Balancing the costs and benefits of increasing information in ecological models

Storlie, Collin James (2014) Balancing the costs and benefits of increasing information in ecological models. PhD thesis, James Cook University.

PDF (Thesis)
Download (2MB) | Preview


Ecology transitioned from observational studies to experimental studies and hypothesis testing, and is now transitioning back again; this reversal is largely due to technological advances in data collection, storage and computation that have enabled mining disparate sources of data to explore broad ecological relationships and theories. While mining these disparate sources of data has facilitated whole new fields of ecology and better understanding of ecological processes, there is a tendency to assume advanced analytics with complex data yields better results or better understanding; this is not always the case. As models and analysis become more complex so do the underlying assumptions, but the increased complexity may not be necessary. Herein, this thesis explores the value of mining disparate data sources, and of increasing model and data complexity, for exploring species-environment relationships (SERs) in ecology.

The spatially explicit models underlying exploration of the SERs typically rely on linking attributes of a species to coarsely interpolated and temporally aggregated information such as 'climate'. Climate is typically a 30 or 50 year average (as opposed to 'weather', which is more temporally discrete, e.g. daily maximum temperature), and spatially explicit estimates of climate and weather are typically interpolated between known locations based on latitude, longitude and elevation. Such climate and weather estimates represent spatial data which are naïve to the importance of key factors (e.g. topography and vegetation) that structure thermal regimes at fine scales. Further, such climate and weather surfaces may be biased in a non-random fashion as a result of estimating the environment at fine scale without reference to certain biotic and abiotic factors. Hence, gridded climate and weather data are often poor predictors of the true environmental conditions to which species are exposed – but how much does this matter in exploring spatiotemporal patterns in species distribution and abundance, and how these SERs may change?

Species tend to experience the environment at very local scales of time and space, thus a major flaw in spatially explicit ecological studies may be temporally aggregated, inaccurate, or spatially biased environmental data. SERs based on spatial environmental layers with any of the above problems will be biased, potentially leading to false inference of how the species interacts with the abiotic environment, and how this in-turn structures the species distribution. Herein, I focus on improving the accuracy of spatial weather and climate layers and proceed to quantify the value and need to improve these estimates using several algorithms of increasing complexity to estimate SERs. I demonstrate increased concordance of model outcomes with ecological niche theory when using accurate spatial data and increased utility for exploring new relationships.

Underlying the entire thesis is a statistical downscaling of broad-scale weather layers for the Wet Tropics Bioregion of north east Queensland. I statistically downscale 30 years of existing spatial weather estimates against empirical weather data and spatial layers of topography and vegetation to produce highly accurate spatial layers of daily weather. The downscaled weather layers are more accurate with respect to empirically measured temperature, particularly for maximum temperature, when compared to current best-practice weather layers. Current best-practice climate layers are least accurate in heavily forested upland regions, frequently over-predicting empirical mean maximum temperature by as much as 7°C. This thesis examines the value of the extra effort, complexity and assumptions required to produce these data with respect to SERs.

Correlative Species Distribution Models (SDMs) combined with spatial layers of climate and species' localities represent a frequently utilised and rapid method for imputing relationships between a species and its environment, as well as generating spatial estimates of species distributions. However, an SDM is only as accurate as the inputs upon which it is based – garbage in, garbage out. Using current best-practice climate data and my improved climate data, I proceed to demonstrate the effect of inaccurately quantified spatial data on SDM outcomes for a group of seven rainforest skinks. Generally, the distributions of the focal species are not visibly different (at a coarse scale) but the predictions generated using the improved climate layers are more fragmented and contain less core distributional area.

To assess a species' vulnerability to climate change, we commonly use mapped environmental data that are coarsely resolved in time and space. Coarsely-resolved temperature data are typically inaccurate at predicting temperatures in microhabitats used by an organism and may also exhibit spatial bias in topographically complex areas. As a result, simple correlations between where a species occurs and mapped environmental data may predict thermal regimes at a site that exceed species' known thermal limits. In this study, I use statistical downscaling to account for environmental and behavioural factors to develop high-resolution estimates of daily maximum temperatures for the preferred diurnal shelter of a group of rainforest frogs (Family: Microhylidae). I then demonstrate that this statistical downscaling provides temperature estimates that consistently place focal species within their fundamental thermal niche, whereas coarsely resolved layers do not. These results highlight the need for incorporation of fine-scale weather data into species vulnerability analyses, and demonstrate that statistical downscaling approaches are valuable for yielding biologically relevant estimates of thermal regimes.

