Download presentation
1
Modeling Species Distribution with MaxEnt
Bryce Maxell, Acting Director, Montana Natural Heritage Program & Scott Story, Nongame Data Manager, Montana Fish, Wildlife and Parks
2
Agenda - Wednesday 8-9 Introduction to MaxEnt
9: Reptile and Amphibian Model Examples 10: Installation and Walkthrough of MaxEnt 11: Preparation of Data Lunch 1-1:55 Thresholds & Model Validation Using models in your DSS Hands-on Session Tomorrow Hands-on, Data Prep, Questions & Discussion
3
About to start again folks on the phone.
4
Installing and Running MaxEnt
INSTALLATION
5
Download & Install http://www.cs.princeton.edu/~schapire/maxent/
Current MaxEnt Version = 3.3.3e Requires Java Version 1.4 or later Type java –version at command prompt Extract the .zip file to a very simple directory No spaces, no strange characters, short C:\maxent Three files are installed Maxent.bat Maxent.jar Readme.txt Download the tutorial Word document
6
Check Java Version
7
Set PATH and customize .bat file
My Computer Properties Advanced Environment Variables System Variables PATH Edit Add to end of the PATH ;c:\maxent Change the maxent.bat file Change the extension to .txt so that you can edit it with Notepad Change line reading java -mx512m -jar maxent.jar to… java -mx512m -jar c:\maxent\maxent.jar Change the extension back to .bat Note that changing the 512 to another number allocates more memory 512 Mb = 0.5 Gb 1024 = 1 Gb 1536 = 1.5 Gb 2048 = 2 Gb
8
Running MaxEnt Basic modeling run
9
Required Inputs Species presence localities (“samples”) file
Environmental feature layers Output directory Note that coordinate systems other than Lat/Long are permitted but the samples file coordinate system must match that of the environmental feature layers We will talk more about preparing feature datasets in the next discussion
10
MaxEnt – Main Screen
11
Supply presence localities
12
The file can have a .txt or a .csv extension.
The file should have a header Multiple species are permitted The x coordinate must come before the y coordinate Note that duplicate points will be dropped (duplicates are those that fall in same grid cell) Keep track of the points that you use in a database if you want (might want to preserve a unique identifier from your point observation database) You will get warnings for any points that fall outside of any of the input feature layers
14
Supply folder containing environmental feature layers
16
Change variable types as necessary
Supply an output directory
17
Ready to Run
18
What MaxEnt Does Reads through each layer to Determine type
Create .mxe file for each layer in maxent.cache Extracts the random background and sample data You will get warnings about points that are “missing some environmental data” Calculates the gain until a threshold is reached Creates the output grids for each species (this takes the longest) Creates the thumbnail .png images
19
Time Required Ten feature layers (3 categorical) 2 Species
46 million pixels 2 Species Intel Core 2 Quad CPU (2.83 GHz) 4.00 GB RAM Windows 7 32-bit Operating System 512Mb of memory specified Without maxent.cache = 38 minutes With maxent.cache = 24 minutes
20
Running MaxEnt Examining output
22
Output plots folder logfile maxentResults.csv For each species .asc
.html .lambdas _omission.csv _sampleAverages.csv _samplePredictions.csv maxentResults.csv contains all of the variable importantce, threshold information, one row per species Html has pointers to a variety of plots omission receiver operating curve table of threshold values pictures of the model analysis of variable contribution raw data outputs and control parameters
23
Logfile Timestamp Version of MaxEnt Samples file name Warnings
Command line to repeat Species Layers Layertypes Directories for: samples file, layers, output Number of samples Maximum gain
24
Gain Closely related to deviance, a measure of GOF in GAM and GLM
Starts at zero and heads toward an asymptote MaxEnt trying to come up with best fit Average log probability of presence samples minus a constant Gain indicates how closely the model is concentrated around presence samples Avg likelihood of presence samples = exp(gain)
25
Gain Examples McCown’s Longspur Olive-sided Flycatcher
Resulting gain: 2.275 Average likelihood for presence points = 9.728 Olive-sided Flycatcher Resulting gain: 1.297 Average likelihood for presence points = 3.658 Average likelihood of the presence sample is X times higher than that of a background pixel
26
Html Analysis of omission/commission
Receiver Operating Curve (AUC calculated) Preset Thresholds Pictures of the Model Analysis of Variable Contributions Raw Outputs
27
Omission Rate vs. Cumulative Threshold
28
Receiver Operating Curve
29
Sample Predictions File
Coordinates for all points Test or Training Predicted values Raw Cumulative Logistic Use this file to calculate deviance Use samples procedure in ArcMap to extract the ones and zeros (above threshold or not)
30
Sample Predictions File
31
Logistic Ouput High probability of suitable conditions
Low predicted probability of suitable conditions White dots = training (1059 points or 75%) Purple dots = test (352 points or 25%)
32
Viewing Data in ArcMap Build Raster Attribute Table (Categorical)
