Practical questions

How much costs Biomapper?

Does Biomapper work with Idrisi32?

I'm working with Arcview. How can I use Biomapper?

How to convert an ESRI grid into an Idrisi/Biomapper raster map?

In which programming language is Biomapper developped?

Could you kindly send me the user manual.

How to quote Biomapper / ENFA?

Why is it not possible to use most of the modules of biomapper with Idrisi16 (.img) files?

Statistics

I don't understand what mean all the validation statistics

What is the "modified Kappa coefficient"?

Do you have an example data set to test Biomapper?

I have very good absence data. Can I use them with Biomapper?

How to use abundance data with Biomapper?

What is a Box-Cox transformation and why is it needed?

What is the "broken-stick advice"?

The number of categories per factor for the making of the HS map changes by steps of two. Is this normal?

I have computed an ENFA model and I would like to extrapolate it on a wider / other area. Is there some equation I can use to do it?

How is computed the score matrix?

Whate are global marginality, specialisation and tolerance? 

How do you compute the HS value for each cell from the scores matrix?

How to interpret factor biological meaning?

Box-Cox normalisation fails with some EGV maps. Should I discard them?

How are the ROC curves and kappa calculated if Biomapper does not use absence data?

How can I compare results from GLM and ENFA?

What is the difference between unidimensional multidimensional histogram algorithm for the computation of HS maps and how strong is the impact on the resulting HS maps?

Can I compare the global marginality and specialisation coefficients of different species if I use the same area but different set of ecogeographical maps, in particular if have had to discard different correlated maps in the process?

You say that sometimes the marginality factor takes also a part of the specialisation into account. Where can I find how much?

I know that the habitat suitability of my species is linearly related to some variable. Nevertheless, in the HS map, the optimum for this variable seems not to be at an extremum. Why?

Procedures

Importing files

• How can I create the species-presence map?
• How to convert a point vector map into a species map?
• How to convert a polygon vector map into a species map?
• How to insert the species-presence map?
• How can I use satellite imagery pictures (or other file format) with Biomapper?

 

 
 
 

Model evaluation
 

• Is it possible to obtain a better model by reducing the amount of explained variance?
• How to compare models obtained by various factor number/class number trade-off?

 

 
 
 

Miscellanies
 

• How can I print a map from biomapper? I guess I can do it from Idrisi but I´d like to use the rainbow palette.
• What are these "Biomapper extensions" used for? Can I ignore them?
• In the Options dialog box, it is possible to switch between correlation and covariance matrix, and to change the norming of the eigenvectors. However, this doesn't seem to affecte ENFA outputs.

Problems

• I get an error message "... is not a valid floating point value"
• When I try to open a project, I get a message "Documentation file not found".

Miscellanies

• Do you have a published paper about ENFA?
• Is Biomapper better suitable for plants or animals modelling?
• How to quote Biomapper / ENFA?

Biomapper is free. You just have to download it.

Yes, Biomapper can read both Idrisi16 and Idrisi32 file format.

Several of Biomapper's users are using Arcview. You will have to prepare your maps in Arcview before converting them to Idrisi/Biomapper format. Once the whole process has been completed, you can reimport the resulting maps into Arcview for further analysis or display.

If you are using Arcview 3.0, you can use the Biomapper module Grid Convertor.

There are several possibilities:

1° ArcView 3.x extension
  There is an extension doing the job on the ESRI site, made by  Holger Schäuble. Look for Grid Converter  (av2idrisi.zip) on http://gis.esri.com/arcscripts/index.cfm. It works for ArcView 3.x.

2° Biomapper's GridConvertor
  If you have Arcview 3.0 or 3.1 and Spatial Analyst, you can use the Biomapper module Grid Convertor. It allows to convert several files in only one operation.

3° Manual conversion
  If you have another version of ArcView, GridConvertor may not work (always problems due to ESRI proprietary policy).You will therefore have to convert your grids manually. In the Biomapper help file you will find a description of the Idrisi/Biomapper file format that should help you to transfer your data.You can try this: in Arcview, go into the file menu and select Export Data Source and export your image as a binary raster (has an flt extension). Exit Arcview and then change the file extension from .flt to .rst. You should then be able to create a document file (*.rdc) for this image with the information supplied in the help file. The flt and rst format are identical. You will probably have to choose "real" data type.

