aip

Bad Std S

Received in revised form 23 November 2012

ctrarinertaken concurrent with image acquisition provided calibration data for chloro-

Remote Sensing of Environment 130 (2013) 254265

Contents lists available at SciVerse ScienceDirect

Remote Sensing o

.eal resources that contribute signicantly to the state's economy andtourism. Of the 17% of the state's river and stream miles assessedfor the 2010 Impaired Waters List, 40% were found to be impaired(Minnesota Pollution Control Agency, 2011). We explored the use ofaircraft-mounted remote sensing systems as a cost-effective way togather information to measure optically active water quality proper-ties of rivers relevant to the issue of river water impairment. Thispaper describes a general approach, as well as specic predictive rela-tionships, that can be used for such measurements.

We have had success previously using multispectral radiance infor-

conditions in rivers and streams are temporally more dynamic andoften spatially more heterogeneous than those in lakes. Second, smallrivers and streams may be so shallow that light penetrates to the bot-tom, such that reectance from thewater is a function of bottom condi-tions in addition to that of the water itself. Third, the spatial resolutionof most satellite sensors, including Landsat, is too coarse for small riversand streams. Finally, to measure water quality conditions other thanclarity, a better set of spectral bands is needed than what Landsat pro-vides. Although the MERIS and MODIS satellite sensors provide suchbands, their coarse spatial resolution makes them suitable only formation from Landsat imagery (e.g., Olmansonlakewater clarity. More recently, we showed t(MERIS and MODIS) can provide accurate

Corresponding author at: Department of Forest Reso1530 Cleveland Avenue North, St. Paul, MN 55108-611405 8081.

E-mail address: olman002@umn.edu (L.G. Olmanson

0034-4257/$ see front matter 2012 Elsevier Inc. Allhttp://dx.doi.org/10.1016/j.rse.2012.11.023) of rivers and streams.orridors and recreation-

expect that similar relationships exist in owing waters, but comparedwith lakes, rivers and streams pose a more challenging set of problemsin applying remote sensing techniques to assess water quality. First,Minnesota has 93,000 miles (150,000 kmThey are highly important as transportation cKeywords:AirborneHyperspectral imageryImaging spectroscopyMississippi RiverRemote sensingRiverChlorophyllTurbidityMineral suspended sedimentWater qualityInherent optical propertiesCDOM

1. Introductionthose characteristics to develop models to map water quality characteristics in optically complex waters.For phytoplankton related variables (volatile suspended solids (VSS) and chlorophyll a (chl a)), the ratios ofthe scattering peak at the red edge (~700 nm) with the reectance troughs caused by chlorophyll absorptionat ~670 nm and other plant pigment absorption peaks at 592 and 620 nm all were strong predictors of chl aand VSS (r2 values of 0.730.94). The scattering peak at ~700 nm was a strong predictor of variables related towater clarity (total suspended solids (TSS), turbidity and turbidity tube (T-tube)) (r2 values of 0.770.93). Formineral-based variables (nonvolatile suspended solids (NVSS) and the ratio NVSS:TSS), combinations of theTSS and chl a relationships described above were strong predictors (r2 values of 0.730.97) and the most robustbecause thismodel corrects for the scattering of phytoplankton at ~700 nm. Application of themethods to quan-tify spatial variations in water quality for stretches of the Mississippi River and its tributaries indicate thathyperspectral imagery can be used to distinguish andmap key variables under complex IOP conditions, particu-larly to separate and map inorganic suspended sediments independently of chlorophyll levels.

2012 Elsevier Inc. All rights reserved.

levels in large and moderately sized lakes (Olmanson et al., 2011). WeAvailable online 30 December 2012

phyll, suspended solids, turbidity and other measures of water clarity. Our approach identied the spectralcharacteristics that distinguish waters dominated by several inherent optical properties (IOPs), and we usedAccepted 27 November 2012Ground-based sampling undAirborne hyperspectral remote sensing tocharacteristics in large rivers: The Mississ

Leif G. Olmanson a,, Patrick L. Brezonik b, Marvin E.a Department of Forest Resources, University of Minnesota, St. Paul, MN 55108-6112, Uniteb Department of Civil Engineering, University of Minnesota, St. Paul, MN 55108-6112, Unite

a b s t r a c ta r t i c l e i n f o

Article history:Received 27 March 2012

Aircraft-mounted hypersperesolution for use in measu

j ourna l homepage: wwwet al., 2008) to measurehat other satellite sensorsestimates of chlorophyll

urces, University of Minnesota,2, United States. Tel.: +1 651

).

rights reserved.ssess spatial distribution of water qualitypi River and its tributaries in Minnesota

uer a

atestates

l spectrometers were used to collect imagery with high spatial and spectralg optically active water quality characteristics of major rivers of Minnesota.

f Environment

l sev ie r .com/ locate / rsevery large rivers or impoundments of large rivers.Hyperspectral sensors, mounted in small aircraft can collect land-

scape images with high spatial and spectral resolution. Such airbornesystems have been available for over two decades and have been usedfor mineralogical exploration (e.g., Abrams et al., 1977; Clark et al.,1990; Goetz & Srivastava, 1985), aswell as to determine the type, healthand condition of vegetation for environmental quality, forestry andagriculture purposes (e.g. Carroll et al., 2008; Gitelson & Merzlyak,

1996; Gitelson et al., 2002; Haboudane et al., 2002, 2004; Shah et al.,2002, 2004;Wessman et al., 1988). Several publications have describedthe application of such systems to measure water quality conditions inlakes (e.g., Ammenberg et al., 2002; Chipman et al., 2009; Hakvoortet al., 2002; Hoogenboom et al., 1998 and Moses et al., 2011), but onlya few publications (e.g., Senay et al., 2001 and Shaque et al., 2003)have focused on potentially more complex river systems. Phytoplankton,mineral suspended sediment, humic color or combinations of theseconstituents may dominate the optical properties of rivers depending

Land cover in the basin is mixed (37% forest and 21% cropland).In the TCMA (7700 km2) the land cover is 40% agricultural rowcrops, 30% urban and 13% forest. The Minnesota and St. Croix Riversare large tributaries that ow into the Mississippi in the TCMA andincrease the total river ow on average by more than a factor of two(Table 1).

The drainage basin of the Minnesota River covers 42,000 km2, in-cluding large portions of south-central and southwestern Minnesota,as well as small portions of Iowa and South Dakota. The basin has rel-

uguisch

190770170

255L.G. Olmanson et al. / Remote Sensing of Environment 130 (2013) 254265on watershed and ow conditions. Strong relationships were found inprevious studies between chlorophyll concentrations and turbidity orsuspended solids concentrations and reectance data, but noneattempted to separate competing inherent optical properties (IOPs).

In this paper we describe the use of aircraft-mounted hyperspectralsensors on threemajor rivers inMinnesota (Mississippi,Minnesota, andSt. Croix) and some associated oodplain lakes with distinct and com-peting IOPs. Sufcient informationwas acquired in three separate aerialdata acquisitions to evaluate the usefulness of aircraft-mounted remotesensing as a supplement/complement to conventional ground-basedriver monitoring programs. For calibration purposes, water sampleswere collected concurrently with the remote sensing data acquisition,and in order to obtain a range of conditions for calibrations, we focusedour measurements around the conuences of river systems that havedifferent water quality characteristics. Imagery was collected duringAugust of 2004, 2005 and 2007, and each period represented differentow and water quality regimes.

