Evaluation of Underwater Rainfall Measurements ... - IEEE Xplore

3 downloads 0 Views 2MB Size Report
Abstract—Rainfall on the sea surface generates a loud and dis- tinctive sound underwater. This sound propagates downward and attenuates, producing an ...
2936

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 10, OCTOBER 2008

Evaluation of Underwater Rainfall Measurements During the Ionian Sea Rainfall Experiment Marios N. Anagnostou, Jeffrey A. Nystuen, Emmanouil N. Anagnostou, Efthymios I. Nikolopoulos, and Eyal Amitai

Abstract—Rainfall on the sea surface generates a loud and distinctive sound underwater. This sound propagates downward and attenuates, producing an effective listening area or an equivalent “catchment basin” for a listening device that is a function of depth and frequency. Acoustical measurements of rainfall are reported from four passive aquatic listeners (PALs) at 60-, 200-, 1000-, and 2000-m depths from a mooring in the Ionian Sea off the southwestern coast of Greece (37N, 21.5E) from January to April 2004. These measurements are compared to colocated high-resolution X-band dual-polarization (XPOL) radar rainfall measurements. The XPOL radar reports the spatial distribution of rainfall variability over the listening areas of the PALs. Four quality-controlled rainfall events, including drizzle, squall line, and heavy rainfall, are presented in this study. The radar rainfall is spatially averaged over the mooring and compared with the four different acoustic measurements at different depths. To understand the issue of spatial averaging, quantitative comparisons are presented, showing a high correlation between the acoustic measurements and the area-averaged radar estimates at corresponding resolutions. Index Terms—Radar measurements, rainfall, sound level, underwater acoustical measurements, X-band dual-polarization (XPOL) radar.

I. I NTRODUCTION

T

HE AMBIENT sound field in the ocean contains significant information about the physical, biological, and anthropogenic processes in the ocean. The interpretation of the ambient sound field can be used to quantify these processes. In particular, the underwater ambient sound that is generated by raindrops striking the ocean surface and trapping bubbles has a unique and distinctive signal in the frequency band of 100–50 000 Hz that can be used to detect and quantitatively measure rain at sea [5]. Furthermore, different raindrop sizes produce a distinctive sound underwater, allowing the inversion of the sound to retrieve the drop size distribution (DSD) within Manuscript received September 30, 2007; revised January 23, 2008 and March 19, 2008. Current version published October 1, 2008. This work was supported by the National Science Foundation through the Physical Oceanography Division under Grant 0241245. M. N. Anagnostou is with the Institute of Inland Waters, Hellenic Center for Marine Research, 19013 Anavissos, Greece (e-mail: managnostou@ath. hcmr.gr). J. A. Nystuen is with the Applied Physics Laboratory, University of Washington, Seattle, WA 98105-6698 USA. E. N. Anagnostou and E. I. Nikolopoulos are with the Institute of Inland Waters, Hellenic Center for Marine Research, 19013 Anavissos, Greece, and also with the Department of Civil and Environmental Engineering, University of Connecticut, Storrs, CT 06269-2037 USA. E. Amitai is with the National Aeronautics and Space Administration Goddard Space Flight Canter, Greenbelt, MD 20771 USA, and also with Chapman University, Orange, CA 92866 USA. Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TGRS.2008.2000756

Fig. 1. Overview of the experimental setup showing the ground validation (2-DVD and gauges), XPOL radar, and mooring locations.

the rain [11], [13]. This allows the potential for rainfall and radar reflectivity detection, classification (drizzle, stratiform, and convective), and quantification over the sea using underwater acoustic measurements [12]. One interesting feature of the acoustical measurement is that the listening area for a hydrophone, its effective “catchment basin,” is proportional to its depth, and yet, the signal should be independent of depth if the sound source is uniformly distributed on the sea surface. Thus, the acoustical measurement of rainfall has an inherent spatial averaging that can be compared to the beam filling assumption of radar or satellite measurements of rainfall. By making sound measurements at different depths (in this study, we used 60, 200, 1000, and 2000 m) and comparing those measurements to simultaneous highresolution dual-polarization X-band (XPOL) radar observations, the spatial averaging of the acoustic signal can be explored and evaluated. In Section II, we describe the study area and data associated with the Ionian Sea Rainfall Experiment (ISREX). In Sections III, we describe the acoustic and radar analysis, whereas in Section IV, we describe the acoustic and radar rainfall algorithms. In Section V, we discuss the aspect of acoustic spatial averaging effect on the rainfall signal, whereas in Section VI, we present radar versus acoustic rainfall comparisons, focusing on four rainfall events, including drizzle, squall line, and heavy rainfall. In Section VII, we provide our conclusions. II. ISREX A. Experimental Site From February to March 2004, ISREX took place in the Ionian Sea off the southwestern coast of Greece (Fig. 1). This