4° Otherwise...
  You can also find useful programs at http://www.pierssen.com/idrisi/grid.htm

I'm using Borland Delphi for all my programming work. It allow me to quickly conceive fast running procedures cloaked in a user-friendly interface.

There is no user manual for Biomapper. There is the help file though, and on the web, you will find the FAQ ( www.unil.ch/biomapper/faq.html ) and more general information.

Biomapper has no importing capabilities. It can build a map from raw data but cannot convert alien file formats. It works only with the Idrisi (16 and 32) file format but Idrisi itself has an extensive set of conversion tools.

You can either quote the main paper:

Hirzel, A., Hausser, J., Chessel, D., Perrin, N., in press. Ecological-Niche Factor Analysis: How to compute habitat-suitability maps without absence data? Ecology.

or the Biomapper software itself:

Hirzel, A., Hausser, J., Perrin, N., 2001. Biomapper 1.0. Lausanne, Lab. for Conservation Biology. URL: http://www.unil.ch/biomapper.

  Only the main Biomapper program has a button to switch from one format to the other. The other module use one of them according to the folder in which they are placed (it is not very elegant, I agree, and I will have to modify that some time in the future).

  When the module is launched, it checks if there is an "idrisi.env" file in the same folder as the module executable. If there is one, it switches to the Idrisi16 mode, else it switch to Idrisi32 mode.

  You can therefore enforce one mode or the other by placing a fake "idrisi.env" in your Biomapper folder.

  I agree that this validation part is not yet very well documented... It is the fruit of hard work to find a way to evaluate a model without absence data. I tried thus many methods that are still present in the output, without any further explanation... I shall try here to unveil a few details on this subject. This will finally be incorporated into the help file.

  You must have parted your sample into two sets. For validation purpose, you must use as "validation map" the set you did NOT used for model calibration. This observed map is therefore a boolean map indicating species presence. You can then evaluate the habitat suitability map produced with the ENFA model.

  When no reliable absence data are available, evaluation consist to compare various statistics computed on the "predicted map" (the habitat suitability map): 1° On the whole study area 2° Only on the validation points.

  The best way to understand this is to look at the box-plot displayed after the validation process. A good model should produce high species HS-value (80-100). The global box-plot  gives you insight on how marginal the species is in the studied area and thus how much these results could have been obtained by chance only.

  The results window gives a few statistics to resume this box-plot. It is composed of three parts: 1° species, 2° global statistics and 3° Comparisons.

  The two first part give identical information: first, common distribution statistics are computed (mean, median, quartiles, etc.). Then a few more specifics statistics: the one I find the most useful is the "proportion of presence cells >50": this is the proportion of validation points that have a predicted HS-value over 50. The higher this value, the better your model.The "Prob. to be above this value by chance" statistic use a bootstrap procedure to assess how much this value could have been obtained by chance. Practically, it is not very interesting as I have allways got here a 0.000 probability (good news)... Then you can see the 90th percentile and 95th percentile. You can compare all the previous values between global and validation sets, but only the validation set gives a really objective idea of the quality of your model.

  The last part give three comparison values between the two sets: The "Kappa" coefficient is a modified kappa statistics that integrate both how good is the model and how far from random it is. "Prop.of pix being significantly above 50" is the difference between the two "Prop.>50" values. And the "Probability to be over 50%" is the probability to be over 50% by chance, computed on the global set distribution histogram.

  These three last values are interesting to assess how the model is different from what could be achieved by a random model but it says nothing about the absolute quality of your model. They are highly related to the global HS of the study area and thus, if your species is not very marginal nor very specialised, the model could be very good but get a very bad "far from random" score.

I gave this name to a home-made statistic whose behaviour was to be similar to the Cohen's Kappa coefficient. It is computed as follows: 

K_mod = ( M_S-M_G ) / (1- M_G )

Alas, none of our data are in public domain and we cannot give them away.

