i ieee proof - cornell university · although radar images, particularly in c-band such as en-60...

IEEE

Proo

f

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 1

Suitability and Limitations of ENVISAT ASAR forMonitoring Small Reservoirs in a Semiarid Area

1

2

Jens R. Liebe, Nick van de Giesen, Marc S. Andreini, Tammo S. Steenhuis, and M. Todd Walter3

Abstract—In semiarid regions, thousands of small reservoirs4provide the rural population with water, but their storage volumes5and hydrological impact are largely unknown. This paper analyzes6the suitability of weather-independent radar satellite images for7monitoring small reservoir surfaces. The surface areas of three8reservoirs were extracted from 21 of 22 ENVISAT Advanced Syn-9thetic Aperture Radar scenes, acquired bimonthly from June 200510to August 2006. The reservoir surface areas were determined with11a quasi-manual classification approach, as stringent classification12rules often failed due to the spatial and temporal variability of the13backscatter from the water. The land–water contrast is critical for14the detection of water bodies. Additionally, wind has a significant15impact on the classification results and affects the water surface16and the backscattered radar signal (Bragg scattering) above a17wind speed threshold of 2.6 m · s−1. The analysis of 15 months18of wind speed data shows that, on 96% of the days, wind speeds19were below the Bragg scattering criterion at the time of night time20acquisitions, as opposed to 50% during the morning acquisition21time. Night time acquisitions are strongly advisable over day time22acquisitions due to lower wind interference. Over the year, radar23images are most affected by wind during the onset of the rainy sea-24son (May and June). We conclude that radar and optical systems25are complimentary. Radar is suitable during the rainy season but26is affected by wind and lack of vegetation context during the dry27season.28

Index Terms—Bragg scattering, radar, remote sensing,29reservoirs, resource management, water, water resources,30West Africa.31

Manuscript received March 4, 2008; revised May 19, 2008. This workwas supported by the Small Reservoirs Project, which is funded through theChallenge Program on Water and Food (CP46).AQ1

J. R. Liebe was with the Department of Biological and EnvironmentalEngineering, Cornell University, Ithaca, NY 14853-5701 USA. He is nowwith the Center for Development Research, University of Bonn, 53113 Bonn,AQ2Germany (e-mail: [email protected]).

N. van de Giesen is with the Water Resources Section, Faculty of CivilEngineering and Geosciences, Delft University of Technology, 2628 CN Delft,The Netherlands (e-mail: [email protected]).

M. S. Andreini is with the International Water Management Institute(IWMI), Accra, Ghana, and also with the United States Agency for Inter-AQ3national Development (USAID), Washington, DC 20523-1000 USA (e-mail:[email protected]).AQ4

T. S. Steenhuis is with the Department of Biological and Environmen-tal Engineering, Cornell University, Ithaca, NY 14853-5701 USA (e-mail:[email protected]).

M. T. Walter is with the Department of Biological and Environmental En-gineering, Cornell University, Ithaca, NY 14853-5701 USA, with the MontanaState University Northern, Havre, MT 59501 USA, and also with the Universityof Alaska Southeast, Juneau, AK 99801 USA (e-mail: [email protected]).AQ5

