Dear all
I am faced with a question on selecting species occurrences as input
data for a biomod run. Specifically, I have a number of beaver (/Castor
fiber/) observations which I want to use to model future beaver range
expansion. The data amount to about 1.800 occurrences, representing 72
territories. The number of occurrences per territory varies between 1
and 65 (mean: 25). There are two obvious choices I can make: 1) use all
1800 data, increasing sample sizes -- but running the risk that the
results will (too?) strongly be influenced by the territories with a
large number of occurrences. 2) only use 1 occurrence per territory,
allowing an 'equal weight' for each territory -- but reducing sample
size (i.e. reducing how good territories are 'sampled').
An alternative would be to do multiple model runs whereby I randomly
select 25 occurrences (the mean number) from the territories with > 25
observations while using all available occurrences for the other
territories. Another way would be to weight or scale the occurrence data
-- for example downweighting the influence of occurrences belonging to a
territorial with a large number of occurrences.
While I can implement standalone R scripts to sub select/downweight
data, I am not sure how to feed this into the biomod flow. Any
suggestions on how to tackle this are much appreciated!
Best wishes and thanks in advance,
Diederik
-- Dr.Diederik Strubbe Evolutionary Ecology Group Department of Biology
do you want to use only occurrence-data in the modeling of the potential expansion?
Diederik,
I would consider filtering your records a bit to reduce the potential of oversampling in some areas of your presence points. Unfortunately, presence-only data that comes from multiple sources is usually biased due to oversampling near known organismal hotspots and areas such as biological research stations. Check out Kramer-Schadt et al. (2013; Div. and Dist.) for more information and support on filtering presence records. You can also look into what is known as a bias file, which is really helpful for reducing supposed sampling bias. I have never implemented a bias file in biomod, but I think if you look around you might be able to find a solution.
I would also consider using a replicated-runs approach if you plan to use maxent via biomod. This will allow you to sub-sample different folds of your dataset to derive a more appropriate average distribution, which may be more appropriate based on the amount of data you have. Hope this helps.
Bill
I think you can weight the occurrences thinking on beaver biological traits ¿do they live in groups? or in pairs? ¿or alone? number of occurrences under the preferred form of social group I think lower probabilities of dispersion. ¿which is the minumum territory for a group?, if there are more individuals or more groups in a an area than the possible for a living the probability of dispersal may trigger . May be you can transform you.r data adjusting a functional regression, with a level of dispersion under the limit densities and a dispersive one above the limit. Sorry for my english. Good luck.
Hi Diederik!
I'm facing the same issue. I am trying to find a way to weight occurrence data (or to use a sampling bias file for all algorithms) in ensemble models. I am used to work with Biomod2 but maybe other packages provide a better opportunity to achieve this. Did you, 5 years ago (...), find a solution? Best, Marine
Hi Diederik Strubbe and Marine Le Louarn .
I stumbled across this thread while looking into the question of weighting observations- but I see I'm rather late to join the discussion!
I am using BIOMOD to implement a model including presence and absence locations, and wanted to weight absence locations based on the degree of certainty associated with the recorded absence.
I assigned weights to observations using the Yweights argument of the BIOMOD_Modeling function and found useful information on this here:
https://www.rdocumentation.org/packages/biomod2/versions/3.3-7.1/topics/BIOMOD_Modeling
Hope that helps. Interested to hear how you managed to achieve this.
Best wishes, Hope
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size