The overall goal was to develop reliable techniques for synopticmeasurements of key indicators of river water quality that can beused to complement data obtained by conventional ground-basedmethods. Our specic objectives were to: (1) identify spectral charac-teristics that distinguish waters dominated by different IOPs indicativeof important water quality characteristics; (2) develop predictive rela-tionships for these characteristics based on their individual spectral(reectance) characteristics (or combinations thereof); (3) determinewhether and how accurately chlorophyll a, total suspended solids(TSS), and nonvolatile suspended solids (NVSS) can be mapped inde-pendently when competing IOPs are dominant and whether theproportion (NVSS/TSS) can be identied and mapped quantiably;(4) develop an overall approach to measure and map water qualityvariables using aircraft-mounted spectrometers; and (5) evaluate theaccuracy and usefulness of aircraft-based hyperspectral remote sensingfor water quality studies of rivers. The work was conducted usingimages collected under clear (cloudless) conditions and processed bymethods similar to those developed for regional assessments of waterclarity and chlorophyll a from satellite imagery (e.g., Olmanson et al.,2008, 2011).

2. Background information: major rivers in Twin CitiesMetropolitan Area

The Mississippi River, which originates in Lake Itasca in northernMinnesota, is a moderate sized river (average ow of 11,700 cfs(3163 m3 s1), Table 1) by river mile 871, where it reaches theMinneapolisSt. Paul (Twin Cities) Metropolitan Area (TCMA). Thedrainage basin of the Mississippi River above the TCMA (~49,000 km2

in area) includes much of central and north central Minnesota.

Table 1River discharge for Minnesota, Mississippi and St. Croix Rivers in the TCMA.

River Site Meandischarge (cfs) 26,604

Ad

Minnesota Jordan 881033% 3Mississippi Anoka 11,70044% 2St. Croix Stillwater 609423% 3atively at to gently rolling hills with highly productive soils and isprimarily (75%) row-crop agricultural land. The river carries a highburden of suspended solids from stream bank and soil erosion thatis attributed to a combination of articial drainage and steep slopesfrom deep incising of the Minnesota River Valley by Glacial RiverWarren (Belmont et al., 2011; Gran et al., 2009; Thorleifson, 1996).

The drainage basin of the St. Croix River covers 20,000 km2 ineastern Minnesota and northwestern Wisconsin. The basin is theleast developed of the three considered here and consists of 50% for-est, 15% grassland or pasture, and 15% agricultural row crops. Amongthe three rivers, the St. Croix has the lowest nutrient and suspendedsolids concentrations but has higher levels of natural coloreddissolved organic matter (CDOM).

Aside from their commercial and recreational importance, the riv-ers play important roles in transporting pollutants downstream. Mostof the treated wastewater from the TCMA enters the Mississippi Riverdownstream of a major treatment plant in St. Paul; among other con-sequences, this leads to high nutrient (nitrogen and phosphorus)concentrations that promote summertime algal blooms in the riverand downstream Lake Pepin (Engstrom et al., 2009).

The water chemistry/quality of these rivers is more complex thanthat of most lakes. Dominance by both phytoplankton, usually mea-sured in terms of chlorophyll a concentrations (sometimes called greenphase) and inorganic sediment, sometimes called brown phase andmea-sured in terms of Secchi depth, turbidity, or suspended solids concentra-tions,may occur depending onowand seasonal dynamics. Typically, theMississippi River accounts for 4045% of ow and 20% of incoming TSSload; theMinnesota River accounts for 2530% of ow but 75% of incom-ing TSS; and the St. Croix River accounts for 2530% ofowbut only 5% ofthe incoming TSS load (Metropolitan Council, 2004). Water quality con-ditions in the three rivers are highly diverse, and conditions in theMissis-sippi River downstream of the conuences with the two tributariesdepend on the relative ows of the rivers and the degree towhichmixingof the waters has occurred.

Water quality impairment issues occur in the Mississippi River as aconsequence of urban and agricultural runoff and inputs of treated mu-nicipal wastewater as the river ows through the TCMA.Major concernsexist about the effects of suspended solids and nutrient loadings on LakePepin, a natural lake in the main channel of the river ~40miles (70 km)downstream of the TCMA (Belmont et al., 2011; Engstrom et al., 2009;Mulla & Sekely, 2009). At a larger scale, there also is concern about con-tributions from Minnesota to the total nutrient load of the MississippiRiver to the Gulf of Mexico (Goolsby et al., 2001; Rabalais et al., 2002a,b; Turner & Rabalais, 1994). Under normal conditions Minnesota con-tributes ~3% of the total nitrogen ux and 2% of the phosphorus ux de-livered to the Gulf of Mexico by the Mississippi River (Alexander et al.,2008).

st 19, 2004arge (cfs) 9130

August 15, 2005discharge (cfs) 7370

August 30, 2007discharge (cfs) 8070

35% 184025% 416052%30% 410056% 193024%35% 143019% 198024%

3. Methods

3.1. Imagery

Imagery was collected by the Center for Advanced Land Manage-ment Information Technologies (CALMIT, University of Nebraska,Lincoln) using AISA (Airborne Imaging Spectrometer for Application)hyperspectral imaging systems in a Piper Saratoga airplane using pro-cedures described by Chipman et al. (2009). The CALMIT systemevolved over the course of the study, as new equipment, softwareand techniques became available. Having different image characteris-tics complicated our analysis. AISA sensors collect spectral-radiancedata and compile imagery in the visible and near infrared portionsof the electromagnetic spectrum. They are push-broom systemswith a global positioning and inertial navigation system that simulta-neously record the aircraft position and attitude, along with upwell-ing radiance and downwelling solar irradiance. These raw dataproducts are used to create geospatially and radiometrically correctedimagery.

During the rst acquisition period (August 19, 2004), an AISA-Classic sensor collected hyperspectral imagery in 16 (for 1 and 3 mresolution imagery) or 24 (for 2 m imagery) narrow bands (~5 nmbandwidth) selected to measure key water quality characteristics oversix segments (Fig. 1) that covered ~65 km of the Mississippi River andselected tributaries. A collaborative sampling effort involving crewsfrom the Minnesota Pollution Control Agency, Metropolitan Council,Minnesota Department of Natural Resources, Minnesota Departmentof Agriculture and the University of Minnesota collected water samplesand took in situ measurements including GPS coordinates at ~39 loca-tions along the river concurrently with the imagery acquisition.

During the second acquisition (August 15, 2005), an AISA-Eaglehyperspectral imager collected hyperspectral data at a spatial resolution

of 2 m in 86 contiguous bands: 2.5 nm bandwidth from 435 to 730 nm,20 nmbandwidth from730 to 900 nm, and 30 nm from900 to 960 nm.The data were recorded for 10 segments over a 60 km stretch of theMississippi River from Spring Lake (south of St. Paul) to Lake Pepin(Fig. 1). Sampling crews from the Minnesota Pollution Control Agencyand Metropolitan Council concurrently collected water samples at 26locations along the river.