0196-2892/$25.00 © 2008 IEEE

ANAGNOSTOU et al.: EVALUATION OF UNDERWATER RAINFALL MEASUREMENTS

2937

TABLE I RAINFALL EVENTS RECORDED IN ISREX

location offers deep water (more than 3 km deep) within the coverage area of a mobile high-resolution XPOL radar. Four acoustic sensors were deployed at 60-, 200-, 1000-, and 2000-m depths on a single mooring. The coastal radar was located at Methoni, approximately 17 km east of the mooring. A dense rain gauge network and a 2-D video disdrometer (2-DVD) were deployed at Finikounda, 10 km east of the radar but in the opposite direction from the mooring. The acoustical measurements and rain gauge network measurements were continuous from mid-January to mid-April, whereas the radar was manually operated during storm events. Six precipitation events were captured by all three systems. Table I lists the rainfall events that were detected and recorded during the experiment. It also gives the rainfall accumulation totals (in millimeters) for each system (i.e., XPOL radar, rain gauge stations at Finikounda, and rain gauge at the Methoni station). Several storms had long periods (several hours) of continuous coverage. For these events, 1099 radar scans were recorded. The total gauge accumulations at times that the radar was operating were between 90 and 110 mm, with only four events associated with more than 10-mm accumulation. Most of the rain events recorded by the radar were classified as widespread (weak rain intensities) with only a few events exhibiting radar returns (ZH ) above 40 dBZ (e.g., February 12). During March 8 and 9, the rain systems that passed over the mooring were characterized as squall lines. A more organized system, moderate in intensity, with a duration greater than 2 h, is the March 12 case.

B. Acoustic Data The acoustic data were collected on four passive aquatic listeners (PALs). PALs consist of a low-noise wideband hydrophone, signal preamplifiers, and a recording computer. The nominal sensitivity of these instruments is −160 dB relative to 1 V/μPa with an instrument noise equal to an equivalent oceanic background noise level of about 28 dB relative to 1 μPa2 Hz−1 . Bandpass filters are present to reduce saturation from low-frequency sound (high pass at 300 Hz) and aliasing from above 50 kHz (low pass at 40 kHz). The hydrophone sensitivity also rolls off above its resonance frequency, about 40 kHz. A further sensitivity correction due to the depth of deployment is also present. A data collection sequence takes about 15 s and consists of four 10.24-ms time series, each separated by 5 s. Each of these time series is fast Fourier transformed

(FFT) to obtain a 512-point (0–50-kHz) power spectrum. Geophysically generated sounds from rain, drizzle, or wind are generally stationary over a 15-s time interval, whereas banging from ships or moorings or chirps, whistles, or clicks from biological sources are sound signals that are usually nonstationary over a 15-s time interval. Thus, a preliminary evaluation of the sound source is to remove nonstationary data samples. The four spectra are then averaged into a single spectrum and evaluated to determine the acoustic source (rain, wind, or drizzle). A sensitivity adjustment of instrument, depth, and local ocean conditions (overall higher to lower sound levels) is determined by choosing a sound condition where the signalto-noise ratio is high and assuming a uniform sound source at the surface. At low wind speeds, the recorded signal includes a component from the ambient background and from instrument noise. At high wind speeds, there is a change to the spectral shape of the wind signal due to attenuation of the signal from ambient bubbles in the water. However, at moderate wind speeds (4–8 m/s), the sound signal is well above the background noise and has a uniform spectral slope between 1 and 40 kHz. This signal is adjusted for absorption and depth. After adjusting for absorption, an unexpected depth-dependent offset remains. At 2 and 10 kHz, there is a 0.5-, 3-, and 6-dB offset for the PALs at 200-, 1000-, and 2000-m depth, respectively. Ocean currents will bend the mooring, causing a horizontal displacement of the acoustic sensors. In order to determine this potential displacement, a pressure sensor was placed on the mooring 2 m below the shallowest acoustic recorder. During most of the experiment, the mooring line can be assumed to be vertical, but there are a few episodes when the top of the mooring dips and horizontal displacement is assumed to be present.

C. Radar Data High-resolution radar data were collected with the National Observatory of Athens (NOA) mobile XPOL radar located at the Methoni meteorological station (see Fig. 2). XPOL is a low-power (50-kW peak power) radar with selectable pulse and simultaneous transmission of signal at horizontal and vertical polarization. The antenna is mounted on the back of a truck with 8-ft radius and has a 0.95◦ (3-dB) beamwidth. During each storm event, the radar was in the planar position indicator (PPI) scan mode at a low elevation (∼ 2◦ ) above the horizon. The pulse repetition frequency was 1000 Hz with 150-m gate length

2938

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 10, OCTOBER 2008

A sound source at a free surface is an acoustic dipole. Thus, the sound intensity recorded at an omnidirectional hydrophone is given by  (1) I = Io cos2 θ atten(p)dA

Fig. 2. NOA mobile XPOL radar during ISREX deployment in Southwest Greece.

