Adaptive track designs

So far we have considered only the pre-planned survey in which the cruise track is decided on the basis of prior decisions. All adaptive survey strategies require some knowledge of fish distribution. The decisions required for the adaptive approach cannot be made in the absence of a knowledge of the spatial distribution.

There are circumstances in which it might be desirable to adjust the cruise track as the work proceeds (e.g. if it is important to locate commercially exploitable fish concentrations. It may be decided that areas of high density should be surveyed more intensively than elsewhere. However, the adaptive survey is not necessarily appropriate when the principle objective is to determine the stock abundance. The problem is that the acoustic data may not be considered as entirely random samples because the position at which each measurement is made has been determined to some extent by the earlier observations, and the abundance estimate may be biased. In addition, it will almost certainly be very difficult to estimate the confidence limits from such strategies. It is not easy to allow for the bias in the analysis, and whether this can be done at all depends upon assumptions about the fish distribution which may be difficult to validate. It is very important to obey strict procedures in the execution of adaptive surveys and to use only appropriate analysis methods for estimation of biomass.

We shall consider three kinds of adaptive survey. One technique is to begin with an outline survey which is a rapid investigation of a large area using widely-spaced transects. This is followed by more intensive examination of particular regions where fish concentrations have been detected. Another approach is to fix the transect spacing in advance, but to allow the length of the legs to be changed during the survey. Thirdly, the transect lengths may be decided beforehand while the spacing is varied according to the observations made at the time.

There are two important considerations to be borne in mind when making the real-time decisions required during an adaptive survey. Firstly, when the sampling intensity is reduced, the coverage must still be sufficient to provide good enough information to decide the subsequent sampling strategy. Secondly, when the plan is to return to regions where large fish concentrations have been observed, success depends upon the assumption that the fish distribution does not change with time, so that the concentrations can be relocated without difficulty.

3.2.2.1 Outline survey

The survey is conducted in two stages. First, the vessel covers the area of interest on a widely spaced grid, to detect regions of high fish density. This stage should occupy no more than say 25% of the time available.

The vessel then returns to the regions where fish have been observed, and the remainder of the time is spent in surveying these regions more intensively (Figure 7). This technique is not useful if the fish are likely to migrate or disperse in the time between the initial sweep of the area and the return visit. Furthermore, if the initial sweep is too widely spaced, some localized concentrations may not be detected at all. The outline survey works best when the area to be examined is not too large (e.g. within a fjord), and the fish are believed to be concentrated in a few large and static schools. The same technique has also been employed in tropical areas by Str0mme and Sretersdal (1990). All fish stocks show some evidence of temporal change and the use of an outline survey with mapping methods such as geostatistics can cause problems if both outline and high density surveys are combined. The outline survey is best used for abundance measurements when a stock occupies a small proportion of the possible area.

3.2.2.2 Variable transect length

This technique may be applied when the spatial distribution is well defined in one direction. For example, suppose there is a coastline along one edge of the area to be surveyed, and the stock is located mainly in the shallow water near the coast. The survey is designed initially as a grid of transects running between turning points on the inshore and off-shore boundaries. During each run in the off-shore direction, it may be decided to terminate the transect once the observed fish density has declined to a small proportion of that observed near the coast (Figure 8). The acoustic data may be analysed in the normal way, by calculating the abundance in elements of area, on the assumption that negligible quantities of fish would have been observed along the abandoned parts of the cruise track. To facilitate the analysis, once the decision to turn has been taken, the transect should nevertheless be continued to the edge of the current area element that will be used for data analysis such as a rectangle or depth stratum.

3.2.2.3 Variable transect spacing

Suppose that the fish are expected to occur in local aggregations, but in regions which are unknown in advance (e.g. clusters of migrating schools). The general plan is to increase the sampling of any region where the observed fish density is much higher than the average, by reducing the transect spacing (Figure 9). The transects continue to run for the full length to avoid gaps in the coverage.