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TGRS.2008.2004805

I. INTRODUCTION 32

IN MANY semiarid regions of the developing world, access 33

to reliable water sources is the single most important factor 34

for the agricultural economy. Thousands of small reservoirs dot 35

the landscape, providing large volume water supply at the vil- 36

lage level, improving food security, and stimulating economic 37

development, particularly in rural areas. 38

Small reservoirs have been largely neglected in hydrological 39

and water resource research because of the combination of sev- 40

eral key characteristics: small size, existence in large numbers, 41

and widespread distribution. These characteristics constitute 42

their main advantages for the scattered rural population but 43

make their monitoring difficult. Adequate ground-based data on 44

small reservoir storage volumes are commonly not available, 45

and conducting ground-based surveys and measurements is 46

prohibitively expensive and time consuming on a regional scale. 47

To overcome the lack of baseline data, Liebe et al. [1] classified 48

the extent of small reservoir surface areas from Landsat ETM 49

imagery and determined regional small reservoir storage vol- 50

umes with a regional area–volume equation. Recently, further 51

studies on regional area–volume relations of small reservoirs 52

have been published, i.e., [2] and [3], indicating an interest in 53

information on small reservoir storage volumes. Such regional 54

storage volume estimates, however, depend on the ability to 55

extract reservoir surface areas from satellite images. Optical 56

satellite data yield good results in delineating small reservoir 57

surface areas under cloud-free conditions, but the often cloudy 58

conditions inhibit their use in an operational setting. 59

Although radar images, particularly in C-band such as EN- 60

VISAT’s Advanced Synthetic Aperture Radar (ASAR), have 61

become routinely available, the classification of distributed 62

inland water bodies has hardly been studied. Radar remote sens- 63

ing is capable of penetrating clouds and is seen as a promising 64

alternative to optical sensors. Successful application of radar 65

in determining small reservoir extents would not only facili- 66

tate transferring this methodology to other areas for regional 67

assessment of small reservoir storage but also allow regional 68

monitoring of storage volumes. 69

This paper analyzes the suitability and limitations of radar 70

remote sensing to determine small reservoir surface areas from 71

a sequence of 22 ENVISAT ASAR images acquired bimonthly 72

from June 2005 to August 2006. In contrast to the common 73

analysis of single images, or image pairs, this larger image 74

sequence ensures taking into account the seasonal variations, 75

i.e., the changing vegetation context, and the large variability 76

of backscatter from water surfaces, i.e., through wind-induced 77

roughness. The surface areas are extracted from the image 78

0196-2892/$25.00 © 2008 IEEE

IEEE

Proo

f

2 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

sequence for three reservoirs in the Upper East Region of Ghana79

and compared with in situ measurements based on bathymetric80

reservoir models and water level measurements.81

II. RADAR REMOTE SENSING OF OPEN WATER82

The detection of surface water on radar images is usually83

described as a simple task [4]. Smooth water surfaces act84

as specular reflectors and reflect most of the incoming radar85

signal away from the sensor. This is equivalent to very low86

radar backscatter signal returning to the sensor, which makes87

surface water bodies usually appear dark on radar images. This,88

however, is an oversimplification [5], as the surface roughness89

of water bodies is very variable, both spatially, within a water90

body, and temporally, leading to a wide range of backscatter. As91

will be shown here, this variability in backscatter can greatly92

affect the operational value of radar images for monitoring93

of small reservoirs. It is necessary to understand in some94

detail how contrast in backscatter between open water and95

surrounding land surface changes as a function of wind speed96

and direction, and vegetation density.97

Wind-induced regularly spaced waves and ripples can lead98

to Bragg scattering [6], which results in elevated backscatter99

signals from the water surface. While wave crests oriented100

orthogonally to the look direction can produce Bragg scattering,101

wave crests oriented in line with the look direction may have no102

significant effect on the radar backscatter. The threshold wind103

speed value causing Bragg scattering in C-band is estimated104

to be at ∼3.3 m · s−1 at 10 m above the surface [7]. This105

corresponds to a wind speed of 2.6 m · s−1 at 2 m, using106

Sutton’s [8] equation for wind speed profiles.107

Literature on open water delineation focuses on flood detec-108

tion and presents various methods. Henderson [5] presented a109

study on the extraction of lakes from X-band radar in different110

environments, using manual interpretation to allow the inclu-111

sion of context and other interpretation clues in the analysis.112

Barber et al. [9] and Brakenridge et al. [10] visually interpret113

flood extents for the 1993 Assiniboine River flood in Manitoba,114

Canada, and the 1993 Mississippi River flood, respectively.115

Henry et al. [11] use band thresholds to classify inundated areas116

of the 2002 Elbe river flood. Likewise, Brivio et al. [12] map117

the extent of the flooded areas of the 1994 flood in the Regione118

Piemonte, Italy, based on visual interpretation and band thresh-119

olds. van de Giesen [13] mapped flooding in a West African120

floodplain during the dry and wet seasons with L- and C-band121

SIR images, distinguishing between open water and water with122

reeds. Nico et al. [14] compare flood detection from amplitude123

change detection to coherence methods from multipass SAR124

data. Horrit et al. [15] use a statistical active contour model to125

delineate flood boundaries, and Heremans et al. [16] compare126

flood delineation results from and active contour model to that127

of an object-oriented classification technique. Context is an128

important factor for the delineation of water bodies. The degree129

of accuracy that small water bodies can be extracted from the130

radar images largely depends on the land–water contrast. For131

a distinct land–water contrast, a low and coherent backscatter132

from the water body is desirable that stands in distinct contrast133

to its surroundings, ideally producing higher signal returns.134

Due to the high dielectric constant of water, the penetration 135

depth of the radar signal into the water and, hence, volume 136

scattering and depolarization is low [4]. Reflections off of the 137

water surface are thus predominantly like-polarized. The return 138

from the water bodies in the HV band is therefore expected to 139

be low. Tall reeds growing on the sides of the reservoirs during 140

the rainy season can act as corner reflectors, which lead to 141

high backscatter signals in radar images due to double bounces 142

which can also partially depolarize the radar signal [4]. In 143

the radar image, this accentuates the land–water boundary and 144

facilitates its detection [5]. In the HH band, water bodies can 145

also be classified well when the water surface acts as a specular 146

reflector, i.e., ideally under calm conditions. The vast portion of 147

the radar burst is then scattered away from the sensor, leaving 148

the water body to appear dark in the image. Under windy 149

conditions, however, a rough water surface reflects more of the 150

incoming radar signal back to the sensor. These elevated re- 151

turns under windy conditions, particularly in the like-polarized 152

bands, are again due to the high dielectric constant of water. As 153

wind speeds are not always uniform over the entire water body, 154

elevated backscatter can occur in patches or affect larger parts 155

of the reservoir. Although elevated backscatter from the water 156

surface is detrimental to its classification in most cases, it can 157

also be seen as a signal typical for water bodies, which can be 158

helpful in classifying reservoirs. 159

Images acquired in dual-polarization mode can therefore 160

provide further clues for the land–water separation. In general, 161

like-polarized images have a better overall image contrast [4], 162

but VV is affected much more by Bragg scattering relative to 163

the HH and HV response [17]. 164

In this paper, the different subtleties of open water delin- 165

eation with ENVISAT ASAR will be explored, leading to a 166

comprehensive overview of the strengths and drawbacks. 167

III. STUDY REGION 168

The study is conducted in a 23-km2 watershed surrounding 169

the village of Tanga Natinga in the Upper East Region of 170

Ghana, West Africa (Fig. 1). Three small reservoirs, referred 171

to as Reservoirs 1, 2, and 3, supply the population of the 172

villages of Tanga, Weega, and Toende with water for irrigation 173

and gardening, livestock watering, household use, building, and 174

fishing [18], [19]. Maximum depths are 5.2 m for Reservoir 1, 175

4.7 m for Reservoir 2, and 4.3 m for Reservoir 3. Climat- 176

ically, the research area is located in the semiarid tropics 177

and is characterized by a monomodal rainy season from July 178

to September, with 986 mm of average annual rainfall and 179

2050 mm of average annual potential evaporation [20]. The area 180

lies in the northern Guinea savanna zone, and the vegetation 181

is characterized by open woodland, interspersed with annual 182

grasses [21]. Due to high population pressure, large areas 183

are under agricultural use. Between reservoirs and agricultural 184

land, there is usually a grass buffer of 10–30 m. The vegetation 185

dynamics in the vicinity of the reservoirs are largely driven by 186

the rainfall patterns. After the first rains, grasses grow around 187

the reservoirs. During the rainy season, the grasses can grow 188

up to 2 m tall, and extensive reeds are found in the tail parts of 189

the reservoirs. In the dry season, the grasses are often harvested 190

IEEE

Proo

f

LIEBE et al.: SUITABILITY AND LIMITATIONS OF ENVISAT ASAR FOR MONITORING SMALL RESERVOIRS 3

Fig. 1. Location of the three reservoirs, and weather stations within the Tanga watershed, in the Upper East Region of Ghana (demarked in NE corner of Ghanain the inset map). Weather stations’ recording wind speed and direction are located on Reservoir 3.

for roofing material, etc., burned, or they deteriorate, leaving191

behind bare dry soil with sparse knee-high grass tussocks as192

vegetation. The direct vicinity of the reservoirs is then free of193

vegetation, exposing the bare banks of the reservoirs.194

IV. DATA SETS AND METHODS195

A. Reservoir Bathymetry and Areas196

Bathymetric reservoir models and water level measurements197

serve as ground reference data for comparison with reser-198

voir surface areas determined with ENVISAT images. The199

bathymetric models were generated from GPS tagged water200

depth measurements and reservoir outlines as described by201

Liebe et al. [1]. Water level measurements are used together202

with the bathymetric models to determine the surface area and203

storage volumes of the reservoirs.204

Water levels were measured with pressure transducers at205

15-min intervals. These were used to determine ground ref-206

erence data of reservoir surface areas at the time of image207

acquisition, which are compared with the radar-based results.208

For Reservoirs 1 and 2, water level data are available from209

June 6, 2005 to February 21, 2006 and for Reservoir 3 from210

June 6, 2005 to August 3, 2006.211

B. Wind Speed and Wind Direction Data212

Wind-induced waves and ripples may influence the radar213

signal return from water surfaces. Besides wind speed, the214

wind direction is of importance, as it determines the crest215

orientation of the wind-induced waves. Assuming that wind216

produces waves with crests orthogonal to the wind direction,217

high wind speeds with wind directions orthogonally to the look218

direction may affect the backscatter from the water bodies less 219

than wind directions in line with the look direction. Wind speed 220

was measured on the center of Reservoir 3 and is available 221

from October 2005 onward. In addition, wind speed and wind 222

direction were measured at three locations on the shore of 223

Reservoir 3 (Fig. 1) at 2-min intervals, starting in August 2005. 224

C. ENVISAT Satellite Data 225

In this paper, 22 ENVISAT ASAR acquisitions are used with 226

a roughly bimonthly coverage from June 2005 to August 2006. 227

ENVISAT ASAR is a C-band radar. We used APG images, 228

with a nominal spatial resolution of 30 m and a pixel spacing 229

of 12.5 m. ENVISAT’s ASAR instrument can acquire images 230

in dual-polarization mode and produce HH- and VV-, HH- 231

and HV-, or VV- and VH-polarized image pairs [22]. In this 232

paper, we have chosen dual-polarized acquisitions with the 233

band combinations HH and HV. This combination has also 234

been found useful in flood delineation by Henry et al. [11]. 235

The scenes were acquired from different swaths (IS1–IS6) with 236

incidence angles ranging from an overall minimum of 12.57◦ 237

to a maximum of 43.77◦ throughout the scenes. The look 238

direction is 81.45◦ on ascending and 261.45◦ on descending 239

images. Images were acquired in both ascending (night time 240

acquisition) and descending (morning acquisition) nodes to 241

take into account the diurnal wind patterns, which determine 242

the occurrence of Bragg scattering. 243

D. Classification of Small Reservoir Surface Areas 244

Several approaches were tested to determine the surface areas 245

of the reservoirs, such as the active contour method or snake 246

IEEE

Proo

f


Fig. 2. (a) Example of a training area. (b) Resulting reservoir area aftergrowing by 3.1 standard deviations.