(range resolution) and 400 gates to give a total range of 60 km. Thus, the nominal backscattering volume at the range of the mooring is 150-m range by 300-m azimuth at an elevation of roughly 500 m above the ocean surface. The scanning rate was less than 1 min. D. Ground Validation Network XPOL rainfall measurement validation is based on the dense network of rain gauges (typical tipping bucket) and a 2-DVD deployed within a 1-km2 area and roughly 10 km away from XPOL (see Fig. 1). The gauge network consists of six dualgauge clusters and a site consisting of three gauges and the 2-DVD. These data are used to correct the radar rainfall rate estimate for biases during each rain event. There is a weather station at Methoni that reports on an hourly basis the wind speed direction and rainfall amounts. These data are available during local working hours (0500– 1800 local time). The station location is 17 km east of the mooring site and is next to the XPOL deployment site. Wind speed comparisons from the Methoni station and PALs showed up to 2-m/s differences. Acoustic wind speed measurements are typically ±0.5 m/s when compared to colocated surfacemounted anemometers [10], and thus, this discrepancy is likely to have an orographic explanation. III. D ATA A NALYSIS A. Acoustical Data Ambient sound in the ocean is a combination of natural and man-made sounds. Various physical processes, including wind, rain, and drizzle, are the primary sound sources in the frequency range from a few hundred hertz to 50 kHz. These are sound sources at the sea surface. The microphysics of the sound generation is resonating bubbles created during the splashing of wind waves or raindrop splashes [7], [9], [11]. These bubbles are very near the free surface of the ocean and, consequently, are assumed to behave as vertically oriented acoustic dipole sources. Man-made devices, which include ships and sonar, and different marine animals, particularly cetaceans, produce sound underwater in this same frequency band. An underwater acoustic recorder will hear all of these sounds. Acoustic monitoring requires for the sound to be recorded, for its source to be identified, and then for it to be quantified.

where I is the sound intensity measured at the hydrophone, Io is the sound intensity at the surface, cos2 θ is the directional radiation pattern for a surface source (vertically oriented dipole), and atten(p) is the attenuation along the acoustic path p. The integral is taken over the surface area A. The attenuation is due to geometric spreading, scattering, and absorption. Scattering is assumed to be negligible. Geometric spreading is affected by the sound speed profile of the ocean. Thus, the attenuation from a surface source to an acoustic recorder is given by atten(p) =

exp(−αp) r 2 + h2

(2)

where p2 = r2 + h2 , r is the horizontal range to the sound source, h is the depth of the recorder, and α = α(f, S, T ) is the absorption of sound in seawater and is a function of frequency, salinity, and temperature [8]. The temperature and the salinity structure of the ocean affect the sound speed profile, which, in turn, will cause a refraction of sound. The temperature, salinity, and sound speed structure of the upper 500 m at the mooring site were measured during the deployment (January 14) and recovery (April 14) of the mooring. If one makes the assumption that I0 is uniform over the sea surface, then (1) can be used to estimate the effective listening area for each sensor. Fig. 3(a) shows the fraction of the total energy received as a function of frequency for sensors at 60-, 200-, 1000-, and 2000-m depths. If the listening radius for a sensor is defined as the area receiving 90% of the sound, then the listening radii for the sensors at 2 kHz are 172, 620, 2800, and 5000 m, respectively, roughly three times the depth of the sensor. At 20 kHz, the radii are 165, 550, 2000, and 3400 m, respectively. The principal rainfall signal is at 5 kHz. At 5 kHz [Fig. 3(b)], the radii are 170, 610, 2735, and 4800 m, respectively. Fig. 3(a) and (b) also shows the weighting function of the listening radius for a uniform surface source. Within the defined listening area, most of the energy is arriving from a much smaller area centered over the mooring. For example, at 5 kHz, 50% of the energy is arriving from the surface area with radii of 58, 200, 970, and 1860 m, respectively, which is roughly the depth of the sensor. Different sound sources are identified by their spectral characteristics. Features of sound source spectra that can be used to identify the source include spectral levels at various frequencies, ratios of these levels, spectral slopes, and the temporal persistence of the sound source. The data were examined to find times when the sound source could be confidently assumed. Long periods (hours) of steady uniform sound were assumed to be periods of constant wind. Short loud events consistent with typical ship spectra (very loud at low frequency) during nonrainy periods were assumed to be ships. Distinctive rain and drizzle spectra were identified and confirmed with radar. These “typical” sound sources are shown in Fig. 4, taken as an example from February 12, and were used to build an acoustic classification algorithm that can be used to objectively identify

ANAGNOSTOU et al.: EVALUATION OF UNDERWATER RAINFALL MEASUREMENTS

2939

Fig. 4. Typical sound spectra for different periods of February 12, 2004 from PAL N at 200-m depth. Curves A–J refer to the different periods reported in Table II. TABLE II TIME PERIODS WITH ACOUSTIC INTERPRETATION FOR F EBRUARY 12, 2004

Fig. 3. (a) Listening radius for the four PALs (M at 58 m, N at 208 m, O at 1008 m, and P at 2008 m) as a function of frequency. The listening radius is reduced at higher frequency by higher absorption of sound. Each group of curves shows this frequency effect from (left) 50 kHz to (right) 1 kHz. (b) Listening radius of the four PALs (M at 60 m, N at 200 m, O at 1000 m, and P at 2000 m) for the 5-kHz frequency.

the sound source in the remaining data (Table II). The goal is to reliably detect the sound source so that the subsequent analysis is not contaminated by sound generated by another source. Fig. 5 shows the relationship between 8 and 20 kHz for these test-case sound sources. This comparison of sound levels at two frequencies is particularly illustrative for demonstrating the ability to use ambient sound to identify the sound sources. In fact, multiple measures are actually used. For example, there is an ambiguity for the sound source “wind = 12 m/s” and “ships.” Other features of the sound field are needed to separate these two sources. For this situation, periods of high wind are usually of long durations (hours), whereas ships pass the mooring in minutes. However, there may be other spectral measures that are diagnostic. Other useful measures are the sound levels at 2 kHz, the ratio of sound levels from 1 to 2 kHz, the slope of the spectrum from 2 to 8 kHz, the slope of the spectrum from 8 to 15 kHz, and the sound level at 20 kHz.

