CHAPTER 4 ADAPTIVE DISTANCE SAMPLING WITH FIXED
4.5 P oint T ransect T heory
4.6.2 Coping with Poor Coverage
So far the chapter has concentrated on increasing the effort in areas of high density to increase the sample size. However, the formulae do not require the effort factor to be greater 1. Thus the approach has a second use as it can also be used to compensate for poor survey coverage. Suppose a conventional line transect survey had been conectly designed so as to ensure good coverage, however due to weather and time, or some other such constraints, surveying was only completed for 70% of the trackline in one area. Then the adaptive methods of this chapter could be used, with the effort factor set to 0.7 for transects in the affected ai’ea and at 1.0 for all other areas.
Although no simulation has been tried for this approach, effort factors of less than one were recorded in the experimental harbour poipoise survey in Chapter 5. These occuned where the boat went off-effort, for lunch or to investigate a species of interest, and returned to the original transect further along than the position it had originally left. This resulted in one or two legs having an effort factor of less than one, and did not present any issues in the analysis.
4.7 Discussion
Overall the simulation results indicate that conditioning on the effort factors only introduces small bias and that the PB method for line transect sampling offers potential for improving density estimator precision for clustered populations. They also indicate a congelation between the degree of clustering and the adaptive efficiency. Adaptive sampling, in its basic foim, will offer no benefit for a population which is not spatially clustered and will in fact be detrimental to
populations with high clustering, a benefit is certainly appaient. An indication of clustering is provided by the relative variance, v{n)lE{n) (Cressie, 1993: p590), with the three simulated population types having mean values of 1 for the CSR, 12 for the clustered and 31 for the highly clustered.
The simulations were sensitive to changes in the adaptive pattern. In paiticulai' if the adaptive track was too lai'ge, so that it frequently stepped outside an animal cluster, this introduced (small) bias into the encounter rate estimate. This is due to violation of assumption (iv), that the expected encounter rate for the adaptive track is the same as the expected encounter rate of the corresponding nominal track. In reality, extra effort is more likely to be triggered when passing neai’ the centre of a cluster, so that adaptive legs may tend to have a slightly lower expected encounter rate than the coiTesponding nominal legs. It should also be noted that the higher the effort factor the more acute the turn on the zigzags, which may introduce both navigation issues and heterogeneity through problems such as double counting.
The ad hoc approach to handling heterogeneity in j{Qi) peifoimed reasonably well. For the bad weather simulation, the mean adaptive density estimator RMSE was lai'ger than the mean conventional RMSE, although this can partly be explained by the smaller increase in adaptive obseiwations compaied to the other surveys. Poor sighting conditions were simulated for 400 units, meaning that the adaptive survey seldom triggered during this time. Thus the majority of the adaptive triggering occuned during the remaining 900 units of nominal effort, causing larger adaptive zigzags, which for much of the time would then step outside clusters.
Investigation is necessary into how to handle a survey involving multiple species, where similar issues arise to the adaptive methods of eaidier chapters. Users could select to trigger on one species only, but then the weights will result in inefficient estimation of density of the other species, unless their areas of high density conespond to those for the trigger species. If the primaiy species and the secondary species are not spatially con elated by habitat, feeding or other factor, it may be acceptable to treat the sightings for the secondary species as conventional sightings and analyse appropriately.
The effort factor calculation does not adjust for changes in expected encounter rate as the survey progresses, so if there was a density gradient in the survey area, there is the potential for the adaptive algorithm to be inefficient. This may adversely affect precision but bias should be unaffected. To minimise this effect, nominal tracldines should run roughly peipendicular to known density contours. If the gradient was extreme and there were few tracklines, such that there was an excess of additional effort remaining at the end of the survey, there is the potential for the adaptive track to step outside clusters and so induce a small amount of bias.
The efficiency is dependent on appropriately selecting the trigger and stopping function; adaptive pattern; and amount of excess effort available. Thus further work is necessary to estimate the degree of clustering for which adaptive sampling is beneficial, and how to tune the adaptive settings to maximise efficiency. In paiticulai', a number of aieas waiTant further investigation.
