n is the number of wavelengths at which the quantity


S was


measured, m is the dimension of the concentration vector C); F t(C) is a


6.1
transposed matrix F (C); µkis the direction of minimization; and ,\ is the optimiza- tion step length.

The method of reduction is the simplest and most efficient method of choosing the optimization step length. This method consists of the following: (1) set ,\ at a certain high value (e.g., 50), and (2) test the inequality f (Ck+1) < f (Ck). If the inequality does not hold, reduce ,\ by a factor of two and test the inequality again. This iteration procedure is to be repeated until the inequality is reached or

,\ becomes too small.

The search for the desired minimum of f (C) is only successful when the starting value of the concentration vector is close to the quaesitum. If the starting value of C0 proves to be far away from the quaesitum and the remote sensing reflectance spectrum is contaminated with noise, the minimum found for f (C) might correspond to some unrealistic values of C. To avoid such problems, a number of random starting values of C0 can be chosen, and the method of multivariate optimization initiated to search for the deepest global minimum. However, there is no guarantee that any particular initial vector C0 will result in the convergence of the iterative procedure, or else the determined concentration vector will prove to be physically sound (e.g., negative concentrations of one or several CPAs). To obviate such difficulties, the following constrains are imposed on Ci:

Ci min ::; Ci ::; Cimax (6.10)

where the subscript i refers to the ith component of the natural aquatic medium. Figure 6.1 illustrates the results of numerical experiments conducted by

Pozdnyakov et al. (2003). In these experiments the R parameterization suggested by Jerome et al. (1988) was employed. The retrievals have been obtained using the spectral values of absorption and backscattering cross sections (otherwise called a

hydro-optical model ), which have been established for surface mesotrophic waters (see e.g., Kondratyev et al., 1990). The results thus obtained indicate that the L-M procedure is capable of retrieving the CPA concentration vector with very-high accuracy within a wide range of concentrations of each CPA.

It follows from Figure 6.1 that given a zero-level of noise in the input data, the retrieval error assured by the L-M procedure can be as low as 1% even for the concentration range 1 ::; chl ::; 10 µg l-1; at higher concentrations of chl the retrieval error is even less. Importantly, Figure 6.1 displays no offset, which is also an explicit indication of the very good performance of this approach.


 

 

(a)

 

(b)

 

 

Simulated (c)

Figure 6.1. Results utilizing the L-M procedure to retrieve concentrations of (a) chl (µgl-1),

(b) sm (mg l-1), and (c) doc (mgC l-1), and their comparison with the reciprocal simulated/ input data.

For comparison reasons, analogous simulations were conducted using the NASA OC4 algorithm for a variety of combinations of CPA concentrations. As seen from Table 6.1, the chl retrieval accuracy is significantly less than the one assured by the L-M procedure for the simulated waters, which only slightly differ from case I waters, and it proves to be absolutely irrelevant for the retrieval of chl in the case of highly absorbing/scattering (i.e., case II) waters.


6.1

As was indicated above, the L-M performance efficiency is most sensitive in terms of the degree of closeness of the starting concentration vector C0to its quaesitum. The realistic limits for C0should either be known a priori or determined otherwise. Therefore, a judicious choice of C0is crucially important in terms of both retrieval success and reduction of the required computation time.

Neural networks (NNs) are known as very fast tools for the solution of inverse problems (for references see e.g., Atkinson and Tatnall, 1997). There are many different types of NNs, but one of the most commonly used in remote sensing is the multi-layer perceptron (MLP): the MLP generally consists of three types of layers. Importantly, for the performance of the MLP the requirement for a priori knowledge of C0is lifted. The input layer neurons are the elements of a feature vector, which might consist of radiances at certain wavelengths. The second layer is the internal or hidden layer. In the third layer, the number of neurons equals the number of parameters to be determined. Each neuron in the network is connected to all neurons in both the preceding and subsequent layers by connections with asso- ciated weights.

The input signals are transferred to the neurons in the next layer in a feed- forward manner. As the signal propagates from neuron to neuron, it is modified by the appropriate connection weight. The receiving neuron sums the weighted signals from all neurons to which it is connected in the previous layer. The total input that the ith neuron receives is weighted in the following way:

 

N

neti= wiioi (6.11)

i=1

where wii is the weight of the relationship between neuron i and neuron i, and oi is the output from neuron i. The output from a given neuron i is then obtained from:

oi = f (neti ) (6.12)

The function f is usually a non-linear sigmoid function. It is applied to the weighted sum of inputs before the signal reaches the next layer. When the signal reaches the output layer, the network output is produced. The created network should first be trained so that it can generalize and predict outputs from inputs that it has not processed before. A training pattern is fed into the NN and the signals are forwarded. After that, the network output is compared to the true output, the error is then computed and back-propagated through the network. As a result, the connection weights are modified following the general rule:

!iwii (n + 1)= T(bioi)+ o:!iwii (n) (6.13)

where T is the learning-rate parameter, bi is an index of the error-change rate, and o: is the momentum parameter. The training is conducted until the output error reaches a desired level of accuracy. The trained NN is then tested against some verification data to assess the network performance.

Applying the same hydro-optical model and the same R parameterization as the ones used for testing the L-M procedure, the performance of the built-up NN


Date: 2016-03-03; view: 710


<== previous page | next page ==>
Aquatic ecosystem profile | Satellite oceanography: New results 2 page
doclecture.net - lectures - 2014-2024 year. Copyright infringement or personal data (0.015 sec.)