These relationships change as a function of depth because of absorption (see the panel comparison in Fig. 5). Since the relationships are used to identify the sound sources, a correction for absorption needs to be applied. For uniform surface sound sources such as wind, the relationships become independent of depth. For nonuniform sound sources such as a ship, relatively less low-frequency sound is present, and the relative position of the sound source on the diagram changes. The classification of loud ships at depth is more difficult as their spectra begin to sound like heavy rainfall. B. Radar Data Radar observations at X-band suffer from a number of error sources that manifest themselves as either systematic or random. Both types of errors are important, and they need to be assessed and accounted for because they can introduce significant uncertainty and lead to erroneous results in radar rainfall estimation. The radar data used in this study were corrected for atmospheric attenuation effects based on the algorithm described in [1] and were calibrated using the 2-DVD as a reference, in order to account for the systematic measurement errors.

2940

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 10, OCTOBER 2008

Fig. 5. Scatter diagrams as a function of depth (60, 200, 1000, and 2000 m). The solid curve is at the same location on each subfigure. Absorption is greater at higher frequencies.

A comparison between disdrometer and radar’s time series of horizontal reflectivity ZH and differential reflectivity ZDR (ratio of the horizontal to the vertical polarization of the electromagnetic wave) revealed a bias. In order to account and correct for this bias, we adjusted the radar measurements according to ZCAL = ZRAW + C

(3)

C = E{ZDisd − ZRAW }

(4)

with

where ZCAL (in decibels) is the calibrated radar variable (ZH or ZDR ), ZRAW (in decibels) is the raw (uncalibrated) radar variable, ZDisd (in decibels) is the simulated radar variable from the disdrometer, and C (in decibels) is the calibration constant. The calibration scheme was separately applied to each storm, and on the average, the results revealed a calibration constant of 1.5 dBZ for ZH and 0.5 dB for ZDR . Sample PPI plots of XPOL reflectivity measurements after bias adjustment and attenuation correction for the four storm events used in this study are shown in Fig. 6. In Fig. 7, we present an example of the comparison between radar and disdrometer measurements after bias adjustment. As we can observe, both ZH and ZDR show good agreement, with that of ZDR exhibiting more scatter. In Table III, the comparative results for all storms are quantified by showing the correlation and the standard deviation of the error (defined as the difference) between

radar and disdrometer measurements. The correlation values for ZH are very high (> 0.9), whereas those for ZDR vary from moderate (∼0.6) to high (∼0.86). The standard deviation provides a measure for the observed scatter, and the values obtained are within an acceptable range (1.4–1.8 dBZ and 0.3–0.5 dB for ZH and ZDR, respectively). Overall, those results give us confidence in the performance of the radar and its subsequent use as an independent reference for the comparison with the PAL estimates. IV. R AINFALL A LGORITHMS A. Acoustical Rainfall Algorithm The acoustic rainfall rate algorithm is a simple power-law relationship between sound intensity and rainfall (in millimeters per hour), which can be written in the form of b I = a · Rpal

(5)

where I is the sound intensity, Rpal is the PAL’s rainfall rate, and α and b are empirically determined parameters [5]. At 5 kHz, these are determined by Ma and Nystuen [5] to have values of 42.5 and 15.4, respectively. Taking the 10 log10 of (5) at 5 kHz, this becomes   SP L   SP L 5kHz −α 5kHz −42.5 β 15.4 = 10 (6) R = 10

ANAGNOSTOU et al.: EVALUATION OF UNDERWATER RAINFALL MEASUREMENTS

2941

Fig. 6. Sample PPI plots of horizontal polarization reflectivity (ZH in decibels of equivalent reflectivity) that are bias-adjusted and corrected for rain path attenuation for the four storm cases.

where SP L5kHz is the sound pressure intensity at 5 kHz in decibels relative to 1 μP2a Hz−1 , and the coefficients α = 10 log10 (α) = 42.5 and β = 10 · b = 15.4 are from [5]. The acoustic rainfall signal at 5 kHz is chosen because this frequency has a large dynamic range associated with the rainfall rate, and the influence of wind on the rain-induced sound level is thought to be small [11]. A point to note is that the above relationship is based on empirical rainfall comparisons from a tropical ocean environment. B. Radar Rainfall Algorithm In this paper, we used a multiparameter radar rainfall estimation algorithm that is divided into two separate parts.