The transect spacing should be decided on the basis of objective criteria. The variance of the density measurements might be used to determine the spacing, as proposed by Stolyrenko (1988), on the grounds that precision is improved by sampling more intensively in regions of high variance. A simpler alternative technique

is to observe the mean fish density along each and to make the to the next transect nrt,nfWI'I\n

to l/(mean density) subject to the calculated spacing being contained within practical limits. When few fish are observed, the variance is also small and the effect is to increase the coverage of the main concentrations. The two methods may not be much different in practice. Jolly and Hampton (1990) found little difference in the two methods when analyzing data from a complete survey to give predetermined strata. Aglen (1989) analyzed data from a number of surveys using a coefficient of variation and obtained consistent results that suggest that the standard deviation of the population is proportional to the mean of the fish density.

Another approach is to design the survey grid in advance with a fixed transect spacing, but ensuring that the pre-planned track does not require all the time available for the survey. The spare time is used to cover extra transects in regions of high density. Whenever the observed density exceeds some limit, one extra transect is traversed half way between those of the pre-planned grid, so that the sampling intensity in that region is doubled. These techniques may cause bias. As the sampling strategy changes, transects are no longer placed evenly throughout the area but are concentrated towards the centre of the aggregation giving the possibility of bias unless the area allocated to each transect is chosen carefully. If this technique is used along with conventional processing methods such as those described in sections 4.2.2-6, bias may occur. The bias in the estimate of the mean density may be removed by randomizing the transect spacings.

We begin with a pre-planned set of parallel transects placed systematically or randomly in one or more strata leaving some extra time for transects to be inserted during the survey. When it is decided to increase sampling intensity, a new transect is included in the current strata and those ahead. This transect must be located randomly because all previously selected locations will have already been occupied by transects; thus no 'systematic' location for an extra transect exists. At this point in the survey, a trend has been found thus requiring increased coverage and indicating that the random relationship between systematic transect position and stock no longer holds and care must be taken to ensure that additional transects do not cause bias in the results. This procedure may result in the first additional transect being out of the normal sequence. On average, half the time, the track progression will reverse to run the extra transect. However, all the transects in each subsequent zone may be taken in the normal progressive sequence. When it is decided to reduce the sampling intensity, all the transects placed in the current zone must be completed first. Alternatively, the transects within each zone may be taken in random order - in which case the sampling intensity may be reduced at any time. Both these procedures are free of bias provided that strata with different sampling intensities are treated as separate regions in the analysis. Figure 10 illustrates the randomized adaptive method. The survey initially progresses from east to west. When the fish concentration is observed on transect 3, the sampling intensity is increased. By chance, transect 4 is back to the east, but the later transects are taken in the normal sequence. The method may appear complicated at first sight, but it is simple to implement in practice and provides an estimate of the abundance with minimal bias and without the need for doubtful assumptions about the spatial distribution of the stock.

The adaptive methods discussed above all suffer from the problem of possible bias in the estimates of abundance. If the assumptions required by the adaptive process are incorrect, bias will occur. Both the pilot survey and the adaptive transect lengths may miss some parts of the stock. The increased sampling may lead to more precise estimates of the parts of the stock that have been detected, while underestimating the total. In all cases, because the transects are located with reference to the observed fish densities, estimates of confidence limits may be difficult to calculate.

3.2.2.4 Summary

Adaptive techniques should not be used where the stock is mobile and can move significantly during the survey. Adaptive track designs are only recommended for surveys of highly contagious stocks (Table 2). By choosing an adaptive strategy, it must be accepted that estimating the precision of the survey cannot be attempted without further important assumptions about the stock distribution. If, following the data collection, it is uncertain whether the criteria required for the adaptive design have been met, such as the absence of

migration, great care must be taken in the analysis. Where double coverage has been used, it may be possible to use the data as two separate estimates.