How does map error impact priorities

How do we effectively protect biodiversity if the maps we are using are inaccurate?

Vivitskaia Tulloch completed this project as part of her Honours thesis at UQ. She is currently a PhD student at the Centre for Biodiversity and Conservation Science (CBCS), jointly funded by the Wildlife Conservation Society and CSIRO. Her collaborative project with WCS involves modelling catchment run-off to coastal waters in Papua New Guinea to investigate the impacts of oil palm on coral reef habitats, and to optimally plan for conservation and sustainable resource development in the region.

Vivitskaia Tulloch completed this project as part of her Honours thesis at UQ. She is currently a PhD student at the Centre for Biodiversity and Conservation Science (CBCS), jointly funded by the Wildlife Conservation Society and CSIRO. Her collaborative project with WCS involves modelling catchment run-off to coastal waters in Papua New Guinea to investigate the impacts of oil palm on coral reef habitats, and to optimally plan for conservation and sustainable resource development in the region.

Reserves are one of the most successful systematic conservation planning tools used to protect biodiversity. But choosing the best location for a reserve is never easy. Reserve placement relies on ecological data, such as the distribution of species or habitats. But field data is expensive and difficult to obtain, so cheaper data options such as habitat maps are often used instead. For planning involving the conservation of coral reefs, remotely-sensed habitat maps are one of the most commonly used data sets, usually because of the remoteness or extent of these marine systems. Using these types of habitat maps, however, can cause big problems. Most people assume habitat data is accurate – but it never is! There are all types of errors associated with the making and using of habitat maps, and almost never is this error accounted for in conservation planning.
Why does this matter? Well, ignoring data uncertainty can cause all sorts of problems when this data is used to plan marine reserves. We might end up placing the reserve in the wrong location if we assume the data is right, when it is actually wrong. Planners however are forced to use inaccurate data regularly, which could be leading to misinformed management decisions. We might not be protecting the right areas, and it also might lead to us to spending our already stretched conservation budgets unwisely.
Making decisions with what we have
To avoid issues of error in our data, some suggest we should collect MORE data (through field surveys for example). However, besides being very costly, such efforts also take time meaning delays in protecting biodiversity that might be under considerable threat. Many coral reef regions, for example, are already facing significant losses from activities connected to human activities. We simply can’t delay our conservation effort. Consequently, we need ways to include data uncertainty in our planning processes now.
And that’s where my work comes in. Working with colleagues at CEED at UQ, I have recently developed a new method of accounting for data errors in conservation planning, using a modified version of the decision-support tool Marxan. In a nutshell, we are seeking to minimise the risk of using inaccurate data in order to be confident in conservation decisions.
To do this, we designed a series of marine reserve networks that included or ignored habitat-mapping error. We used a case study of the Kubulau traditional fishing grounds, or ‘qoliqoli’, in Fiji. The Kubulau qoliqoli is located in the South-West of the island Vanua Levu, and covers an area of around 260 square kilometres. The researchers worked closely with the Wildlife Conservation Society ( WCS) in Fiji, a non-government organization working with communities in Kubulau to initiate marine management projects.
WCS has been working with local stakeholders in the region to build a resilient MPA network design that protects 30% of all coral reef habitat types. The collaboration between UQ and WCS has enabled researchers to obtain detailed knowledge of the cultural and socio-economic needs of local stakeholders in the Kubulau qoliqoli, which were integrated into the new reserve design methods.

How good’s the map?

Figure 1: Accuracy values for a map of marine ecosytems derived from remote-sensing imagery of the Kubulau Fiji fisheries  management area (qoliqoli).

Figure 1: Accuracy values for a map of marine ecosytems derived from remote-sensing imagery of the Kubulau Fiji fisheries management area (qoliqoli).

A number of coral reef habitat maps at different scales have been developed for the region. These have been derived from high spatial resolution satellite imagery using hierarchical object-based image analysis and segmentation. For the spatial prioritization they used a fine-scale benthic habitat map describing coral, reef, algal, seagrass and sediment habitats.
So, a habitat map exists for this area. But planners also needed information on the accuracy of this map to include in the spatial prioritization. Errors in map classification are common, due to detection issues and spectral mixing. These values were obtained using an error matrix, also called a confusion matrix, which basically tells you how many pixels in the map were classified correctly by comparing the map to reference images, and how many pixels were inaccurately assigned to the wrong habitat class.

Using this error matrix, the accuracy of each habitat classification was determined (Figure 1), which was just the number of times a pixel was correctly classified divided by the sum of all reference images for that class. For instance, a habitat type called ‘coral’ was correctly classified 34 times. This number was then divided by the total number of reference images (66 for coral), giving an accuracy of 51.5% for that habitat type.
These accuracy values were included in a new approach to spatial prioritization using Marxan with Probability (or MarProb for short). The traditional version of Marxan aims to achieve biodiversity objectives for the lowest cost. This new modified version of Marxan, MarProb, solves the same problem but is also able to maximise the probability of protecting every conservation feature, given inaccurate data. It does this by applying a penalty to those habitats that do not meet a target certainty, and so is less likely to include them in the final solution as doing so will increase the objective function score.