Image acquisitions on August 30, 2007 also used an AISA-Eaglehyperspectral imager to collect hyperspectral data at 2 m spatialresolution in 97 contiguous bands (2.5 nm bandwidth from 435 to730 nm and 10 nm bandwidth from 730 to 950 nm) in 12 segmentsover a 100 km stretch of the Mississippi River from the Rum Rivernorthwest of the TCMA to downstream of the St. Croix River (Fig. 1).Sampling crews from the Minnesota Pollution Control Agency andthe Metropolitan Council simultaneously collected water samplesfrom 26 sites.

3.2. Water quality measurements

Because of the dynamic nature of rivers, sampling crews were onthe river to collect water samples and take in situ measurements asclose as possible to the time the aircraft collected imagery. All sam-ples were collected within a few hours of the remote sensing data ac-quisition and away from mixing zones subject to rapid temporal andspatial changes. Field measurements at each site included water clar-ity by turbidity tube (Myre & Shaw, 2006) and, if possible, Secchi disk(SD); high ow rates and high turbidity levels limit the feasibility ofSD measurements in some rivers. Samples were collected for turbidi-ty, total suspended solids (TSS), volatile suspended solids (VSS), chlo-rophyll a (chl a) and phaeophytin (Pheo) and were analyzed bylaboratories of the Metropolitan Council using standard methods:

256 L.G. Olmanson et al. / Remote Sensing of Environment 130 (2013) 254265Fig. 1. Hyperspectral imagery ight lines for 2004, 2005 and 2007.

polygons were imported into Microsoft Excel, ratios and differences for

257L.G. Olmanson et al. / Remote Sensing of Environment 130 (2013) 254265band combinations were calculated, and data from the ground-basedmeasurements and samples were linked to the appropriate imageryAPHA-10200h (Greenberg et al., 1992), ASTM D3731-87 (ASTM, 1994),and USEPA-445 (U.S. EPA, 1992).

3.3. Image processing and classication

Image classication procedures were based on the methods ofOlmanson et al. (2001) and Kloiber et al. (2002), modied for thehyperspectral nature of the imagery and advances in software andcomputer hardware that enabled simpler or improved procedures.We used Leica Geosystems ERDAS Imagine and Esri ArcGIS forimage processing. Acquiring a representative sample for each sam-pling site from the image was a primary objective.

Each ight line image was corrected geospatially and radiometrical-ly to at platform reectance using the raw data from the AISA system,but additional image preprocessing was necessary. Although the auto-mated geometric correction process yielded close results, additionalgeometric correctionwas performed using ~20well-distributed groundcontrol points (GCPs) from available orthorectied aerial photographywith a positional accuracy (RMSE) on the order of 1 pixel (13 m).Nearest neighbor resampling was used to preserve the original pixelvalues.

Atmospheric correction was conducted on the images acquiredin 2007 using the FLAASH module in ENVI and standard atmos-pheric variables for mid-latitude summer and water retrieval usingthe band at 820 nm. Radiometric normalization, necessary for only afew nonconforming images which were identied by visuallyinspecting the imagery after mosaicking the images, was accomplishedusing pixel spectral data selected from the overlap area of both(conforming and nonconforming) images representing differentland cover and water areas. Regression analyses were performedwith the base (conforming) image as the dependent variable andthe image needing correction (nonconforming) as the independentvariable. The resulting models were used to normalize thenonconforming image to the base image.

Once image preprocessing was complete, a water-only imagewas produced by performing an unsupervised classication methodbased on ISODATA clustering. Because water has spectral characteris-tics distinct from terrestrial features, water pixels were groupedinto one or a few distinct classes that could be easily identied.We masked out terrestrial features to create a water-only image,performed an unsupervised classication on this image, and generat-ed spectral signatures of each class. We used these signatures, alongwith the locations where the pixels occurred, to differentiate classescontaining open water and shallow water (where bottom sedimentsand/or macrophytes affected spectral response). These areas tend tohave high spatial variability compared to open-water portions of therivers. Based on this analysis, we removed the affected pixels.

Next, the spectral-radiometric data from the open water imageswere extracted from the areas immediately around the sampling lo-cations to develop relationships with measured water quality data.For this purpose, we used the imagery and the GPS locations of eachsample to delineate a polygon in a spectrally-radiometrically similararea (identied visually using different band combinations stretchedto range of water only pixels) around each sample location. Kloiberet al. (2002) found that using at least nine pixels to extract spectralradiometric data from Landsat 30 m imagery improved correlationstrength with eld data. The polygons we used had a minimum of200 pixels in narrow portions of the rivers or areas with high hetero-geneity and up to 1700 pixels in wider, more homogeneous areas.The signature editor in ERDAS Imagine was used to extract spectralradiometric data from the polygons. Mean band values from thesamples.To develop models to predict water quality variables from the im-agery products, we performed step-wise forward regressions of theband datasets using the JMP 9 64-bit version statistical software pack-age of SAS Institute, Inc. Because different band sets had been usedin each acquisition period, a subset of 18 well-positioned bands(16 bands for 2004 1 and 3 m imagery) that occur at the same oralmost the same central wavelength for each dataset was used. Thebasis for selection of the bands is described in the Results anddiscussion section. Raw and log-transformed values of each waterquality variable were the dependent variables and single bands,band ratios and band differences were the independent variables.From these results, we used the one or two independent variablesthat were consistent for all three datasets and that contributedmost to the regression t and applied them in a second multiple re-gression, as described by Olmanson et al. (2011). The models devel-oped for each variable and each image product were applied to eachpixel to create pixel-level water quality maps.

4. Results and discussion

4.1. Spectral characteristics of river water

Hyperspectral imagery can provide sufcient information to recon-struct the visiblenear IR reectance spectra of water bodies that inturn can be used to identify the spectral characteristics (e.g., reectancepeaks and troughs) that distinguish waters dominated by the inherentoptical properties (IOPs) indicative of different water quality variables.Fig. 2 shows the characteristic reectance spectra extracted from theAugust 30, 2007 hyperspectral imagery. Spectra from four locationsrepresent phytoplankton dominated waters: (1) hypereutrophic LakeConley (chl a=200 g/L; TSS=48 mg/L with 42% NVSS), which is aMississippi River backwater lake upstream of the conuence with theSt. Croix River; (2) hypereutrophic Pig's Eye Lake (chl a=230 g/L;TSS=51 mg/L with 58% NVSS), a backwater lake just downstreamfrom the Metropolitan Wastewater Treatment Plant (Metro); (3) theMississippi River upstream of the Rum River at river mile 872 (chla=78 g/L; TSS=13 mg/L with ~60% NVSS); and (4) the MississippiRiver upstream of the Minnesota River at river mile 848 (chl a=45 g/L; TSS=15 mg/L with ~40% NVSS). Flow at the last locationwas unusually low at the time of sampling because of ow restrictionsat St. Anthony Falls locks and dams (river mile 853.9) in Minneapolisnecessitated by a major bridge collapse downstream of the lock anddam (http://en.wikipedia.org/wiki/I-35W_Mississippi_River_bridge).The reectance spectra of these waters show characteristic absorptionby chlorophyll and other plant pigments in the blue (400500 nm)and red (600700 nm) wavelengths, which results in low reectance(i.e., troughs) at ~430 nm (not shown because of noise in that region)and ~670 nm and a phytoplankton scattering peak at ~700 nmresulting from a combination of scattering and decreasing absorp-tion by chlorophyll and increasing absorption by water (Vasilkov& Kopelevich, 1982).