algorithm [23], band thresholds, and classification on HV–HH247

scatter plots. None perform well on images with less than248

optimal land–water contrast and/or water surfaces with patches249

of different backscatter intensities. Therefore, we used a quasi-250

manual method to determine reservoir sizes.251

In this quasi-manual method, a reservoir’s radar signal is252

classified by digitizing a training area within the reservoir and253

growing the region to neighboring pixels using a threshold254

specified by increments of standard deviations from the mean255

value of the radar signal within these digitized training areas.256

The training areas were defined by digitizing a polygon on the257

water surface not affected by wind-induced roughness [e.g.,258

Fig. 2(a)]. The standard deviations chosen for growing the259

regions from the digitized area vary for each reservoir and260

acquisition and were chosen to grow the region to the shoreline.261

Generally, the higher the number of standard deviations for262

growing the training areas, the better the contrast between land263

and water of the used band. The size of the training area can264

also have an influence on the standard deviations chosen for265

growing, depending on the variability of the backscatter values266

covered in the training area. Although there is no common rule,267

it may be noted that the standard deviations used for grow-268

ing the reservoir areas ranged from 1.0 to 7.8. The resulting269

reservoir areas often contain holes, but only areas defined by270

the boundaries of the reservoirs were considered. Fig. 2(b)271

shows the result for growing the previously defined training272

area by 3.1 standard deviations on the HV band acquired on273

September 2, 2005.274

Due to the variability of the backscatter from the water275

bodies, and the variable land–water contrast, the band or band276

combinations with the best visual land–water contrast were277

chosen for the classification. The reservoirs were classified on 278

either the HV or HH band alone, or an image generated by 279

multiplying the bands −1 × (HV × HH), prior to classification. 280

V. RESULTS 281

A. Reservoir Classification 282

While some of the images clearly showed the reservoirs 283

on both the HV and HH bands [e.g., Fig. 3(a) and (b)], their 284

visibility was often drastically better in one of the bands [e.g., 285

Fig. 3(e) and (f)]. Given the large variability in radar backscat- 286

ters from water surfaces, obtaining classification results from 287

a large number of image acquisitions requires a relaxed clas- 288

sification scheme like the quasi-manual approach used here. 289

The three following cases outline the importance of land–water 290

contrast, its variations, and the effect of incoherent backscatter 291

from a water body. 292

An example of excellent contrast between land and water for 293

both bands is the scene acquired on September 2, 2005. The 294

water bodies appear dark in the images, have sharp borders, and 295

can be classified from the low radar backscatter range of the HV 296

and HH bands [Fig. 3(a) and (b)]. At the time of image acqui- 297

sition, weather station 1 recorded no wind, whereas stations 2 298

and 3 recorded wind speeds of 1.3 and 3.0 m · s−1, respectively. 299

The wind direction produced wave crests expected to be in line 300

with the look direction [Fig. 3(a) and (c)], which is adverse 301

to Bragg scattering, and therefore, the water surface does not 302

produce elevated backscatter. In the scatter plot in Fig. 3(d), 303

“water pixels” aggregate in a cluster with low HV and HH 304

backscatter. Part of the reason why the reservoirs in this image 305

are so obvious is that, during this part of the year, the reservoirs 306

are filled to their full extent and are flanked by tall vegeta- 307

tion. To delineate the water bodies, the active contour method 308

[Fig. 4(a)], classification on the HH–HV scatter plot [Fig. 4(b)], 309

and growing of a training area [Fig. 4(c)] all worked similarly 310

well. Comparably excellent land–water contrast was only found 311

in the image acquired on October 7, 2005, and less distinct, but 312

still well on July 11, 2005, July 29, 2005, November 28, 2005, 313

July 11, 2006, August 3, 2006, and August 15, 2006. 314

An example of an image that only shows the reservoirs 315

well in one band was acquired on September 19, 2005. The 316

land–water contrast is distinct in the HV band [Fig. 3(e)], but 317

in the HH band [Fig. 3(f)], the major portions of the water 318

surface of Reservoirs 2 and 3 show elevated backscatter. While 319

the vegetation surrounding the reservoirs is still as tall as on 320

September 2, 2005, the reservoir outlines cannot be readily 321

determined from the HH band [Fig. 3(f) and (g)]. The elevated 322

backscatter is likely to be caused by wind-induced ripples. At 323

the time the image was acquired, wind speeds of 1.3 m · s−1 and 324

gusts up to 1.86 m · s−1 were measured at weather station 1. 325

The wind direction recorded at weather station 1 produces wave 326

crests expected to be roughly orthogonal to the look direction 327

[Fig. 3(e) and (g)], a favorable constellation for Bragg scatter- 328

ing. While the water surface in the HV band does not seem to be 329

affected, the surfaces of Reservoirs 2 and 3 [zoomed reservoir 330

in Fig. 3(g)] are made up of two distinct patches: a dark strip on 331

the windward side, where the wind may not have formed ripples 332

IEEE

Proo

f


Fig. 3. ENVISAT ASAR bands and scatter plots. Left [(a)–(d)]: Acquisition from September 2, 2005. The reservoirs are distinct in both (a) the HV band and(b) the HH band. The zoom on Reservoir 3 in (c) band HH shows good land water contrast. Wind speeds, wind directions (arrows), and expected crest orientationof wind-induced waves are depicted for (black) weather station 1, (blue) weather station 2, and (white) weather station 3. Wave crests are roughly in line withthe look direction and, thus, adverse for Bragg scattering. The (d) HV–HH scatterplot shows a water cluster in the lower backscatter ranges. The colors indicatethe frequency of backscatter ranges, with frequency increasing from blue to green, yellow, and red. Right [(e)–(h)]: Acquisition from September 19, 2005. Thereservoirs are distinct in (e) the HV band, but in (f) the HH band, reservoirs 2 and 3 are not distinctly visible. The zoom on Reservoir 3 in (g) band HH showsextensive areas of the water surface affected by Bragg scattering. Wave crests are at an angle to the look direction, which is more likely to produce Bragg scattering.The (h) HV–HH scatterplot shows a water cluster in the lower backscatter ranges and a second cluster with low backscatter in HV and elevated backscatter in HHdue to Bragg scattering.

IEEE

Proo

f


Fig. 4. Classification results on Reservoir 3. Left: September 2, 2005. (a) From active contour algorithm [Active contour parameters (Xu and Prince [23]):Elasticity (alpha) = 0.1, Rigidity (beta) = 0.25, Viscosity (gamma) = 1.0, External Force (kappa) = 1.25, Gradient Scale Factor = 1.75, Delta Min (dmin) =0.25, Delta Max (dmax) = 5.5, Noise Parameter (mu) = 0.1, GVF Iterations = 30, Contour Iterations = 120, Gaussian Sigma (sigma) = 1] on HH (11.3 ha).(b) From classification on the HH–HV scatter plot as shown in Fig. 4(left) (11.3 ha). (c) From quasi-manual approach on HH (11.9 ha). Right: September 19, 2005.(d) From active contour algorithm on HV (11.4 ha). (e) From classification on the HH–HV scatter plot as shown in Fig. 4(right) (3.7(red) + 5.1(green) = 8.8 ha,displayed on HH band). (f) From quasi-manual approach on HV (11.0 ha). AQ6

yet, and elevated backscatter from the major part of the surface333

area toward the leeward side. In the scatter plot in Fig. 3(h),334

there are two distinctive clusters corresponding to open water:335

one cluster with low backscatter in both bands (red outlines),336

similar to Fig. 3(d), and a cluster with low backscatter in the337

HV band, but elevated backscatter in band HH [green outlines338

in Fig. 3(g); scatter plot in Fig. 3(h)] due to Bragg scattering.339

The active contour method [Fig. 4(d)] and growing of a training340

area method [Fig. 4(f)] still work well on the HV band, while341

the reservoir area obtained through combining the two clusters342

that delineate in the HV–HH scatter plot [Fig. 3(h)] is too343

small [Fig. 4(e)]. Good land–water contrast in the HV band and344

patches of both low backscatter and elevated Bragg scattering345

in the HH band are present in the acquisitions from August 15,346

2005, September 19, 2005, and October 24, 2005. On the347

images acquired on January 27, 2006, February 21, 2006,348

March 3, 2006, and June 13, 2006, the land–water contrast349

was good in the HH band and poorer in the HV band.350

This emphasizes the need for a flexible methodology for351

categorizing the parts of the radar images that correspond to352

reservoirs, i.e., one rigid method will incorrectly categorize the353

reservoir pixels on many of the images.354

Fig. 5 shows an image acquisition of the same area from355

April 20, 2006, in which both bands fail to distinctly distinguish356

the water bodies from their surroundings. This image acquisi-357

tion falls into the dry season when the reservoir levels and sur-358

face areas have decreased significantly and the water bodies are359

surrounded by previously water-covered smooth banks, which 360

have a surface roughness similar to the water surface. The tall 361

vegetation, which surrounded the reservoirs in the rainy season, 362

is now essentially absent. Additionally, high wind speeds and 363

wind directions favorable for generating wave crests orthogo- 364

nally to the look direction (Fig. 5), and thus Bragg scattering, 365

are recorded. The loss of land–water contrast in the dryer and 366

less vegetated environment is reflected in the HV–HH scatter 367

plot with the signal from water bodies scattered throughout 368

(Fig. 5). Although we were unable to delineate the reservoirs in 369

the images from April 20, 2006, we were able to do so for all 370

the other days of radar image acquisition, even several images 371

with similar contrast issues, e.g., images from June 6, 2005, 372

June 24, 2005, November 11, 2005, December 16, 2005, 373

May 9, 2006, and June 29, 2006. 374

In analogy to van de Giesen’s [13] flood plain analysis, image 375

histogram characteristics were used for a qualitative image rat- 376

ing (Table I). The band histograms were calculated for the zoom 377

window on Reservoir 3, for the extent as shown in Fig. 3(c). 378

In cases of pronounced land–water contrast in a band, the his- 379

togram shows a double peak [Fig. 6(a)], where the peak in the 380

lower backscatter range is due to the water body, and the peak at 381

higher backscatter range is due to the vegetation. In bands with 382

low or no land–water contrast, the histogram produces only a 383

single peak [Fig. 6(b)], where the water and land pixels pro- AQ7384

duced backscatter at similar intensities. Acquisitions with dou- 385

ble peaks in both bands [i.e., Fig. 6(b)] were rated as “excellent” 386

IEEE

Proo

f


Fig. 5. (a)–(d) ENVISAT ASAR acquisition from April 20, 2006. In neither(a) the HV band nor (b) the HH band, the reservoirs are distinctly visible. Thezoom on Reservoir 3 in (c) band HH shows poor land water contrast. Windspeeds, wind directions (arrows; no record for weather station 2), and expectedcrest orientation of wind-induced waves are favorable for Bragg scattering. The(d) HV–HH scatterplot shows no distinct water cluster that is different from thebackscatter from the land.

