25nd Scandinavian Symposium on Physics Acoustics Ustaoset 27-30 January 2002 Fish detection in rivers with split-beam sonars Helge Balk and Torfinn Lindem University of Oslo, Department of Physics, Pb 1048 Blindern, 0316 Oslo, Norway. Fax: 22856422. Email: helge.balk@fys.uio.no, torfinn.lindem@fys.uio.no. Introduction For the management, knowledge of the fish stocks is of great interest. Migrating fish in a river can be monitored with mechanical, conductivity, or optical sensors. These sensors have short range and the fish has to be led close by. When this is inconvenient or not possible, acoustic methods provide an alternative. Especially the split beam transducer with its capability of directly detecting the targets acoustic size and 3D-position has become popular [8]. Placed horizontally normal to the river at a suited place, the conical sound beam can cover major parts of the river cross-section and register passing fish. This method has been studied through a four years Dr. Scient. project:"Development of hydroacoustic methods for fish detection in shallow water" [2]. A summery of experiences and results is given in this proceeding. Method The study was carried out by in turns performing fieldwork and analysing recorded material. Data collected by other scientists was studied as well. The data was basically recorded with a Simrad EY500 scientific split-beam echo sounder equipped with a 120 kHz, 4x10 deg. transducer ES120-4 [1]. Sound pulses were transmitted with duration of 0.1 and 0.3 mS, with a power of 63 watt, and a repetition rate of 5-10 pulses per. sec. Long term monitoring projects was carried out in River Tornionjokki (Fi), River Tana and River Lysaker (No). Controlled experiments was carried out on the ice of lake Semsvann (No), and in a pond in Rimov (Cz). To perform the analysis a software program (Sonar5) was developed by the application of Borland's Delphi programming environment. Methods for visualising the data, for noise reduction and for automatic fish detection were implemented. Tests were performed with simulated data and with data from rivers and lakes. Why it is difficult to count fish in a river a) We can not cover the total cross section area of the river with the sonar beam. This can be solved by the application of area expansion [6]. However, area expansion is complicated because we will have to adjust for changes in the water level, transducer alignment and position. Even the river bottom might change during a monitoring period. The beam has a range dependent height and width, and it is necessary to make assumption about the fish density distribution. b) In the recorded echograms, the noise level will depend on time, range and position. At higher range the beam will approach the bottom and /or the surface, which will increase the amount of reflected energy. Reflections from the surface will change with the weather (waves and rain) while the reflections from the bottom will depend on the transducer position and pan. Changes in the tilt will affect the reflection from both boundaries. c) In a river there are more than fish to be detected. Debris, air bubbles, boat wakes and birds diving for food can produce fish like tracks as well. Multiple fish species and fish schools are also difficulties one has to consider [9]. It may be fruitful to discuss the points in (a), (b) and (c) in terms of a detection probability function (DPF). The DPF is a multidimensional function with variables such as time, range, off-axis position, fish-size, fish distribution etc. The monitoring systems efficiency will vary with the DPF. Hence, lack of control and knowledge of its variations may easily result in meaningless results. Figure 1. Horizontal transducer placement in a river. Results Interpretation In order to solve stated problems, two approaches are possible. i) Gain control of the DPF by monitoring or estimating its variables and ii) reduce variations in the DPF by reducing the influence from the variables. The temporal variations in the covered cross section area can be calculated by monitoring the water level and by mapping the bottom profile. The fish density distribution can be estimated from alternative studies of the fish migration pattern or one can apply a "thin" beam and let it scan the water cross section in steps. Reducing the possibility for fish to slip under the sound beam or behind the transducer can be done by modifying the bottom or by guiding the fish with ropes and nets. In River Tana we mounted a coloured rope along the lower edge of the sonar beam at a place where the river bottom had an unreachable trench. From video recordings we got indications that fish preferred to raise from the bottom when they passed the rope. Due to the nature of the sound beam the systems DPF will be low at short and long range. Hence, limiting the fish to pass through the most effective parts of the beam will reduce the influence from the DPF. The surface is responsible for a major part of the temporal fluctuations in the DPF. From data recorded in River Tana it was clearly seen that the noise level increased during periods of waves or rain. Keeping the beam lower in the water reduced the influence from the surface, but this led to increased disturbance from the bottom. Selecting a site sheltered from wind and waves will help, and it is also possible to smooth the surface with floaters. In periods with high noise level, smaller fish will tend to disappear