R=

⎧ ⎨ ZH < 5,

R= 0

⎩ ZH ≥ 5,

R=

The first part uses the standard Z–R relationship to convert the measured horizontal polarization reflectivity ZH (in decibels of equivalent reflectivity) to rainfall rate R (in millimeters per hour), and it is applied to low reflectivity values (< 20 dBZ). The second part combines the information of the differential reflectivity ZDR (in decibels) and ZH to derive an estimate for the rainfall rate. The details of the combined algorithm are summarized in (7), shown at the bottom of the page. The algorithm coefficients were derived based on least square regression between 3-min disdrometer-measured rainfall values R and disdrometer-calculated ZH and ZDR parameters from the 2-DVD-observed spectra. There is variability in the determined coefficients. The coefficient values for the different rain events are summarized in

ZH > 20 dBZ and ZDR > 0.1 dB, else,

b1 c 1 R = a1 ZH ZDR b2 R = a2 ZH

(7)

2942

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 10, OCTOBER 2008

Fig. 7. Scatterplot of calibrated (left) ZH and (right) ZDR parameters measured by radar and derived from disdrometer (2-DVD) spectra for the storm events of (upper panel) February 12 and (lower panel) March 12. TABLE III STANDARD DEVIATION OF THE DIFFERENCE AND CORRELATION OF ZH AND ZDR PARAMETERS FROM DISDROMETER AND RADAR

parameters to represent all storms would introduce biases in rainfall estimation. However, even on a storm-by-storm basis, the derived parameters do not provide a perfect model to relate radar measurements with the actual surface rainfall, which introduces a random error in rainfall estimation. V. S PATIAL A VERAGING E FFECT ON THE R AINFALL S IGNAL

TABLE IV COEFFICIENTS OF RADAR RAIN RELATIONSHIPS

Table IV. For the Z–R parameters, the multiplier α2 varies from −47% (February 12) to +200% (March 12), and the exponent b2 varies from −41% (March 12) to +22% (February 12) with respect to the parameters from “all storms combined.” Similarly, for the multiparameter (ZH –ZDR ) relationship, the α1 varies from −19% (March 12) to +3% (February 12), whereas b1 varies from −10% (February 12) to +30% (March 12), and c1 varies from −30% (March 8) to approximately 126% (March 12). The variability of the above parameters indicates variability in the raindrop size distribution of the different storms. Consequently, the use of a single set of

Four examples of coincident observations of relative continuous rainfall are used to investigate the spatial averaging of the acoustic rainfall signal (February 12 and March 8, 9, and 12). The rainfall rates from PALs are correlated to averaged rainfall rates from the radar for different averaging radii in a circle centered over the mooring location (Fig. 8). As Amitai et al. [2] have found, the averaging radii producing the highest correlation between the radar and the PALs increase from 1 to 4 km for the measurements at 60 m, 200 m, 1 km, and 2 km, respectively. The predictions assuming a uniform weighting of the acoustic signal within the listening area are 170, 610, 2735, and 4800 m, respectively, which is consistent with the general XPOL/PAL correlation results. Thus, the signal from the shallowest hydrophone is most highly correlated to the radar rainfall measurement averaged over a larger listening area than expected, and that from the deeper hydrophones is most highly correlated to smaller areas than expected. It was noted that the weighting of the acoustic signal is not uniform but is weighted toward the center of the listening area,

ANAGNOSTOU et al.: EVALUATION OF UNDERWATER RAINFALL MEASUREMENTS

2943

compared to the averaged radar rainfall rates combining all four rain events. The averaging radii producing the highest correlation between the radar and the PALs increase from 600 to 1000, 2300, and 3400 m for the PAL depths at 60, 200, 1000, and 2000 m, respectively, showing that the trend is as expected. VI. XPOL/PAL R AINFALL C OMPARISON

Fig. 8. Correlation coefficients between the acoustic rainfall measurements at 60-, 200-, 1000-, and 2000-m depth compared to the averaged radar rainfall rates averaged over circles of different radii centered over the mooring. The radar cell size is 150 m.

Once rainfall has been acoustically detected [11], [12], a comparison of rainfall rates with the radar rainfall is possible. In this section, we compare time series of rainfall estimates from the four major rainfall events of ISREX. Both XPOL and PAL rainfall rates are averaged at 3-min temporal resolution. To better assess and evaluate the comparison between XPOL and PAL, we selected the average radii of the radar that gives the highest rainfall correlation between XPOL and PAL as indicated in the correlation analysis presented in Section V (Fig. 8). PALs from the four depths are compared to the corresponding average XPOL radii rainfall estimates through time series and bulk statistics such as bias, relative root-mean-square error (rRMSE), and correlation Bias =

Total Rref Total Rest



N (R (i)−R (i))2 est ref

(8)

i=1

N

rRMSE = 1 N

N

Rref (i)

i=1

N 

ρ = i=1 N 

Rref (i) − Rref

Rref (i) − Rref

i=1

Fig. 9. Correlation coefficients between the acoustic rainfall measurements at 60-, 200-, 1000-, and 2000-m depth compared to the averaged radar rainfall rates averaged over circles of different radii centered over the mooring combining the four dates.