(x), whereas those acquisitions with a double peak in one band 387

and a single peak in the other band were rated as “good” (g). 388

Acquisitions with single peaks in both band histograms [i.e., 389

Fig. 6(b)] were rated as “poor” (p) for land–water separation. 390

The quasi-manual classification method has been arrived at 391

through trial and error. The tested snake algorithm and band 392

threshold approach gave very poor results on images with 393

less than excellent land–water contrast, often not providing 394

any sensible delineation. For this reason, no quantitative 395

comparison of the different methods is provided. 396

B. Comparison of Radar- and Bathymetry-Based 397

Reservoir Sizes 398

Reservoir surface areas could be extracted from 21 of the 22 399

ENVISAT scenes used in this paper. These are compared with 400

surface areas determined from the reservoirs’ bathymetrical 401

models, and water levels at the time of image acquisition. The 402

overall performance of the reservoir size extraction compared 403

well to bathymetry-based reservoir sizes (r2 = 0.92, Fig. 7). 404

The classification performance, however, varied for the indi- 405

vidual reservoirs. The reservoir area classification of Reservoir 406

1 (r2 = 0.83) is mainly affected by the extensive reeds found 407

in the tail part of the reservoir. These are not identified as part 408

of the water body, whereas they are included in the bathymetric 409

model, causing a discrepancy. Reservoir 2 was the most difficult 410

to classify. Its southern shore is composed of a very smooth bare 411

sandy loam, which diminishes the land–water contrast. Nev- 412

ertheless, the highest coefficient of correlation was achieved 413

for Reservoir 2 (r2 = 0.95). Reservoir 3 was often affected by 414

wind, which leads to patches of elevated backscatter. 415

C. Wind Speeds and Scene Acquisition Times 416

The influence of wind speed at the time of image acquisition 417

on the land–water contrast is apparent from the average wind 418

speeds. Table I presents ENVISAT acquisition characteristics, 419

and averaged wind speed and direction records. The acquisi- 420

tions grouped in the aforementioned example of acquisitions 421

with excellent land–water contrast (“x,” Table I) are associated 422

with wind speeds of 0.6, 1.3, 1.4, 1.5, 1.6, and 3.9 m · s−1, 423

whereas images with good contrast but with some Bragg 424

scattering effects (“g,” Table I) show wind speeds of 0.2, 425

0.4, 0.7, 1.4, 1.6, 1.9, and 3.1 m · s−1. With two exceptions 426

(February 21, 2006 and July 11, 2006), these acquisitions are 427

associated with low wind speeds. The acquisitions listed in the 428

third example with poor contrast between land and water (“p,” 429

Table I) coincide with high wind speeds of 3.2, 3.2, 3.6, 3.7, 430

and 4.6 m · s−1. 431

Wind speeds were generally higher during the morning than 432

during the evening (Fig. 8). During the morning acquisition 433

time, on 214 out of 430 days, the 2-m wind speed exceeds 434

the 2.6-m · s−1 Bragg scattering threshold, i.e., on only 50% of 435

the days, wind speeds were below the threshold. For evening 436

overpasses, the Bragg scattering criterion is surpassed only 437

during 18 out of 430 days, i.e., on 97% of the days, wind speeds 438

were below the threshold. 439

IEEE

Proo

f


TABLE IENVISAT ACQUISITION CHARACTERISTICS, AVERAGE WIND DATA, AND BRAGG CRITERIA

VI. DISCUSSION440

The surface extent of small reservoirs could be extracted441

from 21 of the 22 ENVISAT ASAR scenes used in this paper,442

using the best band or band combination. The combination443

of both bands generally improved the classification result. For444

acquisition dates with image pairs consisting of an image with445

good and one with poor land–water contrast or Bragg scattering446

effects, the classification was performed on a single band.447

Based on ESA’s recommendation, image-mode VV-polarized448

images should be used to map open water [24]. Experience with449

a large number of ERS VV images shows that, for the purpose450

of monitoring small reservoirs, this is not the optimal choice.451

Instead, dual-polarized images are preferable for operational452

purposes. In general, HH images give the best results, but the453

HV images form a backup in case Bragg scattering occurs.AQ8 454

Stringent classification rules that allow simple and auto-455

mated surface area extraction only produce results on a small456

number of images with excellent land–water contrast and co-457

herent backscatter from within the water body. Due to the458

great variability and incoherence in the backscatter from water459

bodies, and in land–water contrast, stringent rules, however,460

fail quickly. With more flexible classification rules, such as the461

quasi-manual approach used here, good results were produced462

on all but one image. The fact that we have to manually463

identify the training area is comparable to purely visual tech- 464

niques commonly employed in radar image analysis [5], [9], 465

[10] and should not necessarily be considered a problem. Our 466

quasi-manual approach differs from a fully manual, or visual, 467

approach in that it is still based on backscatter statistics of 468

the training areas and relationships among neighboring pixels, 469

which allows us to set criteria that categorize “water pixels” 470

somewhat less subjectively, particularly when Bragg scattering 471

makes it difficult to visually identify all boundary pixels. 472

A comparison between reservoir sizes extracted from the 473

ENVISAT ASAR scenes and the bathymetry-based outlines 474

(Fig. 9) shows that the ENVISAT results (point markers) are 475

generally lower than the bathymetry-based surface areas (line 476

graphs). Fig. 9 also shows that, as the reservoirs attain their 477

maximum fill level, this difference increases and eventually 478

produces an almost constant offset during the period when the 479

reservoirs are full. As the reservoir levels fall, the difference 480

quickly diminishes. This increasing area differential with in- 481

creasing fill levels, and the eventual offset at the maximum 482

fill levels, is due to the development of wetland vegetation in 483

the inflow part of the small reservoirs. When the fill levels ap- 484

proach the maximum capacity, distinct land–water boundaries 485

diminish, and the open water body often gradually changes 486

into an extensive wetland at its inflow part. At these higher 487

IEEE

Proo

f


Fig. 6. (a) Double peak histogram indicating excellent land–water contrast.Images with distinct land–water contrast show a double peak in the imagehistogram, here in both the HV and HH bands. The extent of the analyzedwindow is shown in Fig. 3(c). (b) Single peak histogram indicating poor land–water contrast. Images with poor or no land–water contrast only show a singlepeak in the image histogram, here in both the HV and HH bands. The extent ofthe analyzed window is shown in Fig. 5(c).AQ9

Fig. 7. Comparison of reservoir sizes based on bathymetry and water levelmeasurements to reservoir sizes determined with ENVISAT. The overall re-lationship of radar-based reservoir sizes to the bathymetry-based referencehas an r2 of 0.92. At greater fill levels with surface areas over ∼10 ha, thediscrepancies become larger (see also Fig. 9).