If you are really sure that your absence data are good and that no historical or spatial factors could have biased them, you will probably get a better model by using a GLM or a GAM. But you can ever put your absence data aside and apply ENFA on presence data only.

So far, the only way to use those data is to "booleanise" them, i.e. transform them into presence data. But as many users seem to have this kind of data, I will probably improve the ENFA to take abundance into account. Stay tuned... (If you want to be alerted when new versions of Biomapper are published, you need to be registered on my "Biomapperians" list).

Box and Cox (1964) developed a procedure for estimating the best transformation to normality within the family of power transformations:

Y' = (Y^L -1)/L  (for L<>0)
Y' = ln Y           (for L = 0)

See "Biometry", Sokal & Rohlf, 1995, pp.417-419 for further explanations.

Biomapper uses the Box-Cox algorithm to normalise as well as possible the ecogeographic variables. Empirically, we have found that normality was not a crucial factor and this step could as well be ignored.

  The distribution of the eigenvalue is compared to the distribution of Mac Arthur's broken-stick. It is the expected distribution when breaking a stick randomly. Therefore, the eigenvalues that are larger than what would have been obtained randomly may be considered "significant". You can also keep only the factors with an eigenvalue larger than 1. These are objective means to choose how many factors to keep for HS map computing. These are just indications designed as a support when selecting the factors.

  Yes. It is because the HS computation is based on the median of the factor distributions and the median must fall between two classes and so  the number of classes must be even.

There is no such equation but the scores matrix computed by the ENFA may be used for extrapolation purposes. Here is the extrapolation procedure:

1. Create a project for the calibration area and compute an ENFA model.
2. Save the score matrix.
3. Create a new project for the extrapolation area. The EGV maps must be equivalent to that used in the calibration model (same units). They also must be sorted in exactly the same order; the best way to achieve this is to give them the same name as in the calibration project as EGV maps are sorted alphabetically.
4. Compute the covariance matrix of these EGVs.
5. Load the calibration score matrix

Compute the HS map. It will be based on the calibration model applied to the extrapolation data set.

  The full mathematical details of this operation are described in a paper currently in press for Ecology. Biomapperians will be alerted when it will be published. You can get an intuitive understanding by looking at the help file or on this site at ttp://www.unil.ch/biomapper/enfa.html . For now, I shall give you here a short view of this process:

  The eigenvalues and eigenvectors are extracted as follows: Compute the matrix W=Rs^-1 * Rg where Rg is the global correlation matrix and Rs the species covariance matrix. From W, extract the marginality factor (I don't give here the mathematical procedure), which gives us the matrix W^*. The specialisation factors are computed by extracting the eigensystem from W^*.

    Global Marginality = M = Sqrt(S_i=1,v[M_i²] )/1.96
    Global Specialisation = S = Sqrt(S_i=1,v(l_i)/V)
    Global tolerance = T = 1/S

where M_i are the coefficients of the marginality factor, Sqrt() is the square root function, V is the number of variables and l_i  are the eigenvalues.

  This is a rather complex procedure. The full mathematical details of this operation are described in a paper currently in press for Ecology. Biomapperians will be alerted when it will be published...

  Shortly said, for each retained factor, a frequency histogram is computed over all the cells of the map. The median of this distribution is computed. The further a cell is from this median, the lower is its suitability for this factor. The global suitability is then obtained by computing a weighted mean on these "partial suitabilities". Marginality has a weight of 1, the sum of specialisation factors has a weight of 1, proportionally to their eigenvalue.

  Look at the scores matrix. The first column of this matrix is the marginality factor. The other columns are the V-1 specialisation factors. (V is the number of variables). The rows are the EGV contributions to each factor.

  For myself, when the Box-Cox fails, I keep the original map. A "Box-Coxised" map gives better results than a "brute" map, but a "brute" map is still better than no map at all. 

  Anyway, you may include it at the beginning (to compute the big time-consuming covariance map. Once it is computed, you can remove easily variables from it, but if you add new variables, the whole matrix will have to be recomputed) and then try to remove it to see how it affects the result.