The Minnesota River and the mixed waters of the Minnesota andMississippi Rivers at St. Paul (river mile 839) represent sediment-dominated waters with TSS concentrations of 61 and 45 mg/L (87 and82% NVSS), respectively. Although chl a levels were 40 and 55 g/L,respectively, they were not high enough to overcome the dominanceby suspended sediment, and their reectance spectra lack the absorp-tion characteristics (troughs at ~430 and ~670 nm) of chlorophyll andhave relatively high reectances in the green and red region especiallyfrom 560 to 700 nm.

Cenaiko Lake, a small, mesotrophic water body east of the Missis-sippi River at river mile 865, represents relatively clear water. It hadlow concentrations of suspended matter (TSS=3 mg/L; NVSS=2 mg/L) and low chl a (8 g/L). Although quantitative information isnot available on CDOM, the lake water was not visibly stained brown.

Its reectance spectrum is characterized by a reectance peak near

hyp

258 L.G. Olmanson et al. / Remote Sensing of Environment 130 (2013) 254265570 nm and declining reectance at higher wavelengths. The lack ofthe characteristic peaks and troughs caused by chl a and theplankton/TSS scattering peak in the red and NIR range is notable. Fi-nally, the St. Croix River with low-to-moderate chl a (22 g/L) and lowconcentrations of suspended matter (TSS=2 mg/L; NVSS=1 mg/L)represents a water body dominated by humic color (CDOM-richwater; color=30 chloroplatinate units, CPU). It had low reectanceacross the entire spectrum, with characteristic absorption by CDOM inthe lower wavelengths; very small reectance peaks are apparent inthe spectrum at 570 and 700 nm.

The distinctive features of reectance spectra for waters dominat-ed by different optically-active constituents (chl a, inorganic sedi-ment, CDOM) provide insights regarding wavelength bands likely tobe useful for predictive models of these characteristics. We were un-able to nd similar reectance spectra in the literature for lakes withIOPs dominated by both phytoplankton and inorganic suspendedsolids, but Han (1997) conducted tank experiments with varying ad-ditions of soil sediment to clear and algae-laden waters and foundhyperspectra similar to those shown in Fig. 2. In general, addition of

Fig. 2. Characteristic reectance spectra extracted from 2007soil sediment to water containing high concentrations of algae causedsmaller increases in reectance between 400 and 700 nm than simi-lar additions to water low in algae, and this was explained by the ab-sorption effects of chlorophyll.

Successful models using spectral-radiometric data to estimate waterquality variables should usewavelengths that identify key spectral char-acteristics without interference from competing optical features. Forchlorophyll, this means that algorithms commonly used for ocean CaseI waters (reectance in the blue and green reectance regions) willnot work well for inland (Case II) waters because these waters areinuenced by suspended solids and CDOM derived from their water-sheds making them optically more complex (Morel & Prieur, 1977).CDOM and inorganic sediment have overlapping absorption featureswith chl a in the blue region. Successful chlorophyll models for thesewaters likely will use absorption characteristics from the red wave-lengthsthe reectance trough at ~670 nm and scattering by phyto-plankton at the red edge ~700 nm, where absorption by CDOM andsuspended solids is minimal (Gitelson, 1992; Moses et al., 2011). Theratio ~700 nmwith ~670 nmhas been shown to have a strong relation-ship with chl a in a variety of waters (including lakes, estuaries and riv-ers) over a wide range of concentrations (0.1 to 350 g/L; Matthews,2011) and has been reported several times (Duan et al., 2007; Gitelsonet al., 1993; Gons, 1999; Menken et al., 2006; Mittenzwey et al., 1992;Moses et al., 2009).Successful models for turbidity and TSS in optically complex waters(where both phytoplankton and inorganic sedimentmay dominate IOPfeatures) would avoid the absorption characteristics of chlorophyll inthe red and CDOM in the blue region and use the scattering peak at~700 or band combinations in the NIR and green regions. The scatteringpeak at ~700 was found to be strongly correlated with TSS by Hrmet al. (2001), Kallio et al. (2001) and Koponen et al. (2007) and for tur-bidity by Senay et al. (2001), and the difference (710740) was founduseful by Shaque et al. (2003). The scattering peak at ~700 by itselfwas also found to work well for NVSS by Ammenberg et al. (2002).

CDOMmodels usually do not use the strong absorption of CDOM inbluewavelengths because chlorophyll and suspendedmatter also affectspectral reectance at these wavelengths. Instead, CDOM (or humiccolor) usually is estimated using ratios of reectance in the green andred regions (typically at wavelengths>550 nm), where CDOM absorp-tion of light still occurs but is much less intense. The physical basis forthese empirical relationships is not well understood.

4.2. Model development

erspectral imagery with 18 selected spectral bands indicated.Having different band sets and radiometric characteristics andcorrections for each set of image acquisitions complicated our analy-sis. Therefore, each acquisition dataset was analyzed individually, anda set of 18 well-positioned bands (16 bands for 2004 3-m imagery)that occur at the same or almost the same central wavelength foreach dataset was used. Because of poor quality 2004 imagery outsideof the TCMA (tree shadows and bottom effects are apparent in theCrow River images and differences in radiometric correction (dueto variable cloud cover) are apparent in the Blue Earth images),only the data from the TCMA for this year were included in the anal-ysis. The bands selected for analysis were derived based on the spec-tral characteristics of the data (Fig. 2) and bands used successfully byothers from the literature, as reviewed by Matthews (2011).

Several models developed from the images showed strong relation-ships between log transformed andnon-transformedwater quality data(T-tube, turbidity, TSS, VSS, NVSS and chl a) and reectance values fromthe hyperspectral data; see Table 2 for a statistical summary for themodels, including ranges of the variables, bands in the best-t relation-ships (consistent for all three datasets), r2, RMSE, and number of datapoints. Color data were collected during the 2004 acquisition, but therange in the data was too small (2040 CPU) to allow development ofpredictive relationships. The non-transformed water quality data werenot normally distributed for some variables, especially chl a, but the

log transformed data were or were nearly so, thus meeting theassumptions of regression. Olmanson et al. (2011) also found thatlog-transformed data produced models with improved t in theimportant chl a range of 020 g/L. Therefore, models based onlog-transformed data were used for map development.

