Tuuli Rissanen has done lots of research on these with tundra plant species

Related to Community Ecology

• We need species occurence and environmental data
• Machine Learning is sometimes used

Occurence data

• GBIF - presence only
• Forest Understory (LUKE) for Finland gives presence-absence for plants

Environmental Data

• Chelsa WorldClim
• CORINE - Europe land cover
• SRTM - topography

Creates a Niche

Hutchinson's n-dimensional hypervolume

Evaluate model

• Split the data into training and test data - train with the test data

Macroecology

Notes from lecture on [Climate Change]] ([Lecture_climate_change.pdf)

Used to:

• Define specis niche
• Define potential distribution
• Identify potential for invation of a spcis
• estimate population size

Biotic-Abiotic Movement (BAM) diagram

!Lecture_climate_change, p.22

• Movement is usually more theoretical since data is often lacking

Modelling steps

• Running an SDM to create a map is usually fairly easy. The hard part is deciding on your assumptions, the data being used, covariates to consider.
• It can be thought of like a cycle since each model will inform the next one, you need to keep iterating on it
  • Conceptualization: Model objective, taxon, location predictors, scale, conceptual underpinning, software availability
  • Gather data
  • Model fitting (choose variable, model settings, etimates, independence analysis)
  • Assessment (preformance statistics)
  • Predictions (maps and uncertainty)
  • Repeat!

Data

• Can come from field data, biodiversity databasees like GBIF, museum records, remote sensing...
• Remove outliers, define spatial extent, select variables, pseudo-absense selection if your data doesn't have absense data.
• Mode with climate data to find the niche (usually tmperature/ppt), create response curve, make projection

Correlative SDMs

• The most common type of SDM. Two main methods:

Presence-absence method:

Logistic regression, Generalized linear models, Generalized additive models, random forest, boosted regression trees

• This method allows us to distringuish between occupied and non-occupied habitats, so we can determine the probability of finding the species at each place
• However, absense data is harder to come by and requires systematic surveys

Presence-background method

Ecological Niche Factor Analysis (ENFA)
Genetic Algorithm for Rule-set Production (GARP)
Maximum Entropy (Maxent)

• WE can distinguish between suitable and unsuitable habitats, but not probability of a species existing here, even with "pseudo-absense" data

Occupancy Models

• Incorporates the detection probability of a species when estimating the probability of occupancy in a site - requires repeated presence-absense data in each site
• Include covariates for prbability of detection like weather
• Bernoulli observation model is used with logit Link function between probability of detection and measurements $\times$ covariates

Hutchinson's n-dimensional hypervolume

See https://damariszurell.github.io/SDM-Intro/ - using mutliple variables to create a n-dimensional response curve

Model evaluation

Cross-evaluation

• Like we did with our Paleo Transfer Function - Model Evaluation, we can use bootstrapping or Leave-One-Out (or other methods) to cross validate our availble data with the training set, since we are usually limited in data
• If we have enough data, external evaluation from independent data is better of course

Threshold-based

!Lecture_climate_change, p.31
How good is the model at true/false positives (TP/FP) and true/false negatives (TN/FN).

TSS = $\frac{TP}{TP+FP}-\frac{TN}{TN+FN}-1$

Usually used with k-fold validation - which method maximises TSS? This is probably the best one

Area under ROC curve:

This can help us determine how good our model is at predicting
!Lecture_climate_change, p.32

Sources of uncertainty in SDM modeling

• Environmental uncertainty from data source
• Model uncertainty (models algorithm, parametrization)
• Future prediction:
  • Projection Scenarios
  • Different global circulation modeling