Reliability of species distribution modelling for wildlife conservation in developing countries

Species distribution models have become essential tools in ecology and wildlife conservation. However, their reliability when used for conservation management is often compromised by many challenges and limitations, as for example the lack of sufficient data-quality and sampling bias. Especially when used for areas in developing countries, where unbiased good-quality data are scarce, the robustness of species distribution models is questionable. In this thesis, I studied some crucial issues affecting the reliability of presence-only species distribution models for wildlife conservation in developing countries.

In the first chapter, I studied the issue of sampling bias in data-poor situations of developing countries and how to correct for it in presence-only species distribution models. I implemented model-based bias correction by incorporating additional bias-predictors describing site accessibility (distance to closest city, road, and protected area) or estimated sampling effort. I showed that bias correction led to improved predictions, with comparable results using the three modelling algorithms (GLMs with subset selection, GLMs fitted with an elastic-net penalty and Maxent, all under the point process modelling framework). The improved prediction due to sampling bias correction was dependent on how well the bias-predictors describe the bias in the available data, with higher improvement when accessibility bias variables were used. However, objectively evaluating sampling bias correction requires bias-free presence-absence testing data, which is typically not available in data-poor situations. Nevertheless, my results showed that model-based bias correction is a useful tool to improve predictions in data-poor situations, in which other bias correction methods may not be applicable.

In the second chapter, I evaluated the adequacy of presence-only data from within one developing country’s boundary to calibrate national species distribution models. I used spatially explicit information representing predictions from the species’ global environmental niche (potential distribution) as an additional predictor (prior) in regional models aiming to improve predictions. The use of priors did not lead to improved regional predictions; meanwhile, the correction for sampling bias led to improved predictions whichever prior was used, making the use of priors less pronounced. Under biased and incomplete sampling, the use of global data did not improve regional model performance. However, the actual improvement could not be quantified without enough bias-free regional data. Nevertheless, predictions from global models, interpolated to regional scales, can still have great potential to guide future surveys and improve regional sampling in data-poor regions.

In the third chapter, I assessed the sensitivity (robustness) of a spatial conservation prioritisation application in an exemplary data-poor developing country to common sources of uncertainty, which are related to the quality of available species distributional data and the choice of some of the user-defined software parameters. I also evaluated the effectiveness of the Egyptian protected areas network for conservation. Conservation planning in data-poor situations was found to be sensitive to the selection of the surrogate group, correction for sampling bias, connectivity parameters, and the choice of modelling algorithm; collectively, these reflect data quality issues. Results showed a lower limit for data quality for the usefulness of the spatial conservation prioritisation approach, demanding improved data quality for data-poor countries. Using currently available data on the Egyptian butterflies, reptiles, and mammals, the Egyptian protected areas network was found inefficient for wildlife conservation. I determined the top priority sites for further on-the-ground field evaluation as potential areas for protected areas expansion.

Despite the promising results of improved predictions after correcting for sampling bias in data-poor developing countries, this improvement is not guaranteed and hence should not be considered a replacement for the urgent need for improving sampling strategies for the collection of biodiversity data on as many taxonomic groups as possible. Improving sampling strategies and data-quality from data-poor countries (mainly from the less visited areas) will consolidate the use of species distribution models for conservation planning in these areas.