Title
Condition Metrics for Elm Forest and Saxaul Ecosystems in the Gobi Desert. Report for Oyu Tolgoi and Arthur Rylah Institute for Environmental Research Technical Report Series No. 310
Author(s)
S.J. Sinclair, O. Avirmed, K. Batpurev, P.A. Griffioen, M.D. White and Kirk Olson
Published
2020
Abstract
Context: Ecological change in the Gobi Desert of Mongolia must be monitored to understand the impacts of changing land use, including the Oyu Tolgoi (OT) mining project. To address this need, metrics were created in 2018 which quantify the condition of five rangeland ecosystems (Avirmed et al. 2018). That work was undertaken as a collaboration between Wildlife Conservation Society (WCS Mongolia and the Arthur Rylah Institute for Environmental Research (ARI). That work, like the current project, was funded by OT. The 2018 metrics were created by collecting stakeholder opinions about the condition of a wide range of computer-generated sites, and then using these opinions to create models of stakeholder condition score. These models can be applied as metrics. They take field-measured data from a site of interest and return a condition score. Given the stakeholder-driven approach, the metric score explicitly reflects stakeholders’ views. While Avirmed et al. (2018) created metrics for five ecosystems, the original project scope only allowed three priority ecosystems to be tested using field data. The metrics for the other two ecosystems (Elm Forest and Saxaul) remained draft metrics. Aims: We aimed to test and refine the condition metrics for Elm Forest and Saxaul, with the aim of producing final metrics fit for field monitoring. Methods: We tested the existing Elm Forest and Saxaul metrics using the same tests used by Avirmed et al. (2018), and diagnosed the issues we found. We implemented three strategies to improve the metrics: • Newly available field-derived evaluation data were added to the training set, • The computer-generated training sites were culled to remove sites with ‘unrealistic’ attributes. This reduction was guided by stakeholder opinions about the ‘plausibility’ (likelihood of encountering) each site. • The training data were stratified and weighted, to ensure that all models encountered data from a range of sites, with a focus on sites with realistic attributes. We tested the new metrics using a cross validation approach. Some field data were used in the above remodelling strategy, while the remaining sites were withheld for testing. We repeated this ten times, with different sub-samples of test data. We appraised model improvement by the difference in r2 between the old and new metrics. Results: When tested against stakeholder field evaluations, the Elm Forest and Saxaul metrics created in 2018 were positively related to stakeholder evaluations, and were thus capable of providing condition scores consistent with stakeholder expectation. The Elm Forest and Saxaul metrics did not perform as well as the metrics previously created for the three desert ecosystems. We suggest this was caused by- 1) a lack of consensus among stakeholders for these ecosystems, making it more difficult to model the consensus. We suggest that it is inherently more difficult for stakeholders to conceptualise condition change in these systems, and 2) bias in the original training data. The computer-generated sites were skewed towards sites with unrealistically high cover and richness values. Real field sites resemble only a small range of the data used in training, at the lower end of the score range. This presumably resulted in the models returning low scores and having low resolution when dealing with the attributes of real sites. The strategy we applied to improve the metrics clearly improved the Elm Forest metric but resulted in a Saxaul metric which was not demonstrably better, and was likely overfit to the field training data. Conclusions and implications: We conclude that the final metrics presented here are fit for monitoring, for both Elm Forest and Saxaul. The Saxaul model is likely to be over-fit to the training data (meaning that the model’s ability to extrapolate beyond the training data is lower than expected) but we argue that this is not a major cause for concern.
Full Citation
Sinclair, S.J., O. Avirmed, K. Batpurev, P.A. Griffioen, K. Olson, and M.D. White (2020). Condition Metrics for Elm Forest and Saxaul Ecosystems in the Gobi Desert. Report for Oyu Tolgoi and Arthur Rylah Institute for Environmental Research Technical Report Series No. 310. Heidelberg, Victoria, Australia: The State of Victoria Department of Environment, Land, Water and Planning.

Access Full Text


Back

DMX4429900000