thus reducing the expected listening area for the hydrophones. For example, the average radius shows maximum correlations with the deepest hydrophone, ranging between 2 and 4 km. In contrast, the averaging radii producing the highest correlation for the PAL at 60 m go from 450 m for the squall line on March 8 to 1 and 2 km, respectively, for the rain events on February 12 and March 12. These events were more widespread with longer spatial and temporal time scales, suggesting that the physical scale of the rain itself may be responsible for this result. Nevertheless, the trend is as expected. The deeper acoustic measurements have higher correlations to the radar rainfall averaged over larger apparent listening areas. The intermediate PALs at 200 and 1000 m show intermediate results. This is clearly exhibited in Fig. 9, showing the correlation coefficients between the acoustic rainfall measurements at different depths

(9)



Rest (i) − Rest

N  2



Rest (i) − Rest

2

.

(10)

i=1

In the above statistics, Rref and Rest represent the 3-minresolution rainfall values from the average XPOL radii and PAL, respectively. The time series comparisons are shown in Fig. 10(a)–(d), where the blue line is the rainfall from the Radarxkm , the black from the PALxkm , and the red line (PALxkm /2) is half the rainfall rate of the PALxkm . The associated bulk statistics are summarized in Table V. In the following, we provide a discussion on the PAL–XPOL comparisons for the four ISREX rainfall events used in this study. A. Case Study: February 12 On February 12, 2004, a frontal system passed over the mooring. This was a complicated system consisting of several rain cells and lulls. The acoustic record of the day contains a wide variety of signals, including distant shipping, a close ship passage, some loud noises, and the sound of high wind conditions. A time-series plot between the different optimum radar radii rainfall rates given for each one of the four PALs is shown in Fig. 10(a). In this case, the XPOL average radius is 1 km for the comparisons with the shallowest PAL and 1.9 km for the 200-m PAL. In the deeper two panels, we used 3- and

2944

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 10, OCTOBER 2008

Fig. 10. Time-series comparison of XPOL versus PAL for the four storm cases. (a) February 12. (b) March 8. PALxkm /2 is defined as 1/2 of the values at x km. (c) March 9. (d) March 12. PALxkm /2 is defined as 1/2 of the values at x km.

4-km average radii. It is clear that in all four depths, there are good correlations (see Table V) with low rRMSE, and the bias of the estimated rainfall of the four PALs is about two times the rainfall of XPOL. The rainfall from all four PALs and the radar matched the peak of the storm at minutes 325 and 355 [Fig. 10(a)].

B. Case Study: March 9 and 12 Both of these events had long periods of rainfall with codetection by both the radar and the PALs. On March 9, shipping contamination is detected from minutes 725–750 at the shallow PALs, and these data are discarded from further analysis. Fig. 10(c) shows the time series between the four different PALs and the different optimum average radar radii shown in Fig. 8. The correlation here is low, and the bias is lower than that in the other events (in the range of 0.65–0.7). On March 12 [Fig. 10(d)], the rainfall starts as a relatively strong rain (6 mm/h) and then tapers off after minute 910. The wind speed is relatively high during this event, with acoustic wind speed

measurements of ∼10 m/s before the start of the event and 8 m/s at the Methoni weather station (17 km to the east). On March 12, a very good correlation is clearly shown, and the bias of the PAL rainfall rates is about two times that of XPOL, which is associated with contamination by strong winds that day. We notice a very good catch of the peak of the storm at minutes 890–900 in the rainfall time series from all the four PALs and the four different average radii of the radar. From this comparison, there is evidence of an increase in effective listening area with increasing listening depth with high correlations (see Table V). C. Case Study: March 8 Perhaps the most serendipitous rain event was an isolated squall line that passed over the mooring on March 8. Based on the echolocation position of the squall from consecutive radar scans, the local wind speed was estimated to be 7 m/s. The acoustic wind measurement at the beginning of the squall is 9 m/s, which is in agreement with the wind measurement at the Methoni weather station. A comparison of time-series rainfall

ANAGNOSTOU et al.: EVALUATION OF UNDERWATER RAINFALL MEASUREMENTS

2945

TABLE V XPOL VERSUS PAL RAINFALL STATISTICS, INCLUDING BIAS (CORRELATION AND rRMSE), PRESENTED FOR THE FOUR RAINFALL EVENTS

rates is shown in Fig. 10(b). From the comparison of rainfall rates between the four different depths and the average radar radii that give the highest correlation, it is evident that both the PALs and the radar match very well the peak of the storm with high correlations, and the bias in all four PALs is in the range of 0.5 (see Table V for the statistics).