Chl a, the most common water quality variable derived from highspectral resolution remote sensing, had strong relationships with theratio ~700:~670 nm for all three datasets, as also was found to bethe case in several previous studies (see previous discussion). Thestrength of this relationship is due to backscattering by phytoplank-ton that causes a peak ~700 nm in phytoplankton dominated watersand the chl a absorption maximum near 670 nm. For this dataset

even with a complex combination of IOP features, the 705:670 ratiowas strongly correlated with chl a, with r2 of 0.750.93 for thethree years of data. Two models that also used the peak at ~700 andother plant pigment absorption features seen in Fig. 2 also workedwell (Table 2): 705:592 and 705:620; e.g., see Fig. 3a. These wave-lengths are not in strong chl a absorption regions (620 nm is in theregion of phycocyanin absorption and 592 nm is in the region ofB-phycoerythrin absorption) and thus cannot be reliable for broadapplication even though they performed statistically well for thisdataset. The strength of these band ratios for chl a in this datasetstrongly suggests the presence of cyanobacteria in these watersbecause phycocyanin and B-phycoerythrin are both indicators of

Table 2Calibration statistics for water quality models.

LN of variable Bands 2004 2005 2007

N Range r2 RMSEa N Range r2 RMSEa N Range r2 RMSEa

T-tube (cm) 705 24 8120 0.91 0.25 19 795 0.77 0.24 19 12160 0.83 0.26Turbidity (NTU) 705 25 225 0.88 0.29 20 250 0.77 0.29 19 346 0.92 0.23TSS (mg/L) 705 25 480 0.83 0.28 20 495 0.77 0.32 19 261 0.93 0.25NVSS (mg/L) 705 & 705/670 25 363 0.95b 0.24 20 060 0.85b 0.27 19 153 0.97b 0.22NVSS/TSS (%) 705 & 705/620 25 5083 0.80b 0.05 20 083 0.73b 0.1 19 3087 0.91b 0.06VSS (mg/L) 705/670 25 117 0.80 0.26 20 481 0.94 0.17 19 128 0.73 0.25Chl a (g/L) 705/670 25 14210 0.75 0.43 20 31830 0.93 0.21 19 8230 0.83 0.32Chl a (g/L) 705/620 25 14210 0.76 0.38 20 31830 0.91 0.21 19 8230 0.92 0.21Chl a (g/L) 705/592 25 14210 0.90 0.25 20 31830 0.92 0.19 19 8230 0.85 0.27

a LN(variable).b R2.

259L.G. Olmanson et al. / Remote Sensing of Environment 130 (2013) 254265Fig. 3. Scatter plots of in situ log transformed biophysical quantities versus corresponding(c) NVSS, (d) T-tube. For NVSS the multiple regression model from Table 2 was used.band, band-ratio or model values for the August 30, 2007 imagery: (a) chl a, (b) TSS,

cyanobacteria. The RMSE values reported for the predictive equationsin Table 2 are in LN(variable) and thus indicate that chl a can be esti-mated with an accuracy of ~21% (based on the 2005 and 2007 data).Some of the uncertainty reected in the RMSE values likely arisesfrom uncertainties in the chl a analyses themselves, as well as fromdifferences between the water mass sampled by the ground crewsand that sampled by the imagery (caused by differences in thetiming of the two sampling methods).

TSS includes both organic and inorganic solids which potentiallyhave a variety of different combinations of IOP features. Remote sens-ing algorithms for TSS reported in the literature (Doxaran et al., 2005;Gitelson et al., 1993; Koponen et al., 2007) thus are less consistentand more condition-specic dependent on the dominant optical fea-tures. For our dataset strong relationships were found between TSSand reectance in the band centered at 705 nm (Table 2); seeFig. 3b. This model avoids the absorption characteristics of chloro-phyll but uses the well-known scattering at ~700 nm of suspendedmaterial (including plankton) in water. As discussed earlier, otherworkers have found that the reectance peak at ~700 works wellfor TSS (Hrm et al., 2001; Kallio et al., 2001; Koponen et al., 2007).

VSS, the mass of combustible (organic) suspended material thatpresumably is plankton and microbial detritus, is generally correlatedwith chl a. The relationships between VSS and band ratios thatworked well for chl a also were strong (Table 2): 705:592, r2=0.800.95; 705:620, r2=0.780.95; 705:670, r2=0.730.94. NVSS,the mass of inorganic suspended material, had strong relationshipswith reectance in the band centered at 705 nm (r2 values of

The ratio NVSS:TSS (the fraction of TSS that is inorganic or mineral)could be useful to identify the dominant IOP (planktonic- versusmineral-based turbidity) in river waters. The predictive model for thisratio that worked best for all datasets used the band centered at705 nm, which worked well for TSS, and the band ratio 705:620,which worked well for chl a, (Table 2). The RMSE values reported forthe predictive equations in Table 2 indicate that NVSS:TSS can be esti-mated within ~510%.

Finally, turbidity and water clarity as measured by T-tube are gen-erally related to TSS. Strong relationships were found for both turbid-ity and T-tube (Table 2; see Fig. 3d for an example) with reectancevalues for the band centered at 705 nm. These models are similar tothe best-t models for TSS.

4.3. Water quality maps

Using the water quality models described above, we created pixel-level maps that captured the spatial heterogeneity of water quality inthe river systems we studied. During the rst image acquisition(August 19, 2004), ows of the Minnesota and St. Croix Rivers wereslightly higher than the ow of the Mississippi River (Table 1), andthe Minnesota River IOP features were dominated by phytoplankton(green phase). As Fig. 4 shows, the high phytoplankton waters of theMinnesota River dominated water quality conditions in the MississippiRiver downstream the conuence with the Minnesota River duringthis time. Similarly, Fig. 5 shows that the phytoplankton (green

260 L.G. Olmanson et al. / Remote Sensing of Environment 130 (2013) 2542650.730.96). As discussed previously, this peak was found to be strong-ly correlated to NVSS by Ammenberg et al. (2002); however, this re-lationship could be condition-specic, depending on the dominantIOPs, and may not be robust in waters with low inorganic sediments.Therefore, for a more robust model we used the band centered at705 nm, which worked well for TSS, and the band ratio 705:670,which worked well for chl a. This model corrects for the scatteringcaused by phytoplankton at 700 nm and had even higher R2 values(Table 2); see Fig. 3c. The RMSE values reported for the predictiveequations in Table 2 indicate that NVSS can be estimated with an ac-curacy of ~2227%.Fig. 4. Maps of turbidity and chl a levels near the conuence ofphase) waters of the Mississippi River were diluted by the relative-ly clear waters of the St. Croix River downstream of the conuenceof those two rivers.

During the second image acquisition (August 15, 2005), theMississippi River dominated ow (Table 1) andwater quality conditions,and the main channel had relatively low phytoplankton and sedimentlevels. However, as Fig. 6 shows, the backwater areas were dominatedby algal blooms. In addition, the gure shows that sediment picked upfrom the channel around Sturgeon Lake apparently settled as the riverow velocity decreased in the upper part of Lake Pepin, and thelow suspended sediment concentrations (clearer water) enabled thethe Minnesota River and Mississippi River, August 19, 2004.

e of

261L.G. Olmanson et al. / Remote Sensing of Environment 130 (2013) 254265development of a large increase in phytoplankton further down-channelin the lake.