in the recorded material. In a long monitoring project lasting for years, noise levels should be measured to find an overall size threshold. In periods when the noise level exceeds this threshold, the number of smaller fish should be estimated by time expansion. Another way of reducing the influence of the DPF is by improving the data analysis methods. Data from rivers are often analysed by interactive methods. An operator looks at the echograms and count fish tracks. In echograms with low SNR this method easily becomes biased. If the operator knows that many fish have been caught in a particular period, noise based tracks will easily be accepted as fish. Hence, automation is important. This system must be able to separate noise from solid targets and to recognise fish from the other targets. Data analysis Single echo detection and tracking If the sound pulse hits one fish without disturbance from any other elements, the returned echo will have nearly the same duration as the transmitted pulse, one single peak, and a stable phase. This has led to the development of the single echo detectors (SED) used by most echo sounders. They work well in open water situations. However, in shallow rivers, echoes from fish frequently avoid the single echo description, while echoes from unwanted targets and peaks in the background reverberation level often are accepted. When the ratio between wanted and unwanted echoes (WUR) becomes low, tracking and fish detection becomes difficult. Our first attempt to improve the ratio was to remove most of the traditional criteria except the echo duration. Setting the echo duration criteria much wider than normal increased the detection of passing fish. Naturally this also increased the amount of unwanted detections, but by applying 2 dimensional low-pas filters in the time and range domain the WUR could be substantially improved relative to the WUR found with traditional methods. Image analysis A second solution to the detection problem was found by means of image analysis. We discovered that different echo sources had different frequency spectra in the time and range domain. A low-pass filter with a low cut of frequency in the range domain could be used to detect the instant background reverberation level. A filter with high cut of frequency in the range domain and a medium cut-of frequency in the time domain could improve tracks from fish and suppress noise pulses. Variations in echo intensity and echo phase from passing fish are well known in rivers and leads to gaps in the tracks. Gaps make tracking difficult [3]. LP-filters enhance the tracks by "moving" energy into weaker parts of the tracks, which reduce the gap problem. With filters it was possible both to highlight and to suppress tracks from passing targets. Based on this observation a new target detection method "Adaptive thresholding by cross filtering and subtraction" was developed. (Patented: PTC/NO00/00288). Let H be a matrix containing lowpas filter coefficients and F the original sampled echogram. Then folding F with H will produce the filtered echograms Q1 and Q2. With an echogram recorded in a shallow river with the described equipment, Q3 will according to our experience mostly contain tracks from passing targets. Q1(m1, m2 ) = Eq. 1 Q2 (m1, m2 ) = 3 1 ∑ ∑ F (r,t ) H (m1 − r + 1, m2 − t ) t =−3 r =−1 1 21 ∑ ∑ F (r, t )H (m1 − r + 1, m2 − t ) t =−1 r = −21 F (m1, m2), Q3 (m1, m2 ) = 0, Q1(m1, m2 ) − Q2 (m1, m2 ) > threshold Q1(m1, m2 ) − Q2 (m1, m2 ) ≤ threshold Figure 1 Left: Up-stream migrating fish track detected in River Tana summer 1999. Right: Same part of the echogram processed b the adaptive cross-filter threshold method. Classification The tracks detected by the described methods might origin from different kinds of targets like reminiscent noise, debris, fish, lures, swimming or diving birds or reflections from stones on the bottom. Hence, classification is necessary. To classify one needs track features. An obvious feature is the velocity. Targets moving directly against the current are more likely to be fish than most of the other targets. Unfortunately the uncertainty in the split beam systems angular estimates are too high in a shallow river for this criterion to be sufficient alone. Echoes from reminiscent noise and even from a stone on the bottom can take on an upstream movement comparable with the average speed of a migrating fish. Other criteria such as the number of times a track changes direction, total moved distance, fluctuation in echo intensity, the track's deviation from a straight line or the track curvature can be calculated. Discriminant function analysis (DFA) provides a way to deal with a multivariate, multiple class system. A training set containing typical tracks from the different classes has to be defined. Then the sum of squares within (W) and between (B) the classes are calculated from the track features. In a multivariate system W and B will be matrices. We want a co-ordinate system optimised so that the ratio between B and W (W-1B) are as large as possible. This can be found by calculating the ratio's eigenvalues (λ) and eigenvectors (ε). (W1 B)ε=λε. The elements in the eigenvectors can now be used to orthogonise the room spanned by the features or variables. Hence, calculating the Euclidean distance between a new item and the class centroids can do classification. In a multiple class system it is common to express the distances as a score