The trigger function is very simple, with effort increased if the number of observations within a section exceeds some value (zero in the simulations). This does not cater for surveys of multiple species, where different trigger functions may be required. The issue is further complicated by the appropriate behaviour of the trigger function during a period of increased effort. Cunently, primaiily in the interests of acceptable field methods, the effort is not increased further when a detection occurs on a zigzag section. If observations are detected on the last leg of a zigzag, then the effort factor is re-calculated, and a new series of zigzags begins. The survey returns to nominal effort, following an adaptive trigger, after a fixed number of zigzags. There is potential to develop more sophisticated stopping functions.
The expected encounter rate is fixed at the beginning of the survey, which requires that either an initial estimate (or guess) is available or a pilot survey is cairied out. Adjustment of the expected encounter rate using the data that accumulate as the survey progresses may prove useful, particularly when a reliable initial estimate is not available.
The design of the zigzag sections (angle and number of zigzags, and length of section) requires investigation. When each leg in a zigzag is not large relative to the truncation distance W, end and edge effects could be problematic, and field procedures need to be carefully defined to minimize bias (this is discussed further in Chapter 5).
Ideally, simulation will be used to identify suitable parameters, prior to any survey. However as a rule of thumb, the zigzag pattern with 2 or 3 complete cycles perforais well and a suitable trigger value could be obtained either from a short pilot survey or by examining previous survey data.
Overall these methods offer a number of advantages over the Thompson-based adaptive methods. Although not design unbiased, the simulations have shown that the bias intioduced by conditioning on the effort factors is small. The notation and formulae for the methods are less complex than Thompson’s and thus significantly easier to understand and use. This in itself is a significant benefit as Thompson’s methods can be extremely involved and it is easy to make a mistake when calculating estimates.
The zigzag adaptive pattern allows surveying to be continuous for a line transect survey , and so removes the need for off-effort. Depending on the survey this may offer a distinct advantage. If observers can move easily between transects, but the transects themselves are surveyed slowly, for example crawling on hands and knees looking for deer dung, then the off-effort will be a negligible factor and may be seen as a welcome brealc. However if the spacing to adaptive transects is large and the speed of travel is comparable to the survey speed, then resources may not be used optimally and potential surveying time maybe wasted.
The hai'bour porpoise survey in Chapter 5 pointed to poor performance of the group size estimator, although the reasons for this are not fully understood. Ideally the simulations would have estimated group size, instead of modelling observations as single animals, although to provide a useful compaiison the simulations would have also needed to model the bias. Use of the more sophisticated estimator utilising
covariates in the detection function estimate may to go some way to addressing these issues. However tests also need developing to detect heterogeneity in/(0).
Applying the methods to point transect surveys needs further work, and in its current fomi may offer less benefit than for line transects. The approach currently requires the effort factor to be a multiple of the effort to survey a single point, and thus it is difficult to tune parameters to a population’s spatial distribution and the mechanism is likely to be highly sensitive to slight changes in the underlying model. It may be preferable to consider adding adaptive points to groups of points, located within a common area, rather than to individual points. The advantage of this is that effort could be increased in smaller increments, and thus be less susceptible to changes in the spatial clustering. The basic concepts remain the same, and using the terminology of the section, you would still have locations, but now a location would consist of more than one initial point. In making this change the estimators will need to be slightly modified, bringing them closer to the line transect formulae. This revised approach would also open up the options for many other adaptive patterns. So far', the focus of the thesis has been the improvement of estimator precision by increasing the sample size. However the PB method also provides the ability to improve survey coverage through the variable effort factor. It is likely that this will prove the main benefit of the approach. Because the effort factor is a function of whether the survey is ahead of or behind schedule, the method can accommodate some loss of effort due to poor conditions. Effort simply resumes when conditions improve, and the effort factor in adaptive legs is reduced accordingly. Additionally, the methods can also accommodate an effort factor of less than one. Thus areas with incomplete surveying, due say to bad weather, can still be included in the analysis without biasing the abundance estimate.
In the next chapter we report on the application of this adaptive approach to an experimental line transect survey and discuss the field methods.