During the acquisition period onAugust 30, 2007, the largerow fromthe Minnesota River (brown phase caused by a large rainfall event fromAugust 18 through August 20, 2007 that caused severe ooding insoutheastern Minnesota) dominated conditions in the Mississippi Riverdownstreamof the conuence (Fig. 7) becauseowupstream in theMis-sissippi (green phase) was restricted by the closing of the St. AnthonyFalls locks and dams because of the I-35 bridge collapse. Further down-

Fig. 5. Maps of turbidity and chl a levels near the conuencstream the relatively clear waters of the St. Croix River diluted theseconditions in the stretch of the river downstream of its entry to the Mis-sissippi (Fig. 8).

Fig. 6. Maps of turbidity and chl a levels in the Mississippi River from Spring LakeThe capability of airborne hyperspectral imagery to capture ne-scale variations in water quality conditions is illustrated by themaps in Fig. 9, which show the transition from sediment-dominatedwaters of the Mississippi River to phytoplankton-dominated waterof Pig's Eye Lake. Although turbidity is similar in the river and Pig'sEye Lake (with some variations within both water bodies (Fig. 9a)),phytoplankton is dominant in most of the lake and is much lower inthe river (Fig. 9b). An inux of water with a high load of inorganicsediment (i.e., high NVSS:TSS) from the river is evident in the south-

the Mississippi River and St. Croix River, August 19, 2004.western portion of the lake (Fig. 9c). The transition fromphytoplankton-dominated water at location a (Fig. 9b) to inorganic sediment-dominated water at location e is captured in the reectance spectra

to Lake Pepin and the conuence with the St. Croix River, August 15, 2005.

extracted from the imagery (Fig. 10). Absorption characteristics of a small area. Airborne sensors are well-suited to capture conditions

Fig. 7. Maps of turbidity and chl a levels near the conuence of the Minnesota River and Mississippi River, August 30, 2007.

262 L.G. Olmanson et al. / Remote Sensing of Environment 130 (2013) 254265chlorophyll are clearly visible at location a but diminish toward loca-tion e. This illustrates the vast amount of information available froma single image that would not be seen by conventional monitoring,which probably would involve only one sample for the entire area.

5. Conclusions

Airborne hyperspectral remote sensing was effective in capturinga much more comprehensive picture of river water quality thanis obtained by typical ground-based sampling, in which large areasare represented by a single sampling point that may represent onlyFig. 8. Maps of turbidity and chl a levels in the Mississippi River from the Ruin systems with extensive backwater areas (like the MississippiRiver and its backwater areas and tributaries in east-centralMinnesota)that are difcult to reach by conventional samplingmethods. Capturingthe spatial heterogeneity of such systems by conventional methodswould be prohibitively time consuming and expensive. High resolutionhyperspectral imagery is especially useful forwaters with complex IOPsbecause it can distinguish between the contributions of phytoplanktonand inorganic suspended sediments to water clarity.

The spectral characteristics of waters dominated by various IOPswere used to develop models to map water quality characteristics inoptically complex waters. Light absorption in the red wavelengths bym River to past the conuence with the St. Croix River, August 30, 2007.

orga

263L.G. Olmanson et al. / Remote Sensing of Environment 130 (2013) 254265chlorophyll in phytoplankton-dominated waters contrasted withthe lack of these absorption characteristics in inorganic sediment-dominated waters. The scattering peak at ~700 nm, above the regionwhere plant pigments absorb light, was a strong predictor for waterclarity variables (e.g., TSS and turbidity), and ratios of this scatteringpeak with the reectance troughs caused by chlorophyll absorptionat ~670 nm and absorption by other pigments at 592 and 620 nmwere strong predictors of chl a and VSS (the last two wavelengthsbecause plankton blooms likely were dominated by cyanobacteria).

Fig. 9. Maps of Pig's Eye Lake showing the transition from conditions dominated by in(c) NVSS/TSS; August 30, 2007.The models that worked for these complex waters also workedwell for other researchers under varying water quality regimes andthus appear to be generally robust predictors. Our results for Pig'sEye Lake indicate that such predictive relationships can distinguish

Fig. 10. Reectance spectra of the transition zone for conditions dominated by inorganic seLake, August 30, 2007 (Fig. 9B). Tabulated chl a, turbidity and NVSS/TSS values were calculkey water quality variables under complex IOP conditions, i.e., inor-ganic suspended sediment and chlorophyll levels can be mappedindependently. For NVSS and the ratio NVSS:TSS, we found that themodel that combines the TSS and chl a relationships was themost ro-bust because it corrects for scattering by phytoplankton at ~700 nm.

Although airborne hyperspectral sensing has many advantagesfor river water quality assessments, it does have limitations, includingcost, which may be too high for local agencies with small budgets,and the need to collect data when atmospheric conditions are clear.

nic sediment to conditions dominated by phytoplankton: (a) turbidity, (b) chl a, andHaze, cloud cover and cloud shadows can alter spectral-radiometricresponses, and such conditions should be avoided. Imagery thus maynot be available when needed. Although hyperspectral data are usefulin developingwater qualitymodels, they are not needed for operational

diment in the Mississippi River to conditions dominated by phytoplankton in Pig's Eyeated from reectance spectra using the best predictive models (Table 2).

Science and Technology, 45(20), 88048810, http://dx.doi.org/10.1021/es2019109.Carroll, M. W., Glaser, J. A., Hellmich, R. L., Hunt, T. E., Sappington, T. W., Calvin, D., et al.

tions from the sediments of Lake Pepin. Journal of Paleolimnology, 41, 563588,

264 L.G. Olmanson et al. / Remote Sensing of Environment 130 (2013) 254265http://dx.doi.org/10.1007/s10933-008-9292-5.Gitelson, A. (1992). The peak near 700 nm on reectance spectra of algae and water:

Relationships of its magnitude and position with chlorophyll concentration. Inter-national Journal of Remote Sensing, 13(17), 33673373.

Gitelson, A., Garbuzov, G., Szilagyi, F., Mittenzwey, K., Karnieli, A., & Kaiser, A. (1993).Quantitative remote sensing methods for real-time monitoring of inland watersquality. International Journal of Remote Sensing, 14, 12691295.

Gitelson, A. A., & Merzlyak, M. N. (1996). Signature analysis of leaf reectance spectra:Algorithm development for remote sensing of chlorophyll. Journal of Plant Physiology,148, 494500.(2008). Use of spectral vegetation indices derived from airborne hyperspectral imag-ery for detection of European corn borer infestation in Iowa corn plots. Journal of Eco-nomic Entomology, 101(5), 16141623, http://dx.doi.org/10.1603/0022-0493(2008)101[1614:UOSVID]2.0.CO;2.

Chipman, J. W., Olmanson, L. G., & Gitelson, A. A. (2009). Remote sensing methods forlake management: A guide for resource managers and decision-makers. Developedby North American Lake Management Society in collaboration with DartmouthCollege, University of Minnesota, University of Nebraska and University of Wisconsinfor the U.S. Environmental Protection Agency.

Clark, R. N., King, T. V. V., Klejwa, M., Swayze, G. A., & Verga, N. (1990). High spectralresolution reectance spectroscopy of minerals. Journal of Geophysical Research Solid Earth and Planets, 95, 1265312680, http://dx.doi.org/10.1029/JB095iB08p12653.

