MaxGeeking Out On MaxEnt/Presence Only Species Modeling

[Sunny_Esri]

Extra! Extra! Read all about it! 

Presence-Only Predictive Modeling (aka MaxEnt) is now available in ArcGIS 2.9 and I am SO EXCITED. Why is this a big deal? Well, for one, it's just super fun... 

Ok but really - it's exciting because there are many phenomenon that occur and we know where they've occurred but not necessarily where they HAVEN'T occurred. Like Bigfoot, for example. (Just kidding... or am I?) 

For wildlife biologists (including those that study cryptozoology), botanists or other -ologists looking to understanding where something COULD be... or more specifically, where the conditions might occur to support the presence of said phenomenon... but they only know where it already exists, MaxEnt modeling is extremely helpful. 

How does it work? 

I am no mathematician, but if you want to know the nitty gritty details you can read the documentation. Or you can CC the @GeoSpatialAnalysis team in your response. Ha! 

But here's my version of how it works: 

Let's say I have a rare plant species and I have just a handful of known populations. I've been tasked with strategizing my field work during its flowering season and I want to maximize my time by targeting searches in locations where the species COULD occur. Maybe I know a few things about what it seems to prefer - a specific geology, proximity to a stream, not-too-steep of a slope, and OH! perhaps some habitat characteristics. Great! 

You can pull all those underlying datasets in to ArcGIS Pro: A geology layer, run some Distance Accumulation on a stream layer, derive some slope and pull in some land cover data. On top of that, let's pull in our presence data for our species of interest. 

Now- we're going to run the MaxEnt tool, which is basically us asking the machine to look at where presence occurs, consider the underlying datasets that might EXPLAIN why this species is here and not there, and then score the rest of our study area based on how close we find similar criteria that match. A better score, means the higher the probability is that the right conditions might occur for this species to also be found there. 

What are some considerations?

Again, I totally recommend reading the documentation for the details, but in short, I always tell folks that running this tool does actually require one to know something about the phenomenon you're trying to run it against. 

Why? Because modeling anything - using MaxEnt or otherwise - requires iteration. You must do this a few times. Here's a good example: I once ran MaxEnt on a salamander species and I got a REALLY GREAT result the first time I ran the model. BUT.. when I actually looked at the output and not just the score, I realized that the model had picked up on what's called "sample bias"... I happened to have a lot of records near ROADS, because I would stop the truck, get out and sample, and find a record. So my first run told me ROADS ARE AWESOME FOR SALAMANDERS! 

Oops. I had to remove my sample bias through thinning (a tool available in the Presence-only Prediction Tool in ArcGIS Pro) and run it again, then start to look at my explanatory variables (geology, slope, etc) and ask myself "Is this layer helping my model at all?" and decided whether or not I wanted to keep it. It's pretty easy to do this, because as part of your output, you get little graphs of each of your explanatory variables that tell you which attribute of that variable matters most and how much for influencing the model. 

Pretty awesome, right? 

And you can do this for plants, animals, or other kinds of phenomena! It's SO EASY and super powerful and has tons of potential use cases (wildlife corridors, I'm looking at YOU!) 

For more info check out this blog post.