.vat.dbf Build Histograms (Classified) .aux Build Pyramids .rrd .xml For species output grids Convert ASCII to Raster (Output Data Type = FLOATING) Output as .bil (Band interleaved by line)
34
MORE Advanced parameters
Running MaxEnt MORE Advanced parameters
35
Running MaxEnt Replicate runs
36
Running MaxEnt BATCH MODE
37
Preparation of Data Scott Story
38
Required Inputs Species presence localities (“samples”) file
Environmental feature layers Output directory Note that coordinate systems other than Lat/Long are permitted but the samples file coordinate system must match that of the environmental feature layers We will talk more about preparing feature datasets in the next discussion
39
Getting Feature Data Ready
Same projection (coordinate system, units, datum) Same resolution Same extent ESRI ascii format
40
Two Raster Datasets Land cover Precipitation
Source = Montana Natural Heritage Program Type = IMAGINE Image Cell size = 30 meters Columns & Rows =33005, 24008 Spatial Reference = Montana State Plane (NAD83) Pixel Type = Unsigned Integer (8-bit) Source = PRISM Climate Center Type = ASCII grid Cell size = Columns & Rows = 7025, 3105 Spatial Reference = undefined (see metadata) Pixel Type = Signed Integer (32-bit)
41
Two Raster Datasets Land cover Precipitation
42
Making Rasters Match Define coordinate systems for both
Set some environment variables Tools Options Geoprocessing Tab Environments General Settings: Extent and Snap Raster Raster Analysis Settings: Cell Size, Mask Project Raster Select target raster to match for output cell size
43
Precipitation Reprojected & Resampled
Same exact extent Same exact number or rows & columns Same exact cell size Real test is…does Maxent throw any errors? In this case…it worked! Getting all your data layers squared away will take some time!
44
Deriving New Raster Data - Ruggedness
45
Types of Environmental Features
Continuous (Quantitative) Interval-scale (interval data, order, linear scale) Ordinal variables (scale unknown-transformed?, rank clear) Ratio-scale (interval data, ordered, not on linear scale, e. g. temp on F or C scale) Categorical (Qualitative) Nominal (e.g. gender) Ordinal (has order, e.g. low to great) Dummy variables from quantitative (classes) Name the ASCII files with CONT or CAT prefix
46
Preparing Point Data Create a separate file for each species
Combine them all\groups of them into one file Probably want to retain a unique identifier May want to setup scripts in ArcGIS to extract presence data Might also want more control of how background data is selected Let’s look at an example script - ExtractModelInputData.py
47
Other “Feature” Layers
Masks useful if you want to train a model using only a subset of the region mask.asc containing a constant value (1, for example) in area of interest and no-data values everywhere else. Bias assumption that species occurrence data are unbiased good understanding of the spatial pattern values should indicate relative sampling effort
48
Representing the output
THRESHOLDS
49
Logistic Output (Ranges 0-1)
50
Reclassify with ArcGIS
51
Preset MaxEnt Thresholds
Cumulative Threshold Logistic Threshold Fractional Predicted Area Training Omission Rate Test Omission Rate Fixed Cumulative Value 1 1 0.043 0.344 0.002 0.000 Fixed Cumulative Value 5 5 0.172 0.255 0.020 Fixed Cumulative Value 10 10 0.260 0.210 0.044 0.082 Minimum Training Presence 0.699 0.029 0.365 10 Percentile Training Presence 17.522 0.351 0.167 0.099 0.151 Equal Training Sensitivity & Specificity 21.989 0.393 0.149 0.148 0.205 Maximum Training Sensitivity Plus Specificity 9.201 0.248 0.216 0.035 0.065 Equal test sensitivity & specificity 18.603 0.361 0.162 0.106 Maximum test sensitivity plus specificity 7.729 0.225 0.228 Balance Training Omission, Predicted Area, &Threshold Value 1.054 0.047 0.342 Equate Entropy of Thresholded & Original Distributions 5.465 0.182 0.250 0.021 0.026 Purple row/column are only calculated in test data is specified. Column of p-values is also present at far right. One-sided p-value for test of null hypothesis that test points are predicted no better than by a random prediction with the same fractional predicted area The “balance” threshold minimizes 6 * training omission rate * cumulative threshold * fractional predicted area
52
Thresholds – Ends of Spectrum
Balance Training Omission, Predicted Area, &Threshold Value Equal Training Sensitivity & Specificity
53
Model Validation MODEL VALIDATION
54
Validation Metrics Receiver Operating Curve – obtained by plotting, for each threshold in this range, the proportion of true positive against the proportion of false positive Area Under Curve – computed by computing the area under the above described curve Deviance – 2 times the log probability of the test data. Absolute Validation Index - the proportion of presence evaluation points falling above the threshold or within the GAP predicted distribution Point Biserial Correlation - The correlation between a model’s predictions and presence/absence in test data (regarded as a 01 variable)
55
_samplePredictions.csv
57
Discussion Point
59
Topics Left Data Prep Output Thresholds Validation Batch Replicates
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.