Methods to predict spatially explicit patterns of species abundance are numerous in form. The most accurate techniques account for variable detection rates, so that we can separate detection from our estimate of abundance. While elegant, these detection models require large presence-absence datasets, derived from repeated surveys across temporal and geographic gradients. In many cases, however, the data are simply not available for these statistical approaches. In these cases, detection-invariant models, which do not require repeated survey effort, represent an alternative. Importantly, if detection rates are unaffected by the predictor variables, then these detection-invariant approaches may yield just as useful a measure of abundance as the more data-intensive models. Thus, by avoiding the use of predictor variables that likely affect detectability, some of the pitfalls of detection-invariant methods can be avoided. To test this, I model the abundance patterns of a group of rainforest skinks using two techniques: occupancy modelling, which accounts for variable detection rate, and a commonly-used presenceonly approach (MaxEnt) which does not. I verify the veracity of model outputs against a large dataset of surveys for skink abundance at 200+ sites over 10 years of time. I find that variable detection models and detection invariant models correlate well with carrying capacity across a number of sites, although variable detection models consistently predict abundance with greater accuracy. This result indicates that detection-invariant models, such as MaxEnt, are not as good as variable detection models but in the absence of repeat survey data, they can come close to the accuracy of a variable detection model. As such, they are still useful for the majority of cases when we require rapid assessment of species abundance patterns in the absence of more robust datasets.

Spatial layers of the weather have applications beyond SDMs and in this section I leverage information from the statistical downscaling of weather maps to demonstrate the effects of vegetation clearance on thermal regimes. The impacts of deforestation are typically measured in terms of habitat: hectares lost, altered habitat fragmentation or connectivity. However, altered habitat extent is just one component of change stemming from vegetation clearance. Climatic conditions too are regulated by vegetation and so are liable to change as well. Vegetation buffers habitats from extreme climate and weather conditions, which are predicted to increase in frequency under global warming scenarios. Despite this, we know surprisingly little about the indirect legacy of deforestation on accelerating the loss of extant climates (and dependent species) projected to 'disappear' under climate change. Here I describe the legacy of deforestation on climatic availability in the Australian Wet Tropics by integrating spatial information on vegetation and weather to quantify 30 years of weather patterns under two alternative scenarios of vegetation extent: prior to European Settlement (ca. 1750) and current (1976-2005). I find that deforestation has on average increased region-wide maximum temperatures by 0.67°C with larger increases in localised areas subjected to more extensive deforestation (0.86-0.90°C). I also show that these modest climate shifts can be underpinned by dramatic reductions in the available area of particular thermal regimes including important cool environments projected to become increasingly scarce under climate change. Moreover, I demonstrate that thermal environments are more fragmented and less connected as a result of deforestation. Finally, I consider the potential for targeted reinstatement of vegetation to reduce range losses and buy time for adaptation to further climate change.

As data sources describing the environment and species localities proliferate, we are left asking what value these data lend to ecological analyses. Observational studies and statistical methods have developed to accommodate ever larger datasets, often assuming that more data will produce better results. The results of this thesis demonstrate that simpler models, with less restrictive datasets and assumptions can utilise large pools of data to form accurate predictions. However, the utility of data sources still needs to be address before they are applied. My research shows that inaccurate or spatially biased environmental data can lead to false inference of SERs, altered patterns of predicted spatial distribution, and a lack of concordance with ecological theory. However, in the process of tailoring these spatial data to suit a variety of ecological analyses, I have further improved our understanding of the interplay between vegetation and the environment. Overall, these results indicate spatially biased climate and weather layers can be corrected with statistical downscaling techniques which explicitly consider abiotic and biotic factors that influence local processes. Downscaled layers meet both statistical (predicting empirical temperatures) and biological (concordance with species thermal limits) criterions of accuracy. Further, downscaling allows for an explicit understanding of how vegetation influences exposure, and the role of forest clearing in shifting thermal regimes.

Item ID: 46519
Item Type: Thesis (PhD)
Keywords: Australia's Wet Tropics; biodiversity conservation; biodiversity; biogeography; Boosted Regression Trees; climate change; climate downscaling; ecological models; exposure; global warming; sensitivity; skinks; spatial climate layers; spatial weather layers; species abundance; species distribution model; vegetation; vulnerability; Wet Tropics
Additional Information:

Publications arising from this thesis are available from the Related URLs field. The publications are:

Chapter 2: Storlie, C.J., Phillips, B.L., VanDerWal, J.J., and Williams, S.E. (2013) Improved spatial estimates of climate predict patchier species distributions. Diversity and Distributions, 19 (9). pp. 1106-1113.

Chapter 3: Storlie, Collin, Merino-Viteri, Andres, Phillips, Ben, VanDerWal, Jeremy, Welbergen, Justin, and Williams, Stephen (2014) Stepping inside the niche: microclimate data are critical for accurate assessment of species' vulnerability to climate change. Biology Letters, 10 (9). pp. 1-4.

Related URLs:
Date Deposited: 30 Nov 2016 02:02
FoR Codes: 06 BIOLOGICAL SCIENCES > 0602 Ecology > 060208 Terrestrial Ecology @ 100%
SEO Codes: 96 ENVIRONMENT > 9608 Flora, Fauna and Biodiversity > 960805 Flora, Fauna and Biodiversity at Regional or Larger Scales @ 33%
96 ENVIRONMENT > 9613 Remnant Vegetation and Protected Conservation Areas > 961308 Remnant Vegetation and Protected Conservation Areas at Regional or Larger Scales @ 34%
96 ENVIRONMENT > 9602 Atmosphere and Weather > 960203 Weather @ 33%
Downloads: Total: 6
Last 12 Months: 4
More Statistics

Actions (Repository Staff Only)

Item Control Page Item Control Page