fill levels, GPS-based outlines obtained by walking around the488

reservoirs included these wetlands because they constitute, in489

reality, open water. As the bathymetric models partially depend490

on these GPS outlines, the offset introduced at the highest fill491

levels is an artifact from the data acquisition for the bathy-492

metric models and, therefore, a direct result of the surveying.493

When the reservoir levels become lower again, this effect 494

quickly diminishes together with the presence of wetlands, 495

and the radar- and bathymetry-based area estimates converge. 496

By taking into account this overestimation of the bathymetric 497

data at full reservoir capacity as compared with ENVISAT 498

classified reservoir areas, the overall classification results are 499

acceptable. 500

The land–water separability is influenced by the vegetation 501

context, and the natural variability of the water surface, often as 502

a response to wind speed, which affects parts of or the entire 503

reservoir. To a large degree, the presence of tall vegetation 504

around the reservoirs drastically improves the delineation of 505

water bodies, as the low backscatter from the water surface 506

itself stands in distinct contrast to the high backscatter from 507

its edges, due to the high potential for double bounces off of the 508

water and the vertical vegetation. Such double bounces however 509

only occur during the rainy season and, shortly thereafter, when 510

the water levels are high and the vegetation is still lush. At 511

the same time, exact delineation of the reservoirs can become 512

difficult in the tail part at full capacity, when there is no 513

clear distinction between open water and wetland. During the 514

dry season, when the water levels have decreased, the water 515

bodies are mainly surrounded by the smoothly transgressing 516

basin sides, which are mostly free of vegetation. Under these 517

conditions, the land–water contrast is less distinct. 518

Wind was identified to affect the backscatter from water 519

bodies. Images with very poor land–water contrast, and the 520

acquisition from which the reservoirs could not be classified, 521

coincided with high wind speeds and wind directions which 522

are propitious for Bragg scattering, whereas most images with 523

excellent land–water contrast were acquired at low wind speeds 524

and/or wind directions which are unfavorable for Bragg scat- 525

tering. Choosing night time acquisitions clearly gives a higher 526

chance of acquisitions at lower wind speeds. As is typical for 527

most semiarid areas, the onset of the rainy season, here from 528

May to June, is a period with relatively high wind speeds. 529

During this period, the chance to acquire scenes with clearly 530

distinguishable small reservoirs diminishes. 531

A threshold value of 2.6 m · s−1 at 2 m has been put forward 532

as criterion for the occurrence of Bragg scatter. In the case 533

of small reservoirs, we do indeed see that when winds are 534

above this threshold, Bragg scatter occurs if the wave crests 535

are perpendicular to the look direction. It should be noted that 536

there were also occasions where minor wind gusts at low wind 537

speed (1.0 m · s−1) in the look direction caused Bragg scatter. 538

In such cases, the use of the dual-polarization mode ENVISAT 539

images is essential, because cross-polarized images are less 540

affected. 541

VII. CONCLUSION 542

The use of radar remote sensing as a tool for water resource 543

monitoring is promising due to its ability to penetrate clouds, 544

but the delineation of water bodies is difficult to automate. For 545

time series analysis, an automated extraction of water bodies 546

would be desirable but is not always possible due to the large 547

variation in the radar backscatter from the water surface and 548

to the changing ambient conditions. The quasi-manual method AQ10549

IEEE

Proo

f


Fig. 8. Measured wind speeds over the course of the year for morning and evening overpasses. (Dark blue) Night time acquisitions yield a much higher chanceto obtain images where the water bodies are not affected by wind-induced waves. Over a period of 15 months, wind speeds recorded during the (dark blue) nighttime acquisitions were below the Bragg criterion on 96% of the days, whereas during the (magenta) morning acquisitions, wind speeds were below the Braggcriterion on only 50% of the observed days. The ten-day averages (morning acquisition on red; night acquisition in light blue) show that wind speeds during themorning acquisitions are generally higher than those recorded at night and also show a seasonal cycle. During the dryer months from December to July, windspeeds are particularly high during the morning acquisitions.

Fig. 9. Comparison of (markers) reservoir sizes from ENVISAT classification with (lines) reservoir sizes from bathymetrical models. Note the almost constantoffset between bathymetry- and satellite-based surface areas at the maximum fill levels, which is due to the inclusion of wetlands in the bathymetric model. Thesedifferences diminish with decreasing reservoir size, when the wetlands fall dry.

presented provides a good combination of computer objectivity550

and human classification skills.551

The analysis of the radar image time series indicates that552

the land–water contrast, which is of greatest importance for the553

detection of water bodies, varies significantly with the seasons.554

Radar images acquired during the rainy season showed the best555

land–water contrast and were most easily classified. In the dry556

season, with smaller water bodies and lack of surrounding veg-557

etation, their classification was more difficult as the land–water558

contrast diminishes. Toward, and throughout the dry season,559

the water detection on radar imagery is most difficult, as the560

land–water contrast suffers from the lack of vegetation context.AQ11 561

Under such conditions, optical systems yield good results [1],562

even on small water bodies, and independent of the surrounding563

vegetation and wind conditions, as long as cloud-free images564

can be obtained. This leads to the important conclusion that565

optical- and radar-based methods can be seen as seasonally 566

complementary for surface water detection, particularly in 567

semiarid areas. 568

The backscatter signal of water bodies is significantly in- 569

fluenced by wind-induced waves and wave crest orientation. 570

The analysis shows that Bragg scattering effects emerge at 571

much lower wind speeds than ESA’s wind speed threshold for 572

Bragg scattering, which translates to wind speeds of 2.6 m · s−1 573

(at 2 m height); however, the heavily affected acquisitions 574

with poor land water contrast are all associated with wind 575

speeds well above this threshold. The analysis of wind speed 576

prevalence at the time of the morning and night time image 577

acquisitions shows a distinct difference. Night time acquisi- 578

tions are much less likely affected by wind than the morning 579

acquisitions. For the delineation of water bodies, selecting 580

night time acquisitions yields a significantly higher chance of 581

IEEE

Proo

f


obtaining radar images at wind speed conditions below the582

Bragg criterion. Although Bragg scattering was also observed583

below the 2.6-m · s−1 threshold, Bragg producing gusts are less584

likely during the night time acquisitions, when the atmosphere585

commonly has stabilized.586

Overall, this paper shows that regional to basin scale in-587

ventories of small inland water bodies are readily possible588

with ENVISAT ASAR images. In combination with regional589

area–volume equations (i.e., [1]–[3]), basin-wide small reser-590

voir storage volumes can be estimated, and the impact of591

further development can be assessed and monitored. With the592

ever improving digital elevation models, it is foreseeable that593

area–volume equations can be determined adequately from594

these data, which will allow for regional and basin-wide small595

reservoir storage volume estimates at any given location. Fur-596

ther research could clarify whether ALOS PALSAR’s L-band597

data yield better land–water separation under the various wind598

conditions and seasonal changes in vegetation context, which599

would allow extracting small reservoirs and other inland water600

bodies at a higher degree of automation.601

ACKNOWLEDGMENT602

The authors would like to thank for the cooperation with603

the GLOWA Volta Project, the Center for Development Re-604

search, University of Bonn, Bonn, and the International Water605

Management Institute, Ghana, for supporting the field work.606

The authors would also like to thank Cornell University for607

funding in the form of an assistantship. The ENVISAT im-608

ages used in this paper were obtained through ESA’s TIGER609

Project 2871.610

REFERENCES611

[1] J. Liebe, N. van de Giesen, and M. Andreini, “Estimation of small reser-612voir storage capacities in a semi-arid environment—A case study in the613Upper East Region of Ghana,” Phys. Chem. Earth, vol. 30, no. 6/7,614pp. 448–454, 2005.615

[2] P. Cecchi, “L’eau en partage: les petits barrages de la Côte d’Ivoire,” in616Collection Latitudes 23 Paris. Paris, France: IRD Editions, 2007.617

[3] T. Sawunyama, A. Senzanje, and A. Mhizha, “Estimation of small618reservoir storage capacities in Limpopo River Basin using geographi-619cal information systems (GIS) and remotely sensed surface areas: Case620of Mzingwane catchment,” Phys. Chem. Earth, vol. 31, no. 15/16,621pp. 935–943, 2006.622

[4] F. M. Henderson and A. J. Lewis, Principles and Applications of Imaging623Radar, 3rd ed, vol. 2. New York: Wiley, 1998.624

[5] F. M. Henderson, “Environmental factors and the detection of open sur-625face water areas with X-band radar imagery,” Int. J. Remote Sens., vol. 16,626no. 13, pp. 2423–2437, Sep. 1995.627

[6] G. R. Valenzuela, “Theories for the interaction of electromagnetic and628oceanic waves: A review,” Boundary Layer Meteorol., vol. 13, no. 1–4,629pp. 61–85, Jan. 1978.630

[7] ESA, Bragg Scattering, 2005. [Online]. Available: http://earth.esa.631int/applications/data_util/SARDOCS/spaceborne/Radar_Courses/Radar_632Course_II/Bragg_scattering.htm.633

[8] O. G. Sutton, “Wind structure and evaporation in a turbulent atmosphere,”634Proc. R. Soc. Lond. A, Contain. Pap. Math. Phys. Character, vol. 146,635no. 858, pp. 701–722, Oct. 1934.636

[9] D. G. Barber, K. P. Hochheim, R. Dixon, D. R. Mosscrop, and637M. J. McMullan, “The role of earth observation technology in flood638mapping: A Manitoba Case study,” Can. J. Remote Sens., vol. 22, no. 1,639pp. 137–143, 1996.640

[10] G. B. Brakenridge, J. C. Knox, E. D. Paylor, II, and F. Magiligan, “Radar641remote sensing aids study of the great flood of 1993,” EOS, vol. 75, no. 45,642pp. 521–530, 1994.643

[11] J. B. Henry, P. Chastanet, K. Fellah, and Y. L. Desnos, “Envisat multi- 644polarized ASAR data for flood mapping,” Int. J. Remote Sens., vol. 27, 645no. 10, pp. 1921–1929, May 2006. 646

[12] P. A. Brivio, R. Colombo, M. Maggi, and R. Tomasoni, “Integration of 647remote sensing data and GIS for accurate mapping of flooded areas,” Int. 648J. Remote Sens., vol. 23, no. 3, pp. 429–441, Feb. 2002. 649