Doxaran, D., Cherukuru, R., & Lavender, S. (2005). Use of reectance band ratios toestimate suspended and dissolved matter concentrations in estuarine waters.International Journal of Remote Sensing, 26, 17631770, http://dx.doi.org/10.1080/01431160512331314092.

Duan, H. T., Zhang, Y. Z., Zhan, B., Song, K. S., & Wang, Z. M. (2007). Assessment ofchlorophyllA concentration and trophic state for Lake Chagan using Landsat TMand eld spectral data. Environmental Monitoring and Assessment, 129, 295308,http://dx.doi.org/10.1007/s10661-006-9362-y.

Engstrom, D. R., Almendinger, J. E., & Wolin, J. A. (2009). Historical changes in sedimentand phosphorus loading to the Upper Mississippi River: Mass-balance reconstruc-monitoring. Our ndings show that only a few key spectral bandsare needed, and a multispectral system with well-positioned bandsthat capture key spectral characteristics and high spatial resolution(~13 m) would be sufcient. Technical advancements in satellites orunmanned drones with these characteristics may reduce the costs ofdata collection and make remote sensing a cost-effective and routinelyused technique for river water quality in coming years.

Acknowledgments

Support for data acquisition, processing and analysis was providedby funding from the Legislative-Citizen Commission on MinnesotaResources (LCCMR) Environment and Natural Resources Trust Fundto the Minnesota Pollution Control Agency and the University ofMinnesota Agricultural Experiment Station. For assistance with elddata coordination, collection and sample analysis, special thanks goto Bruce Wilson (Minnesota Pollution Control Agency), Steve Kloiber(Metropolitan Council), and Kent Johnson (Metropolitan Council)and sampling crews from the Minnesota Pollution Control Agency,Metropolitan Council, Minnesota Department of Natural Resources,and Minnesota Department of Agriculture, which were dispatchedon short notice to collect water samples. Weather forecasts fromKARE 11 meteorologist Jonathan Yuhas were invaluable in planningfor data acquisitions under clear conditions.

References

Abrams, M. J., Ashley, R., Rowan, L. C., Goetz, A. F. H., & Kahle, A. B. (1977). Mappingof hydrothermal alteration in the Cuprite mining district, Nevada, using aircraftscanner imagery for the 0.462.36 m spectral region. Geology, 5, 713718.

Alexander, R. B., Smith, R. A., Schwarz, G. E., Boyer, E. W., Nolan, J. V., & Brakebill, J. W.(2008). Differences in phosphorus and nitrogen delivery to the Gulf of Mexicofrom the Mississippi River basin. Environmental Science and Technology, 42(3),822830, http://dx.doi.org/10.1021/es0716103.

American Society of Testing Materials (1994). ASTM standards. Water and environmentaltechnology (I), Vol. 11.01, West Conshohocken, PA: ASTM International.

Ammenberg, P., Flink, P., Lindell, T., Pierson,D., & Strombeck, N. (2002). Bio-opticalmodel-ling combined with remote sensing to assess water quality. International Journal ofRemote Sensing, 23(8), 16211638, http://dx.doi.org/10.1080/01431160110071860.

Belmont, P., Gran, K. B., Schottler, S. P., Wilcock, P. R., Day, S. S., Jennings, C., et al. (2011).Large shift in source of ne sediment in the Upper Mississippi River. EnvironmentalGitelson, A. A., Zur, Y., Chivkunova, O. B., &Merzlyak,M.N. (2002). Assessing carotenoid con-tent in plant leaves with reectance spectroscopy. Photochemistry and Photobiology, 75,272281.

Goetz, A. F. H., & Srivastava, V. (1985). Mineralogical mapping in the Cuprite miningdistrict, Nevada. Proceedings, Airborne Imaging Spectrometer Data Analysis Work-shop. Jet Propulsion Laboratory. (pp. 2231) Pasadena, California: JPL Publication85-41.

Gons, H. J. (1999). Optical teledetection of chlorophyll a in turbid inland waters. Environ-mental Science and Technology, 33, 11271132, http://dx.doi.org/10.1021/es9809657.

Goolsby, D. A., Battaglin, W. A., Aulenbach, B. T., & Hooper, R. P. (2001). Nitrogen inputto the Gulf of Mexico. Journal of Environmental Quality, 30, 329336.

Gran, K. B., Belmont, P., Day, S. S., Jennings, C., Johnson, A., Parker, G., et al. (2009). Geo-morphic evolution of the Le Sueur River, Minnesota, and implications for currentsediment loading. In L. A. James, S. L. Rathburn, & G. R. Whittecar (Eds.),Managementand restoration of uvial systemswith broad historical changes and human impacts. Geol.Soc. Amer. Spec. Paper, 451. (pp. 119130), http://dx.doi.org/10.1130/2008.2451(08).

Greenberg, A. E., Clesceri, L. S., & Eaton, A. D. (1992). Standard methods for the examina-tion of water and wastewater (18th Ed.). Washington, D.C.: American Public HealthAssociation.

Haboudane, D., Miller, J. R., Pattey, E., Zarco-Tejada, P. J., & Strachan, I. B. (2004).Hyperspectral vegetation indices and novel algorithms for predicting green LAI ofcrop canopies:Modeling and validation in the context of precision agriculture. RemoteSensing of Environment, 90(3), 337352, http://dx.doi.org/10.1016/j.rse.2003.12.013.

Haboudane, D., Miller, J. R., Tremblay, N., Zarco-Tejada, P. J., & Dextraze, L. (2002). Inte-grated narrow-band vegetation indices for prediction of crop chlorophyll contentfor application to precision agriculture. Remote Sensing of Environment, 81(23),416426.

Hakvoort, H., De Haan, J., Jordans, R., Vos, R., Peters, S., & Rijkeboer, M. (2002). Towardsairborne remote sensing of water quality in The Netherlands Validation anderror analysis. ISPRS Journal of Photogrammetry and Remote Sensing, 57, 171183,http://dx.doi.org/10.1016/S0924-2716(02)00120-X.

Han, L. (1997). Spectral reectance with varying suspended sediment concentration inclear and algal-laden waters. Photogrammetric Engineering and Remote Sensing, 63,701705.

Hrm, P., Vepsalainen, J., Hannonen, T., Pyhalahti, T., Kamari, J., Kallio, K., et al. (2001).Detection of water quality using simulated satellite data and semi-empirical algo-rithms in Finland. Science of the Total Environment, 268, 107121.

Hoogenboom, H. J., Dekker, A. G., & De Haan, J. F. (1998). Retrieval of chlorophyll andsuspended matter in inland waters from CASI data by matrix inversion. CanadianJournal of Remote Sensing, 24(2), 144152.

Kallio, K., Kutser, T., Hannonen, T., Koponen, S., Pulliainen, J., Vepsalainen, J., et al.(2001). Retrieval of water quality from airborne imaging spectrometry ofvarious lake types in different seasons. Science of the Total Environment, 268, 5977,http://dx.doi.org/10.1016/S0048-9697(00)00685-9.

Kloiber, S. M., Brezonik, P. L., Olmanson, L. G., & Bauer, M. E. (2002). A procedure forregional lake water clarity assessment using Landsat multispectral data. RemoteSensing of Environment, 82(1), 3847, http://dx.doi.org/10.1016/S0034-4257(02)00022-6.