Mapping the options

To investigate how priority reserve areas change once habitat-mapping error was accounted for, we ran a number of different reserve design scenarios using the traditional version of Marxan, where we assume habitat occurrence is 100% certain. We compared these reserve designs with those using the probabilistic method, MarProb, which included the habitat-mapping accuracy data, to determine what if any changes occurred by using probability data. By varying the conservation representation targets from 10% to 100%, we also aimed to evaluate trade-offs in marine reserve design between representation and cost.
When we compared the results from the standard and probabilistic scenarios, we found that many areas that were a high priority for meeting targets in one scenario could be a low priority in another scenario, and vice versa. Areas important to achieving representation targets for habitats in probabilistic scenarios were largely excluded or considered unimportant in the standard scenarios.
We also found reserve networks that accounted for mapping error were up to 50% larger than those designed using the standard approach. We wanted to look into what was driving these differences, and found that larger amounts of habitats with low mapping accuracy were represented in reserve networks created through the probabilistic approach. When we locked in the priority planning units from the standard approach into a run using the probabilistic method, most habitats failed to achieve their representation targets, and habitats with accuracy values less than 90% would actually never achieve their targets.
Importantly, existing reserves do not meet national conservation targets (30%) once mapping error was accounted for. Six habitats were under­represented including coral and seagrass, which are really important habitats supporting large amounts of biodiversity.

The consequences of a poor map

This study is important as it shows that conservation schemes based solely on the distribution of habitats that don’t have information on mapping accuracy may fail to achieve adequate conservation outcomes. This is because they risk either under-representing conservation features by missing out on protecting high priority areas, or over-representing conservation features by including areas in the reserve network that are of low importance to meeting conservation goals.
Given that we have so little money to invest in conservation actions as it is, we could really be reducing the efficiency of conservation investments if we don’t start accounting for uncertainty. The results from this study show that serious consequences might result by not accounting for mapping error in marine spatial planning. That’s in terms of the benefits for conservation and the cost of management.
Even though larger reserve networks that account for mapping accuracy could be more costly to manage, they are more robust to uncertainty than those that do not consider mapping accuracy. Even though they were larger, the key advantage of the probabilistic solutions with high certainty targets is that they increase confidence in achievement of conservation outcomes, making decisions more robust and less risky.
There is a growing call for decision-support tools to explicitly incorporate uncertainty associated with the data in conservation planning. This new probabilistic approach allows planners to quantify how often targets might be missed when uncertainties are not considered, enabling them to evaluate necessary trade-offs, and understand the implications of not including uncertainty information in the planning process.
If uncertainty data are not readily available, planners have several options to ensure they are not ignoring the error inherent in their habitat maps. First, planners could try to source more information about the habitat information that is often just assumed to be correct. For instance, how were the maps produced? Was there any validation? Which habitats may have been problematic and/or under-sampled?

“Serious consequences might result by not accounting for mapping error in marine spatial planning. That’s in terms of the benefits for conservation and the cost of management.”

Second, planners could set different targets for each habitat, with higher targets set for habitats suspected to be uncertain. For example, some habitats are known to be commonly confused when mapped (eg, because their spectral signatures or textural characteristics are very close or because they tend to occur in the same geomorphic zone in patchy habitats), so planners could set higher representation targets and/or higher certainty targets for these habitats.

Embracing uncertainty in conservation planning and reserve design is important in our search for more robust and defensible conservation decisions – so let’s start doing it! Our challenge now is to ensure data accuracy assessments or uncertainty information is more readily available, and find new methods of dealing with these uncertainties to allow the design of reserve networks that adequately and efficiently represent biodiversity.
Importantly, this research provides an avenue for applied conservation management, as it was done in collaboration with WCS-Fiji, an influential organisation in making conservation decisions in this region. The new methodology provides a potentially useful and cost-effective reserve planning decision-making tool for scale-up to a national level in Fiji. The collaboration with WCS, and in particular WCS-Fiji Director Stacy Jupiter, is ongoing, with current research investigating trade-offs between socio-economic costs and data accuracy in coral reef marine spatial planning.


More info: Vivitskaia Tulloch v.tulloch@uq.edu.au
Reference: Tulloch VJ, HP Possingham, SD Jupiter, C Roelfsema, AIT Tulloch & CJ
Klein (2013). Incorporating uncertainty associated with habitat
data in marine reserve design. Biological Conservation 162: 41-51.

 

Leave a Reply