[13] N. van de Giesen, “Characterization of West African shallow flood plains 650with L- and C-band radar,” in Remote Sensing and Hydrology 2000, 651vol. 267, M. Owe and K. Brubaker, Eds. Paris, France: IAHS Publica- 652tion, 2001, pp. 365–367. 653

[14] G. Nico, M. Pappalepore, G. Pasquariello, A. Refice, and 654S. Samarelli, “Comparison of SAR amplitude vs. coherence flood 655detection methods—A GIS application,” Int. J. Remote Sens., vol. 21, 656no. 8, pp. 1619–1631, May 2000. 657

[15] M. S. Horritt, D. C. Mason, and A. J. Luckman, “Flood boundary de- 658lineation from synthetic aperture radar imagery using a statistical active 659contour model,” Int. J. Remote Sens., vol. 22, no. 13, pp. 2489–2507, 660Sep. 2001. 661

[16] R. Heremans, A. Willekens, D. Borghys, B. Verbeeck, J. Valckenborgh, 662M. Acheroy, and C. Perneel, “Automatic detection of flooded areas on 663ENVISAT/ASAR images using an object-oriented classification tech- 664nique and an active contour algorithm,” in Proc. 1st Int. Conf. Recent 665Advances Space Technol., Istanbul, Turkey, 2003, pp. 311–316. 666

[17] J. J. Van der Sanden and S. J. Thomas, Application Potential of 667RADARSAT-2, 2004. Supplement One. [Online]. Available: http://www. 668radarsat2.info/sartrek/2005/jun/RSAT2_APPS_2004_PDF_Final.pdf 669

[18] J. W. Faulkner, T. Steenhuis, N. V. De Giesen, M. Andreini, and 670J. R. Liebe, “Water use and productivity of two small reservoir irrigation 671schemes in Ghana’s upper east region,” Irrigation Drainage, vol. 57, 672no. 2, pp. 151–163, Apr. 2008. 673

[19] J. Liebe, M. Andreini, N. van de Giesen, and T. Steenhuis, “The small 674reservoirs project: Research to improve water availability and economic 675development in rural semi-arid areas,” in The Hydropolitics of Africa: 676A Contemporary Challenge, M. Kittisou, M. Ndulo, M. Nagel, and 677M. Grieco, Eds. Newcastle, Australia: Cambridge Scholars Publishing, 6782007, pp. 325–332. 679

[20] G. Kranjac-Berisajevic’, T. B. Bayorbor, A. S. Abdulai, and F. Obeng, 680“Rethinking natural resource degradation in semi-arid sub-Saharan 681Africa,” in The Case of Semi-Arid Ghana. Tamale, Ghana: Faculty Agri- 682culture, Univ. Develop. Stud., 1998. Revised January 1999. 683

[21] P. N. Windmeijer and W. Andriesse, Inland Valleys in West Africa: 684An Agro-Ecological Characterization of Rice-Growing Environments, 685vol. 52. Wageningen, The Netherlands: Int. Inst. Land Reclamation Im- 686provement, ILRI Publication, 1993. 687

[22] B. Gardini, G. Graf, and G. Ratier, “The instruments on ENVISAT,” Acta 688Astronaut., vol. 37, pp. 301–311, Oct. 1995. 689

[23] C. Y. Xu and J. L. Prince, “Snakes, shapes, and gradient vector flow,” 690IEEE Trans. Image Process., vol. 7, no. 3, pp. 359–369, Mar. 1998. 691

[24] “ASAR modes for remote sensing science,” ASAR Product Handbook, 692ESA, Paris, France, Feb. 2007. Issue 2.2. [Online]. Available: http:// 693envisat.esa.int/handbooks/asar/CNTR1-1-6.htm 694

Jens R. Liebe received the Diploma M.Sc. degree 695in geography from the University of Bonn, Bonn, 696Germany, in 2002. 697

In 2004, he was with the Department of Biolog- 698ical and Environmental Engineering, Cornell Uni- 699versity, Ithaca, NY, where he conducted his gradu- 700ate research within the Small Reservoirs Project in 701Northern Ghana. His work focuses on the hydrology 702of semiarid areas, small reservoirs, and the use of 703remote sensing in data scarce areas. Since 2008, he 704has been with the Center for Development Research, 705

University of Bonn, where he was appointed as the Scientific Coordinator of 706the Global Change in the Hydrological Cycle Volta Project. 707

Mr. Liebe is a member of the American Geophysical Union, the Euro- AQ12708pean Geosciences Union, and the International Association of Hydrological AQ13709Sciences. He was the recipient of the Hans-Hartwig Ruthenberg Award for AQ14710his thesis. 711

IEEE

Proo

f


Nick van de Giesen received the Kandidaats B.S.712degree and the M.Sc. degree in land and water man-713agement from Wageningen Agricultural University,714Wageningen, The Netherlands, in 1984 and 1987,715respectively, and the Ph.D. degree in agricultural716and biological engineering from Cornell University,717Ithaca, NY, in 1994.718

After a postdoctoral position with the West719Africa Rice Development Association, Bouaké, Côte720d’Ivoire, he was a Senior Researcher for six years721with the Center for Development Research (ZEF),722

University of Bonn, Bonn, Germany, where he was the Scientific Coordinator723of the Global Change in the Hydrological Cycle Volta Project. Since 2004, he724has been with the Water Resources Section, Faculty of Civil Engineering and725Geosciences, Delft University of Technology, Delft, The Netherlands, where he726is currently the “van Kuffeler” Chair of Water Resources Management. He is727the Secretary of the International Commission on Water Resources Systems of728the International Association of Hydrological Sciences, the Chairman of the729Committee on Water Policy and Management of the European Geosciences730Union, and the Chairman of the Netherlands Commission on Irrigation and731Drainage.732

Prof. van de Giesen is a senior fellow of ZEF, University of Bonn.733

Marc S. Andreini is a Civil Engineer and isAQ15 734currently a Senior Researcher with the Interna-735tional Water Management Institute (IWMI), Accra,736Ghana, where he has been the Ghana Coordina-737tor of the GLOWA Volta Project. He is also cur-738rently an Advisor with the United States Agency739for International Development, Washington, DC,740and the Leader of the Small Reservoirs Project741(www.smallreservoirs.org). He has worked in Cali-742fornia and several African countries. He has worked743as a Civil Engineer on a variety of construction and744

water supply projects. He built village water supply systems in Morocco, was a745Physical Planner for the UNHCR in Tanzania, did research on shallow ground-746water irrigation in Zimbabwe, and was a member of the project coordinating747unit supervising the construction of Botswana’s North South Carrier.748

Tammo S. Steenhuis received the Kandidaats (B.S.) 749degree and the Ingenieur (M.S.) degree in land and 750water management of the tropics from Wageningen 751Agricultural University, Wageningen, The Nether- 752lands, and the M.S. and Ph.D. degrees from the 753Agricultural Engineering Department, University of 754Wisconsin, Madison. 755

Since 1978, he has been a Faculty Member with 756the Department of Biological and Environmental En- 757gineering, Cornell University, Ithaca, NY, where he 758is an International Professor of water management. 759

He is currently directing the Cornell Masters program in Watershed Manage- 760ment and Hydrology at the University of Bahir Dar, Bahir Dar, Ethiopia. His 761research encompasses the movement of colloids at the pore- to large-scale 762studies of the distributed runoff response of watersheds such as in New York 763City source watersheds in the Catskill Mountains and the Abay Blue Nile in 764Ethiopia. 765

Dr. Steenhuis is a fellow of the American Geophysical Union. He was the 766recipient of the Darcy Metal of the European Geosciences Union. 767

M. Todd Walter received the B.S. degree in bio- 768logical and environmental engineering and the M.S. 769degree in civil and environmental engineering from 770Cornell University, Ithaca, NY, in 1990 and 1991, 771respectively, and the Ph.D. degree in engineering 772science from Washington State University, Pullman, 773in 1995. AQ16774

He is an Assistant Professor with the Depart- 775ment of Biological and Environmental Engineering, 776Cornell University. His specialization is hydrology 777with particular emphases on transport processes and 778

watershed modeling. Two of his current focus areas are interactions between 779hydrology and biogeochemistry and environmental applications of nanotech- 780nology. The objectives of much of his work is to improve our understanding of 781how chemicals, microorganisms, and sediment move through the landscapes in 782order to develop better strategies for protecting water quality. He has also been 783an Assistant Professor with the Montana State University Northern, Havre, and AQ17784the University of Alaska Southeast, Juneau. 785

Dr. Walter is an active member of the American Society of Agricultural Engi- AQ18786neers, the American Society of Civil Engineers, and the American Geophysical AQ19787Union. 788

IEEE

Proo

f

AUTHOR QUERIES

AUTHOR PLEASE ANSWER ALL QUERIES

AQ1 = Please validate the postal codes provided for the current affiliations of the authors.AQ2 = For consistency purposes, “Bonn University” was changed to “University of Bonn.” Please check if

appropriate.AQ3 = Please validate the address provided for IWMI. In addition, please provide postal code for IWMI.AQ4 = The acronym “USAID” was defined as “United States Agency for International Development.” Please

check if appropriate.AQ5 = “Montana State University Northern” and “University of Alaska Southeast” were captured as another

current affiliations of author M. T. Walter based on the curriculum vitae. Please check if appropriate.AQ6 = “the histogram” was inserted in this sentence for clarity purposes. Please check if appropriate.AQ7 = The asterisk symbol was deleted and, thus, replaced with its corresponding footnote/legend data.