Koponen, S., Attila, J., Pulliainen, J., Kallio, K., Pyhlahti, T., Lindfors, A., et al. (2007). Acase study of airborne and satellite remote sensing of a spring bloom event in theGulf of Finland. Continental Shelf Research, 27, 228244, http://dx.doi.org/10.1016/j.csr.2006.10.006.

Matthews, M. W. (2011). A current review of empirical procedures of remote sensingin inland and near-coastal transitional waters. International Journal of Remote Sensing,32:21, 68556899, http://dx.doi.org/10.1080/01431161.2010.512947.

Menken, K. D., Brezonik, P. L., & Bauer, M. E. (2006). Inuence of chlorophyll andcolored dissolved organic matter (CDOM) on lake reectance spectra: Implicationsfor measuring lake properties by remote sensing. Lake and Reservoir Management,22, 179190.

Metropolitan Council (2004). Regional progress in water quality analysis of water qualitydata from 1976 to 2002 for the major rivers in the Twin Cities. Metropolitan CouncilEnvironmental Services regional report.

Minnesota Pollution Control Agency (2011). Minnesota's impaired waters and TMDLs:Impaired waters list. http://www.pca.state.mn.us/index.php?option=com_k2&Itemid=250&id=859&layout=item&view=item

Mittenzwey, K., Gitelson, A., Ullrich, S., & Kondratiev, K. (1992). Determination of chlorophylla of inland waters on the basis of spectral reectance. Limnology and Oceanography, 37,147149.

Morel, A., & Prieur, L. (1977). Analysis of variations in ocean color. Limnology andOceanography, 22, 709722.

Moses, W. J., Gitelson, A. A., Berdnikov, S., & Povazhnyy, V. (2009). Satellite estimationof chlorophyll a concentration using the red and NIR bands of MERISThe AzovSea case study. IEEE Geoscience and Remote Sensing Letters, 6(4), 845849, http://dx.doi.org/10.1109/LGRS.2009.2026657.

Moses, W. J., Gitelson, A. A., Perk, R. L., Gurlin, D., Rundquist, D. C., Leavitt, B. C., et al.(2011). Estimation of chlorophyllA concentration in turbid productive waters usingairborne hyperspectral data. Water Research, 46(4), 9931004, http://dx.doi.org/10.1016/j.watres.2011.11.068.

Mulla, D. J., & Sekely, A. (2009). Historical trends affecting accumulation of sedi-ment and phosphorus in Lake Pepin. Journal of Paleolimnology, 41(4), 589602,http://dx.doi.org/10.1007/s10933-008-9293-4.

Myre, E., & Shaw, R. (2006). The turbidity tube: Simple and accurate measurement ofturbidity in the eld. Houghton, MI: Michigan Technology University.

Olmanson, L. G., Bauer, M. E., & Brezonik, P. L. (2008). A 20-year Landsat water claritycensus ofMinnesota's 10,000 lakes.Remote Sensing of Environment,112(11), 40864097,http://dx.doi.org/10.1016/j.rse.2007.12.013.

Olmanson, L. G., Brezonik, P. L., & Bauer, M. E. (2011). Evaluation of medium to lowresolution satellite imagery for regional lake water quality assessments. WaterResources Research, 47, W09515, http://dx.doi.org/10.1029/2011WR011005.

Olmanson, L. G., Kloiber, S. M., Bauer, M. E., & Brezonik, P. L. (2001). Image processingprotocol for regional assessments of lake water quality. Public report series #14WaterResources Center and Remote Sensing Laboratory, University of Minnesota, St. Paul,MN, 55108, October 2001 (15 pp.).

Rabalais, N. N., Turner, R. E., & Scavia, D. (2002). Beyond science into policy: Gulf ofMexico hypoxia and the Mississippi River. BioScience, 52, 129142.

Rabalais, N. N., Turner, R. E., & Wiseman, W. J., Jr. (2002). Gulf of Mexico hypoxia, A.K.A. thedead zone. Annual Review of Ecology and Systematics, 33, 235263, http://dx.doi.org/10.1146/annurev.ecolsys.33.010802.150513.

Senay, G. B., Shaque, N. A., Autrey, B. C., Fulk, F., & Cormier, S. M. (2001). The selectionof narrow wavebands for optimizing water quality monitoring on the Great MiamiRiver, Ohio using hyperspectral remote sensor data. Journal of Spatial Hydrology, 1,122.

Shaque, N. A., Autrey, B. C., Fulk, F. A., & Flotemersch, J. E. (2003). Hyperspectralremote sensing of water quality parameters for large rivers in the Ohio RiverBasin. First Interagency Conference on Research in the Watersheds. Benson, Arizona,October 2730, 2003. Washington, DC: USDA Agric. Res. Serv.

Shah, C. A., Arora, M. K., Robila, S. A., & Varshney, P. K. (2002). ICA mixture model basedunsupervised classication of hyperspectral imagery. Proceedings of 31st Interna-

tional Workshop on Applied Imagery and Pattern Recognition. IEEE Computer Society,Washington DC, USA, 1618 October 2002. (pp. 2935) Los Alamitos, CA: IEEEComputer Society Press.

Shah, C. A., Arora,M. K., &Varshney, P. K. (2004). Unsupervised classication of hyperspectraldata: An ICA mixture model based approach. International Journal of Remote Sensing,25(2), 481487, http://dx.doi.org/10.1080/01431160310001618040.

Thorleifson, L. H. (1996). Review of Lake Agassiz history. In J. T. Teller, L. H. Thorleifson,G. Matile, & W. C. Brisbin (Eds.), Sedimentology, geomorphology and history ofthe central Lake Agassiz Basin Field trip guidebook B2; Geological Association ofCanada/Mineralogical Association of Canada Annual Meeting, Winnipeg, Manitoba,May 2729.

Turner, R. E., & Rabalais, N. N. (1994). Coastal eutrophication near the Mississippi Riverdelta. Nature, 368, 619621, http://dx.doi.org/10.1038/368619a0.

U.S. Environmental Protection Agency (1992). Methods for determination of chemicalsubstances in marine and estuarine environmental samples. EPA/600/R-92/121(95 pp.).

Vasilkov, A., & Kopelevich, O. (1982). Reasons for the appearance of the maximumnear 700 nm in the radiance spectrum emitted by the ocean layer. Oceanology, 22,697701.

Wessman, C. A., Aber, J. D., Petersen, D. L., & Melillo, J. M. (1988). Remote-sensing ofcanopy chemistry and nitrogen cycling in temperate forest ecosystems. Nature,335, 154156, http://dx.doi.org/10.1038/335154a0.

265L.G. Olmanson et al. / Remote Sensing of Environment 130 (2013) 254265

Airborne hyperspectral remote sensing to assess spatial distribution of water quality characteristics in large rivers: The ...1. Introduction2. Background information: major rivers in Twin Cities Metropolitan Area3. Methods3.1. Imagery3.2. Water quality measurements3.3. Image processing and classification

4. Results and discussion4.1. Spectral characteristics of river water4.2. Model development4.3. Water quality maps

5. ConclusionsAcknowledgmentsReferences