VII. C ONCLUSION High-frequency (1–50-kHz) acoustic measurements of the marine environment at four depths (60, 200, 1000, and 2000 m) are used to describe the physical, biological, and anthropogenic processes present at a deep-water mooring site near Methoni, Greece, from mid-January to mid-April in 2004. The main focus of the experiment was the detection of precipitation from the acoustic measurements in deep waters at different depths and validated with the rainfall from the coastal highresolution XPOL radar. XPOL radar measurements were quality controlled and corrected for attenuation. A multiparameter rainfall algorithm was used to retrieve rainfall over the mooring site. XPOL rainfall was then averaged at different radii and compared to the PAL rainfall measurements. Eight events were recorded from PALs and six from the radar. Of those, four major rain events with rain accumulations greater than 10 mm were used in this study. The radar data were used to verify the acoustic classification of rainfall and the acoustic detection of embedded shipping noise within a rain event. In this paper, the four-rain-event record of radar and acoustic rainfall measurement was compared to investigate averaging of the acoustic rainfall signal as a function of listening depth. The comparison shows an increase in effective listening area with increasing listening depth. The averaging radius for the radar data that resulted in the highest correlation with the rainfall measurement for the shallowest hydrophones (at 60 and 200 m) was about 1.0 km, suggesting a listening area for the shallowest hydrophone between 1 and 12 km2 , which is larger than expected, whereas the averaging radius for the deepest hydrophone (at 2000 m) was between 3 and 4 km, suggesting a listening area between 30 and 50 km2 , which is less than expected. For the highest correlation PAL/XPOL matching values, we determined an acoustic (PAL) overestimation of rainfall in the range of 100%. One can say that this issue is due to the PAL’s rainfall relation (6) that derived from observations in a tropical ocean environment and applied to observations in the Ionian Sea (with different DSDs) or due to the wind contamination to acoustic rainfall rate measurement. There is a need to continue our experimental effort to enhance our understanding of acoustic rainfall estimation. New

questions include the following: 1) Is the change in the length scale of maximum correlation due to the spatial structure of the rain event? 2) If so, can information about the spatial structure of rain be part of the acoustic rainfall detection process? 3) What is the influence of wind on acoustic rainfall classification? 4) Can the wind effect be incorporated into the acoustic rainfall type classification algorithms? 5) What is the influence of wind on acoustic rainfall rate measurement? The combined influence of wind and rain on sound levels in the ocean has been modeled using data from the tropical Pacific Ocean [5]. This model needs to be inverted to extract the acoustic rainfall signal in the presence of wind. The calibrated radar data from ISREX will be used to model and constrain this inversion. ACKNOWLEDGMENT The authors would like to thank E. Boget for designing and deploying the deep-water mooring, NOA for making the XPOL radar available to the experiment, Dr. G. Chronis at the Hellenic Center for Marine Research for providing the vessel “Filia” used to deploy the mooring, T. Paganis and A. Gomta at the Methoni weather for providing the Methoni meteorological data, and the citizens of Finikounda for allowing rain gauges to be set up in their yards during the experiment. R EFERENCES [1] M. N. Anagnostou, N. Anagnostou, and J. Vivekananda, “Correction for rain path specific and differential attenuation of X-band dualpolarization observations,” IEEE Trans. Geosci. Remote Sens., vol. 44, no. 9, pp. 2470–2480, Sep. 2006. [2] E. Amitai, J. A. Nystuen, E. N. Anagnostou, and M. N. Anagnostou, “Comparison of deep underwater measurements and radar observations of rainfall,” IEEE Geosci. Remote Sens. Lett., vol. 4, no. 3, pp. 406–410, Jul. 2007. [3] P. Barber and C. Yeh, “Scattering of electromagnetic waves by arbitrarily shaped dielectric bodies,” Appl. Opt., vol. 14, no. 12, pp. 2864–2872, Dec. 1975. [4] E. A. Brandes, G. Zhang, and J. Vivekanandan, “Experiments in rainfall estimation with polarimetric radar in a subtropical environment,” J. Appl. Meteorol., vol. 41, no. 6, pp. 674–685, Jun. 2002. [5] B. B. Ma and J. A. Nystuen, “Passive acoustic detection and measurements of rainfall at sea,” J. Atmos. Ocean. Technol., vol. 22, no. 8, pp. 1225–1248, 2005. [6] S. Y. Matrosov, K. A. Clark, B. E. Martner, and A. Tokay, “X-band polarimetric radar measurements of rainfall,” J. Appl. Meteorol., vol. 41, no. 9, pp. 941–952, Sep. 2002. [7] H. Medwin and M. M. Beaky, “Bubble sources of the Knudsen Sea noise spectra,” J. Acoust. Soc. Amer., vol. 86, no. 3, pp. 1124–1130, Sep. 1988. [8] H. Medwin and C. S. Clay, Fundamentals of Acoustical Oceanography. New York: Academic, 1998. [9] H. Medwin, J. A. Nystuen, P. W. Jacobus, L. H. Ostwald, and D. E. Synder, “The anatomy of underwater rain noise,” J. Acoust. Soc. Amer., vol. 92, no. 3, pp. 1613–1623, Sep. 1992.

