I'm trying to predict shorebird distributions using habitat variables, using Maxent to model. I have 4 categorical data (wetlands, waterbodies, tilled, beach) and 1 continuous data layer (topographic roughness).
My categorical data layers are binary (1 = habitat is present, 0 = habitat not present). This is because 3 of these layers were extracted from a larger dataset called SOLRIS and I didn't want everything in the dataset, which had 30 different habitat classes. However after extracting the layers I wanted, I ended up having many zeros in these layers because only few cells were classified as beach habitat, for example. All my layers can be seen below, where D1-D4 consist of the binary categorical layers. Cells that are zero appear to be blank, where there is no colour on the maps. However topographic roughness can be seen covered across the entire extents of the study area.
Additionally, Maxent produces variable percent contributions, that show how topo roughness contributed to majority of the model.
As a result, the model prioritized the topographic roughness layer because each cell in this layer had a value indicating how rough the terrain was, as opposed to the rest of the habitat categorical layers which had a few patches of the defined habitat, and when the habitat was not present, the cell was coded as zero.
How do I make sure Maxent runs on the categorical habitat variables across the entire study area?
May be you
- avoid zero value and take 1 and 2 instead
- reduce the number of categories by lumping (see soil in my majella papers)
Hi,
I would suggest the following:
1) Combine your landcover data. Ideally, you have one categorial data set that has your prefered (=biological meaningful for your targeted species) landcover classes. As I read your data originates from the same landcover, so this should not be a problem. But avoid NoData in your data. If there are a lot of cells that you dont have data for, then the MaxEnt algorythm also cannot calculate occurence probability for those cells.
A solution might be for example to include a additional landcover classes which fills those NoData cells, but be very cautious with that to not alter your landscape.
Also dont forget variables with potential negative effect on species, like maybe human infrastructure.
2) After running some models with a sinle landcover data set, you should get knowledge about which landcover category is prefered and which one is avoided by your species. Then you can go on and included additional "distance to good landcover" and "distance to bad landcover" variables. Those are continuous data spanning your whole study area that represents the positive and negative effects of landcover on its suroundings.
3) I am no expert on shorebirds, but in general I start the modeling process by including topographical variables, landcover variables and climatic variables. You can try to include meaningful climatic variables; MaxEnt will decide itself if those are explaining or not. Dont forget to exclude highly correlated variables though.
4) Dont forget to check "categorial" before running the model. This is trivial, but happens quite often.
Cheers, Florian
Hein Van Gils Thanks for your help. For the first suggestion, I don't think that would work because the reclassified values would still be treated as "NoData". If I classified all zero values as 1 or 2, Maxent would treat this as an insignificant value. Example below shows wetlands across Southern Ontario, where blue raster cells show the presence of a wetland but all that blank space across the province are the NoData cells - and it clearly shows there are more NoData cells than wetland cells. I think my problem is using binary variables for Maxent modelling, as the model will primarily run on the NoData cells, and the model will choose the layer that actually has cell values, such as the topo rough layer.
For the second idea, not sure lumping will work for my case. Unless you mean combining each habitat variable into one layer, but then the NoData cells will still be present.
Thanks for your suggestions!
Hi Rebekah,
I suggest to use proportional habitat variables to allow fitting proper response curves rather than categorical data. E.g. from here: http://www.earthenv.org/landcover which has 12 classes including 'regularly flooded', 'water', 'barren', 'herbacious', and so on.
Otherwise, I would agree with Florian and suggest to include all levels of your categorical variable (habitat type) in a single layer e.g. with integers coding for the different types. You should avoid NA as no predictions can be made to NA.
You most likely need more variables, in particular, you are entirely lacking information on climate, see e.g. chelsa-climate.org, or human impacts (as mentioned by Florian!), e.g. https://sedac.ciesin.columbia.edu/data/set/wildareas-v2-human-influence-index-geographic
Also you should try several topographic variables, e.g. VRM (google sappington) & slope, but not aspect as it is circular.
cheers, rainer
- Badges
- Science topic
- Similar topics
- Filing