Please check if appropriate.AQ8 = The sentence was rephrased or reworded. Please check if appropriate.AQ9 = Please indicate the figure subpanels [(a.) and (b.)] in the artwork of Figure 6.AQ10 = The sentence was rephrased or reworded. Please check if appropriate.AQ11 = The sentence was rephrased or reworded. Please check if appropriate.AQ12 = The acronym “AGU” was defined as “American Geophysical Union.” Please check if appropriate.AQ13 = The acronym “EGU” was defined as “European Geosciences Union.” Please check if appropriate.AQ14 = The acronym “IAHS” was defined as “International Association of Hydrological Sciences.” Please

check if appropriate.AQ15 = Please provide educational background for author M. S. Andreini.AQ16 = The sentence was rephrased or reworded to conform with the IEEE standards. Please check.AQ17 = “University of Montana Northern” was changed to “Montana State University Northern, Havre,

MT.” Please check if appropriate.AQ18 = The acronym “ASAE” was defined as “American Society of Agricultural Engineers.” Please check

if appropriate.AQ19 = The acronym “ASCE” was defined as “American Society of Civil Engineers.” Please check if

appropriate.

ATTN: If you are paying to have all or some of your figures appear in color in the print issue, it is veryimportant that you fill out and submit a copy of the IEEE Page Charge & Reprint Form along with yourproof corrections. This form is available from the same URL where these page proofs were downloadedfrom. Thank you.

END OF ALL QUERIES

IEEE

Proo

f

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 1

Suitability and Limitations of ENVISAT ASAR forMonitoring Small Reservoirs in a Semiarid Area

1

2

Jens R. Liebe, Nick van de Giesen, Marc S. Andreini, Tammo S. Steenhuis, and M. Todd Walter3

Abstract—In semiarid regions, thousands of small reservoirs4provide the rural population with water, but their storage volumes5and hydrological impact are largely unknown. This paper analyzes6the suitability of weather-independent radar satellite images for7monitoring small reservoir surfaces. The surface areas of three8reservoirs were extracted from 21 of 22 ENVISAT Advanced Syn-9thetic Aperture Radar scenes, acquired bimonthly from June 200510to August 2006. The reservoir surface areas were determined with11a quasi-manual classification approach, as stringent classification12rules often failed due to the spatial and temporal variability of the13backscatter from the water. The land–water contrast is critical for14the detection of water bodies. Additionally, wind has a significant15impact on the classification results and affects the water surface16and the backscattered radar signal (Bragg scattering) above a17wind speed threshold of 2.6 m · s−1. The analysis of 15 months18of wind speed data shows that, on 96% of the days, wind speeds19were below the Bragg scattering criterion at the time of night time20acquisitions, as opposed to 50% during the morning acquisition21time. Night time acquisitions are strongly advisable over day time22acquisitions due to lower wind interference. Over the year, radar23images are most affected by wind during the onset of the rainy sea-24son (May and June). We conclude that radar and optical systems25are complimentary. Radar is suitable during the rainy season but26is affected by wind and lack of vegetation context during the dry27season.28

Index Terms—Bragg scattering, radar, remote sensing,29reservoirs, resource management, water, water resources,30West Africa.31

Manuscript received March 4, 2008; revised May 19, 2008. This workwas supported by the Small Reservoirs Project, which is funded through theChallenge Program on Water and Food (CP46).AQ1

J. R. Liebe was with the Department of Biological and EnvironmentalEngineering, Cornell University, Ithaca, NY 14853-5701 USA. He is nowwith the Center for Development Research, University of Bonn, 53113 Bonn,AQ2Germany (e-mail: [email protected]).

N. van de Giesen is with the Water Resources Section, Faculty of CivilEngineering and Geosciences, Delft University of Technology, 2628 CN Delft,The Netherlands (e-mail: [email protected]).

M. S. Andreini is with the International Water Management Institute(IWMI), Accra, Ghana, and also with the United States Agency for Inter-AQ3national Development (USAID), Washington, DC 20523-1000 USA (e-mail:[email protected]).AQ4

T. S. Steenhuis is with the Department of Biological and Environmen-tal Engineering, Cornell University, Ithaca, NY 14853-5701 USA (e-mail:[email protected]).

M. T. Walter is with the Department of Biological and Environmental En-gineering, Cornell University, Ithaca, NY 14853-5701 USA, with the MontanaState University Northern, Havre, MT 59501 USA, and also with the Universityof Alaska Southeast, Juneau, AK 99801 USA (e-mail: [email protected]).AQ5