function for each class so that the most likely class results in the highest score [4],[7] DFA was tested on detected tracks. It was found that it was possible to filter out major parts of the unwanted tracks without "loosing" any of the fish we wanted to detect. SC Fish = −2.68 x14 + 0.46 x24 + 39.72 x28 + 0.66 x39 − 59.86 SC Debri = −3.30 x14 + 0.68 x24 + 14.8 x28 + 1.34 x39 − 84.97 Eq. 2 SC Noise = −2.30 x14 + 0.44 x24 + 1.83x28 − 2.34 x39 − 10.44 Eq. 2 demonstrates the sum-score functions for three target classes. x represent features extracted from a track while the factors have been calculated from the eigenvectors of the training sets. In one test, 800 tracks were detected by a traditional SED-tracking method in River Numedalslågen (Norway). This set contained 8 tracks believed to origin from upstream migrating fish. With DFA, it was possible to reduce the unwanted / wanted track ratio from 800:8 to 24:8. It remains to test DFA on tracks detected by the cross-filtering method. It is reasonable to believe that even better results will be achieved, because the crossfiltering technique allows more features to be calculated and produces tracks with fewer gaps and less surrounding noise. Strange behaviour of the sound-beam During fieldwork in River Tana summer 98 and 99 we observed that a target could be detected in the entire water cross-section and not only within the expected 4-deg. sound cone. This was observed from 6 to 53 m from the transducer and with a water depth stretching from 2.25 to 3.5 m. We have not been able to explain the observations by means of refraction, reflections, wave guides or dipole effects [2]. Summery Figure 3 demonstrates the most important post-processing elements needed to analyse split-beam sonar data from a river. The target detection (a) is based solemnly on the output from the echosounders amplitude detector. This was done because the phase signal was found too unreliable in many rivers. When a target has been located, the angular positions are extracted from the echosounders phase detector (b). A quality factor is calculated from the shape of the echo pulse and from the deviation in the phase samples within each echo pulse. This enables us to apply all echoes in the classification while low quality echoes can be kept out of sensible calculations such as size and velocity estimates. In (c) the targets acoustic size is corrected for offaxis positions and for variation due to the fish aspect [5]. In (d), track features are calculated and the tracks classified before the results are interpreted (e). Echosounder a) target detector b) phase and quality estimation c) size correction d) classification e) interpretation Figure 3. The most important elements in the post-processing chain. In the figure, the echosounder outputs the result from the amplitude detector on the left side and the phase signal on it's right side. Conclusions Split-beam echo sounders can monitor fish in shallow rivers. However, great care has to be taken especially with long term monitoring projects lasting for months and years. In addition to the echo data it is important to monitor parameters such as the water level, bottom profile, sound beam behaviour, and noise level. It will always be possible to set up the system so that the results will have no meaning. Automatic fish detection is possible with the described track detection and classification methods. References [1] Anon, 1996. EY500 Instruction manual. Simrad report no P2473E. Simrad Subsea AS, Strandpromenaden 50, 3191 Horten Norway. 244 p. [2] Balk, H., 2001. Development of hydroacoustic methods for fish detection in shallow water, Faculty of Mathematics and Natural Science, University of Oslo 309 p. [3] Blackman, S. S., 1986. Multiple-Target Tracking with Radar application, Artech House, Inc. 685 Canton St. Norwood, MA 02062. [4] Fisher, R. A., 1936. The use of multiple measurements in taxonomic problems. Ann. Eugenics, 7, 179-188. [5] Kubecka, J., Duncan, A., 1998. Acoustic size vs. relationships for common species of riverine fish. Fish. Res. 35, 115125 [6] Mesiar, C. D., Douglas, M., Eggers, David, M., Gaudet, 1990. Development of techniques for the application of hydroacoustics to counting migratory fish in large rivers. Rapp. P. v. Réun. Cons. int. Explor. Mer. 189, 223-232. Counting results has to be adjusted according to the variations in the detection probability function (DPF) in order to gain a time and range independent index. This includes elements such as time and area expansion, fish distribution estimation, and noise level detection. Ways of reducing and map the influence from the DPF should be investigated further. The method is complicated with all the involved parameters and ways to reduce or automatically detect parameters should be studied. [7] Richard, A. J., Dean, W. W., 1986. Applied multivariate statistical analysis. Prentice-Hall, Englewood Cliffs, NJ, 594 p. [8] Ransom, B. H., Johnston, S. V, Steig, T. W., 1998. Review on monitoring adult salmonid (Oncorhynchus and Salmo spp.) escapement using fixed-location split-beam hydroacoustics. Fish. Res. 35, 33-42. [9] Xie, Y., Cronkite, G., Mulligan, T. J., 1997. A Split Beam Echosounder Perspective on Migratory Salmon in Fraser river: A Progress Report on the Split-Beam Experiment at Mission, BC., in 1995. Pacific Salmon Commission Technical Report No. 8, Pacific Salmon Commission, 600-115 Robson St., Vancouver, BC V6E1B5, Canada. 14 p.