Hi Everyone,
First of all I wanna acknowledge their responses to my previous questions. they have been so useful.
I have my models in MaxEnt and then probably I'll run the same analysis in BIOMOD2. NOw I would like compare the performance of two model or if there are some difference in my results if I compare data in temporal scale. For example, I have a data from 1990-99 and I'd like compare differences in the model with data set from 2000-2010. Someone knows how could I do it?
Any comment will be appreciated!
Cheers
First, I would recommend standardized your model's output before making any comparison. ENMTools provides this function for free and below is the related literature. Second, if you are comparing the model's performance, then it will involve some accuracy assessment. There are many ways to do it. The second literature may served as a starting point for the question.
Warren, D. L., Glor, R. E., & Turelli, M. (2010). ENMTools: a toolbox for comparative studies of environmental niche models. Ecography, 1(September 2009), 607–611. doi:10.1111/j.1600-0587.2009.06142.x
Lobo, J. M., Jiménez-Valverde, A., & Real, R. (2008). AUC: a misleading measure of the performance of predictive distribution models. Global Ecology and Biogeography, 17(2), 145–151. doi:10.1111/j.1466-8238.2007.00358.x
I am not sure if I understood your question but maybe this paper can help you.
Article A comparison of Maxlike and Maxent for modeling species distributions
Beyond the many statistical comparisons possible (see above and many more), the number of environmental variables used by each or all SDM-algorithms to achieve a specific statistical performance is relevant. Standardization of the number of variables would seem prudent as Mang seems to suggest. Also the category (e.g. climate) of variables impacts often on how many variables you need to achieve an acceptable (statistical) performance. In my experience it generally takes more climate variables than other variables at intermediate spatial resolution to achieve the same statistical performance. The attached also provides an example of temporal resolution. Finally, it could be that not all SDM-algorithms have the same optimal number of occurrence records. If you have large numbers of records you may consider to test thinning.
Article Where the bears roam in Majella National Park, Italy
Article Fine resolution distribution modelling of endemics in Majell...
To just look at differences between the two outputs themselves, you could use the raster calculator (if you are using ArcGIS) and subtract one rasterised output from the other to show the areas of difference.
Thank you very much everyone!
It has been very useful. I agree with many of your comments, mainly standardize the results and use some tools. Actually I'm starting to use ENMTools to test my results.
Hi,
I've used the map comparison kit (http://mck.riks.nl/), namely the fuzzy kappa in this paper to compare map projections of two models (https://www.researchgate.net/publication/268443984_Combining_distribution_modelling_and_non-invasive_genetics_to_improve_range_shift_forecasting).
Cheers,
FMestre
Article Combining distribution modelling and non-invasive genetics t...
Thanks Frederike, Jin Li and everyone! I'm doing comparison between niche now, I'm starting with a simple Pearson' correlation and I'm trying with ENMtools (Warren et al 2008). Recently Warren recommend me to use ecospat for niche overlap and other utilities (based on Broennimann et al 2012 - DOI: 10.1111/j.1466-8238.2011.00698.x). I'll being review during these weeks. Cheers again
@Ricardo Alvarez - What is exactly that you are using the Pearson correlation for? I am also looking for ways to compare Maxent models.
Some additional approaches are published in the attached.
Article Transferability of species distribution models: The case of ...
Hello Ricardo,
You can calculate dissimilarities between models using distance metrics. You can read more about it in this paper. Good luck,
Manuel
HI George,
I'm using pearson to measure the correlation of the maxent niche output. I used ENMTools to run the correlation. This tool allows measure the correlation of variables to estimate the autocorrelation (environmental, geographic variables) and also it is posible run the asc files obtained in maxent.
Cheers
You can use several diagnostic tests to compare the goodness of fit of Maxent models. I will list some of these and the software tools that can help you calculate these diagnostics:
1- AIC (Akaike Information Criterion) this is a statistic that favors more pasiomoneous models based on the likelihood of the model and being penalized for higher number of parameters. In other words, the higher the number of parameters the higher the AIC value. The best performing model is the one with the lowest AIC score. AIC difference of less than 2 units is insignificant. AIC = -log(liklihood) +2K, K is the number of parameters. You can use ENMTools by Dan Warren to calculate AIC for different Maxent models.
2- TSS ( Total Skill Statistic) this statistic is based on the specifity and the sensetivity of the model. A TSS of 0.7 or greater is a good performing model. you can calculate this statistic using R package. There is a R script posted on google Maxent group by Alaa Soultan calculates the TSS out of your Maxent outputs. But you need to run your maxent models with the option to calculate predicted background points checked on.
- Badges
- Science topic