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TGRS.2008.2004805

I. INTRODUCTION 32

IN MANY semiarid regions of the developing world, access 33

to reliable water sources is the single most important factor 34

for the agricultural economy. Thousands of small reservoirs dot 35

the landscape, providing large volume water supply at the vil- 36

lage level, improving food security, and stimulating economic 37

development, particularly in rural areas. 38

Small reservoirs have been largely neglected in hydrological 39

and water resource research because of the combination of sev- 40

eral key characteristics: small size, existence in large numbers, 41

and widespread distribution. These characteristics constitute 42

their main advantages for the scattered rural population but 43

make their monitoring difficult. Adequate ground-based data on 44

small reservoir storage volumes are commonly not available, 45

and conducting ground-based surveys and measurements is 46

prohibitively expensive and time consuming on a regional scale. 47

To overcome the lack of baseline data, Liebe et al. [1] classified 48

the extent of small reservoir surface areas from Landsat ETM 49

imagery and determined regional small reservoir storage vol- 50

umes with a regional area–volume equation. Recently, further 51

studies on regional area–volume relations of small reservoirs 52

have been published, i.e., [2] and [3], indicating an interest in 53

information on small reservoir storage volumes. Such regional 54

storage volume estimates, however, depend on the ability to 55

extract reservoir surface areas from satellite images. Optical 56

satellite data yield good results in delineating small reservoir 57

surface areas under cloud-free conditions, but the often cloudy 58

conditions inhibit their use in an operational setting. 59

Although radar images, particularly in C-band such as EN- 60

VISAT’s Advanced Synthetic Aperture Radar (ASAR), have 61

become routinely available, the classification of distributed 62

inland water bodies has hardly been studied. Radar remote sens- 63

ing is capable of penetrating clouds and is seen as a promising 64

alternative to optical sensors. Successful application of radar 65

in determining small reservoir extents would not only facili- 66

tate transferring this methodology to other areas for regional 67

assessment of small reservoir storage but also allow regional 68

monitoring of storage volumes. 69

This paper analyzes the suitability and limitations of radar 70

remote sensing to determine small reservoir surface areas from 71

a sequence of 22 ENVISAT ASAR images acquired bimonthly 72

from June 2005 to August 2006. In contrast to the common 73

analysis of single images, or image pairs, this larger image 74

sequence ensures taking into account the seasonal variations, 75

i.e., the changing vegetation context, and the large variability 76

of backscatter from water surfaces, i.e., through wind-induced 77

roughness. The surface areas are extracted from the image 78

0196-2892/$25.00 © 2008 IEEE

IEEE

Proo

f


sequence for three reservoirs in the Upper East Region of Ghana79

and compared with in situ measurements based on bathymetric80

reservoir models and water level measurements.81

II. RADAR REMOTE SENSING OF OPEN WATER82

The detection of surface water on radar images is usually83

described as a simple task [4]. Smooth water surfaces act84

as specular reflectors and reflect most of the incoming radar85

signal away from the sensor. This is equivalent to very low86

radar backscatter signal returning to the sensor, which makes87

surface water bodies usually appear dark on radar images. This,88

however, is an oversimplification [5], as the surface roughness89

of water bodies is very variable, both spatially, within a water90

body, and temporally, leading to a wide range of backscatter. As91

will be shown here, this variability in backscatter can greatly92

affect the operational value of radar images for monitoring93

of small reservoirs. It is necessary to understand in some94

detail how contrast in backscatter between open water and95

surrounding land surface changes as a function of wind speed96

and direction, and vegetation density.97

Wind-induced regularly spaced waves and ripples can lead98

to Bragg scattering [6], which results in elevated backscatter99

signals from the water surface. While wave crests oriented100

orthogonally to the look direction can produce Bragg scattering,101

wave crests oriented in line with the look direction may have no102

significant effect on the radar backscatter. The threshold wind103

speed value causing Bragg scattering in C-band is estimated104

to be at ∼3.3 m · s−1 at 10 m above the surface [7]. This105

corresponds to a wind speed of 2.6 m · s−1 at 2 m, using106

Sutton’s [8] equation for wind speed profiles.107

Literature on open water delineation focuses on flood detec-108

tion and presents various methods. Henderson [5] presented a109

study on the extraction of lakes from X-band radar in different110

environments, using manual interpretation to allow the inclu-111

sion of context and other interpretation clues in the analysis.112

Barber et al. [9] and Brakenridge et al. [10] visually interpret113

flood extents for the 1993 Assiniboine River flood in Manitoba,114

Canada, and the 1993 Mississippi River flood, respectively.115

Henry et al. [11] use band thresholds to classify inundated areas116

of the 2002 Elbe river flood. Likewise, Brivio et al. [12] map117

the extent of the flooded areas of the 1994 flood in the Regione118

Piemonte, Italy, based on visual interpretation and band thresh-119

olds. van de Giesen [13] mapped flooding in a West African120

floodplain during the dry and wet seasons with L- and C-band121

SIR images, distinguishing between open water and water with122

reeds. Nico et al. [14] compare flood detection from amplitude123

change detection to coherence methods from multipass SAR124

data. Horrit et al. [15] use a statistical active contour model to125

delineate flood boundaries, and Heremans et al. [16] compare126

flood delineation results from and active contour model to that127

of an object-oriented classification technique. Context is an128

important factor for the delineation of water bodies. The degree129

of accuracy that small water bodies can be extracted from the130

radar images largely depends on the land–water contrast. For131

a distinct land–water contrast, a low and coherent backscatter132

from the water body is desirable that stands in distinct contrast133

to its surroundings, ideally producing higher signal returns.134

Due to the high dielectric constant of water, the penetration 135

depth of the radar signal into the water and, hence, volume 136

scattering and depolarization is low [4]. Reflections off of the 137

water surface are thus predominantly like-polarized. The return 138

from the water bodies in the HV band is therefore expected to 139

be low. Tall reeds growing on the sides of the reservoirs during 140

the rainy season can act as corner reflectors, which lead to 141

high backscatter signals in radar images due to double bounces 142

which can also partially depolarize the radar signal [4]. In 143

the radar image, this accentuates the land–water boundary and 144

facilitates its detection [5]. In the HH band, water bodies can 145

also be classified well when the water surface acts as a specular 146

reflector, i.e., ideally under calm conditions. The vast portion of 147

the radar burst is then scattered away from the sensor, leaving 148

the water body to appear dark in the image. Under windy 149

conditions, however, a rough water surface reflects more of the 150

incoming radar signal back to the sensor. These elevated re- 151

turns under windy conditions, particularly in the like-polarized 152

bands, are again due to the high dielectric constant of water. As 153

wind speeds are not always uniform over the entire water body, 154

elevated backscatter can occur in patches or affect larger parts 155

of the reservoir. Although elevated backscatter from the water 156

surface is detrimental to its classification in most cases, it can 157

also be seen as a signal typical for water bodies, which can be 158

helpful in classifying reservoirs. 159

Images acquired in dual-polarization mode can therefore 160

provide further clues for the land–water separation. In general, 161

like-polarized images have a better overall image contrast [4], 162

but VV is affected much more by Bragg scattering relative to 163

the HH and HV response [17]. 164

In this paper, the different subtleties of open water delin- 165

eation with ENVISAT ASAR will be explored, leading to a 166

comprehensive overview of the strengths and drawbacks. 167

III. STUDY REGION 168

The study is conducted in a 23-km2 watershed surrounding 169

the village of Tanga Natinga in the Upper East Region of 170

Ghana, West Africa (Fig. 1). Three small reservoirs, referred 171

to as Reservoirs 1, 2, and 3, supply the population of the 172

villages of Tanga, Weega, and Toende with water for irrigation 173

and gardening, livestock watering, household use, building, and 174

fishing [18], [19]. Maximum depths are 5.2 m for Reservoir 1, 175

4.7 m for Reservoir 2, and 4.3 m for Reservoir 3. Climat- 176

ically, the research area is located in the semiarid tropics 177

and is characterized by a monomodal rainy season from July 178

to September, with 986 mm of average annual rainfall and 179

2050 mm of average annual potential evaporation [20]. The area 180

lies in the northern Guinea savanna zone, and the vegetation 181

is characterized by open woodland, interspersed with annual 182

grasses [21]. Due to high population pressure, large areas 183

are under agricultural use. Between reservoirs and agricultural 184

land, there is usually a grass buffer of 10–30 m. The vegetation 185

dynamics in the vicinity of the reservoirs are largely driven by 186

the rainfall patterns. After the first rains, grasses grow around 187

the reservoirs. During the rainy season, the grasses can grow 188

up to 2 m tall, and extensive reeds are found in the tail parts of 189

the reservoirs. In the dry season, the grasses are often harvested 190

IEEE

Proo

f


Fig. 1. Location of the three reservoirs, and weather stations within the Tanga watershed, in the Upper East Region of Ghana (demarked in NE corner of Ghanain the inset map). Weather stations’ recording wind speed and direction are located on Reservoir 3.

for roofing material, etc., burned, or they deteriorate, leaving191

behind bare dry soil with sparse knee-high grass tussocks as192

vegetation. The direct vicinity of the reservoirs is then free of193

vegetation, exposing the bare banks of the reservoirs.194

IV. DATA SETS AND METHODS195

A. Reservoir Bathymetry and Areas196

Bathymetric reservoir models and water level measurements197

serve as ground reference data for comparison with reser-198

voir surface areas determined with ENVISAT images. The199

bathymetric models were generated from GPS tagged water200

depth measurements and reservoir outlines as described by201

Liebe et al. [1]. Water level measurements are used together202

with the bathymetric models to determine the surface area and203

storage volumes of the reservoirs.204

Water levels were measured with pressure transducers at205

15-min intervals. These were used to determine ground ref-206

erence data of reservoir surface areas at the time of image207

acquisition, which are compared with the radar-based results.208

For Reservoirs 1 and 2, water level data are available from209

June 6, 2005 to February 21, 2006 and for Reservoir 3 from210

June 6, 2005 to August 3, 2006.211

B. Wind Speed and Wind Direction Data212

Wind-induced waves and ripples may influence the radar213

signal return from water surfaces. Besides wind speed, the214

wind direction is of importance, as it determines the crest215

orientation of the wind-induced waves. Assuming that wind216

produces waves with crests orthogonal to the wind direction,217

high wind speeds with wind directions orthogonally to the look218

direction may affect the backscatter from the water bodies less 219

than wind directions in line with the look direction. Wind speed 220

was measured on the center of Reservoir 3 and is available 221

from October 2005 onward. In addition, wind speed and wind 222

direction were measured at three locations on the shore of 223

Reservoir 3 (Fig. 1) at 2-min intervals, starting in August 2005. 224

C. ENVISAT Satellite Data 225

In this paper, 22 ENVISAT ASAR acquisitions are used with 226

a roughly bimonthly coverage from June 2005 to August 2006. 227

ENVISAT ASAR is a C-band radar. We used APG images, 228

with a nominal spatial resolution of 30 m and a pixel spacing 229

of 12.5 m. ENVISAT’s ASAR instrument can acquire images 230

in dual-polarization mode and produce HH- and VV-, HH- 231

and HV-, or VV- and VH-polarized image pairs [22]. In this 232

paper, we have chosen dual-polarized acquisitions with the 233

band combinations HH and HV. This combination has also 234

been found useful in flood delineation by Henry et al. [11]. 235

The scenes were acquired from different swaths (IS1–IS6) with 236

incidence angles ranging from an overall minimum of 12.57◦ 237

to a maximum of 43.77◦ throughout the scenes. The look 238

direction is 81.45◦ on ascending and 261.45◦ on descending 239

images. Images were acquired in both ascending (night time 240

acquisition) and descending (morning acquisition) nodes to 241

take into account the diurnal wind patterns, which determine 242

the occurrence of Bragg scattering. 243

D. Classification of Small Reservoir Surface Areas 244

Several approaches were tested to determine the surface areas 245

of the reservoirs, such as the active contour method or snake 246