2946

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 46, NO. 10, OCTOBER 2008

[10] J. A. Nystuen, M. J. McPhaden, and H. P. Freitag, “Surface measurements of precipitation from an ocean mooring: The underwater acoustic log from the South China sea,” J. Appl. Meteorol., vol. 39, no. 12, pp. 2182–2197, Dec. 2000. [11] J. A. Nystuen, “Listening to raindrops from underwater: An acoustic disdrometer,” J. Atmos. Ocean. Technol., vol. 18, no. 10, pp. 1640–1657, Oct. 2001. [12] J. A. Nystuen and E. Amitai, “High temporal resolution of extreme rainfall rate variability and the acoustic classification of rainfall,” J. Geophys. Res.—Atmos., vol. 108, no. D8, pp. 8378–8388, 2003. [13] J. A. Nystuen, “Using underwater sound to determine drop size distribution,” in Sounds in the Seas: Introduction to Acoustical Oceanography, H. Mdewin, Ed. Cambridge, U.K.: Cambridge Univ. Press, 2005. [14] S.-G. Park, V. N. Bringi, V. Chandrasekar, M. Maki, and K. Iwanami, “Correction of radar reflectivity and differential reflectivity for rain attenuation at X-band. Part I: Theoretical and empirical basis,” J. Atmos. Ocean. Technol., vol. 22, no. 11, pp. 1621–1632, Nov. 2005. [15] J. Testud, E. Le Bouar, E. Obligis, and M. Ali-Mehenni, “The rain profiling algorithm applied to polarimetric weather radar,” J. Atmos. Ocean. Technol., vol. 17, no. 3, pp. 332–356, Mar. 2000. [16] G. Vulpiani, F. S. Marzano, V. Chandrasekar, and S. Lim, “Constrained iterative technique with embedded neural network for dual-polarization radar correction of rain path attenuation,” IEEE Trans. Geosci. Remote Sens., vol. 43, no. 10, pp. 2305–2314, Oct. 2005.

Marios N. Anagnostou received the B.S. and M.Eng. degrees in electrical engineering from York University, Toronto, ON, Canada, and the Ph.D. degree in environmental engineering from the University of Connecticut, Storrs. He is currently a Contractor Research Associate with the Institute of Inland Waters, Hellenic Center for Marine Research, Anavissos, Greece. He has participated in numerous international field experiments deploying a mobile dual-polarization radar. He has lately been involved in underwater acoustics for the quantitative estimation of rain and wind. He is the author or coauthor of more than 13 journal papers in the areas of dual-polarization radar rainfall estimation and hydrometeorological applications. His main research experience and interests are on rainfall microphysics and precipitation remote sensing on the basis of new radar remote-sensing systems such as X-band dual polarization (polarimetric).

Jeffrey A. Nystuen received the B.A. degree in mathematics and physics from the University of Michigan, Ann Arbor, in 1979 and the Ph.D. degree in oceanography from the Scripps Institution of Oceanography, La Jolla, CA, in 1985. He is currently a Principal Oceanographer with the Applied Physics Laboratory, University of Washington, Seattle. He is an internationally recognized leader in the field of acoustical oceanography. His specialty is the analysis of the ambient sound field to detect and measure geophysical quantities, especially precipitation. Dr. Nystuen received the 2003 Medwin Prize in Acoustical Oceanography in recognition of his work for the development of the acoustical measurement of oceanic rainfall rate and precipitation type, and has been elected as Fellow of the Acoustical Society of America.

Emmanouil N. Anagnostou received the B.S. degree in civil and environmental engineering from the National Technical University, Athens, Greece, and the M.S. and Ph.D. degrees in civil and environmental engineering from the University of Iowa, Iowa City. He is currently an Associate Professor with the Department of Civil and Environmental Engineering, University of Connecticut, Storrs. He is also with the Institute of Inland Waters, Hellenic Center for Marine Research, Anavissos, Greece. His research interests are the development of techniques for the remote sensing of precipitation parameters ranging from lightning detection to the retrieval of precipitation profiles and surface rainfall from satellite and ground-based sensors, and the optimal integration of rainfall remote-sensing products in hydrologic modeling systems for the prediction of floods and studying regional water and energy cycles.

Efthymios I. Nikolopoulos received the B.Eng. degree in environmental engineering from the Technical University of Crete, Crete, Greece, and the M.Sc. degree in environmental engineering from the University of Iowa, Iowa. He is currently working toward the Ph.D. degree at the University of Connecticut, Storrs. He is also with the Institute of Inland Waters, Hellenic Center for Marine Research, Anavissos, Greece. His research interests are the remote sensing of precipitation, error propagation of satellite rainfall through hydrologic models to evaluate the potential use of satellite observation for hydrologic applications, and, particularly, flood predictions.

Eyal Amitai received the M.Sc. and Ph.D. degrees in atmospheric sciences from the Hebrew University of Jerusalem, Jerusalem, Israel, in 1991 and 1996, respectively. Since 1996, he has been with the Tropical Rainfall Measuring Mission (TRMM), National Aeronautics and Space Administration (NASA) Goddard Space Flight Center, Greenbelt, MD. He spent two years on a postdoctoral fellowship with the Universities Space Research Association. From 1998 to 2002, he was a research faculty member with the University of Maryland, Baltimore County. From 2003 to 2008, he was a Research Professor with the Center for Earth Observing and Space Research, George Mason University, Fairfax, VA. Since July 2008, he has been with Chapman University, Orange, CA. His research interests include hydrometeorology, radar hydrology, and remote sensing of rainfall from space, ground, and underwater platforms. His research is focused on combining information from a variety of sensors to evaluate and improve precipitation estimates.