  The kappa and the ROC curves that you find in the validation dialog box assume that blank values are true absences. They are therefore not suitable for most data sets where absences are unreliable. I put them here for the cases where you can rely upon absences.

  Comparing ENFA and GLM is a tricky stuff. In my recent paper (Hirzel et al, 2001, Ecological Modelling 145), I was able to compare them because I was using virtual data and so I had access to the "reality", the "truth". But in the general case, we do not know it and so we are constrained to use the standard statistics (Kappa, ROC, etc.). There are three main problems when comparing presence/absence to presence-only methods:

1° If absences are thought to be unreliable to build a model, there is no sense in using them to validate it afterwards. So, the standard statistics are not useful. I tried to develop a few statistics to replace them (see the FAQ on the Biomapper site) but the perfect statistic has still to be invented.

2° As it is based on presence data only, the ENFA is more efficient to model the areas with average to high suitability; its predictions for low-suitabitility regions should be taken with prudence. By contrast, presence/absence methods and in particular GLMs, will tend to model good versus bad areas causing such a kind of "stepped" response which is different from the linear one of the ENFA.

3° Without absences - i.e. bad habitat points - to "fix the floor", ENFA must scale its suitability index to the ceiling. That means that, on an ENFA HS map, you will always, by construction, have at least one cell with a HS of 1 (or 100) ; with GLM, it is generally not the case as it is computing "probability of presence" and so the maximum values are generally quite lower than 1.

  Thus, when comparing visually GLM and ENFA maps (computed on well-known species at equilibrium, i.e., when absence data are largely reliable), the results are obviously quite similar. But if you try to compare them statistically, you get strange results biased either for one or for the other depending on which base hypothesis you use. To compare them you must correct both results to make them comparable:

1° Remove the "ceiling effect" by stretching the GLM results between 0 and 1.

2° Synchronise the "step effects" by transforming both GLM and ENFA results into boolean maps (by choosing a threshold).

3° Then you can compare these results with standard presence/absence statistics.

  I fear you will have to wait the publication of our main paper on ENFA (Ecology, in press, hopefully early in 2002) to fully understand these algorithms, but I can give you here a "feeling" of what they do and how they differ:

Unidimensional algorithm:

  Once the ENFA factors have been computed, it is possible to compute for every cell of the map a value along each of them (in fact, usually, one computes it only for a few of the first factors). The distribution histogram of each factor is then computed taking into account only the presence points. Each histogram will be used to attribute a "partial suitability" value to every cell (the more the cell factor value depart from the median of the histogram, the lower its "partial suitability". Then, a weighted sum of these partial suitability values is made for each map cell, and finally these sum are stretched in order to have their maximum at 100.

Multidimensional algorithm:

  Here we do not address the selected factors independently. By crossing all factors together, the factor space is divided in small units (hypercubes). Then we count how many cells belong to each hypercube : this is the multidimensional histogram. This multidimensional distribution is computed both for all cells (global distribution) and the presence cells (species distribution). Finally, the HS value is computed for each hypercube by dividing the species hypercube by the global hypercube (and multiplied by 100).

  The problem with the multi algorithm is that it need very huge amount of presence data to be accurate, in particular if you want to include more than two factors, which is generally the case. It is also very memory-consuming. So far, I never got good results with this algorithm and it is why it will not be described in your paper. We strongly advise Biomapper users to use only the unidimensional algorithm. In fact, I could well have removed it from the interface...

  I am currently working on new kinds of algorithms, but it is another story... Stay tuned!

  Strictly speaking, you cannot. Practically, the main biasing effect is the study area. When you have to remove a variable, it is because it does not contain more information than is already included into the model, so removing it should not alter significantly the global marginality and specialisation coefficients. Thus, provided the map sets do not differ drastically, you can still comparer them by these statistics. Anyway, do not assume too much significance to small differences in marginality or specialisation between species

  To be true, I never tested this. You could test it by building a common minimal EGV set and apply it to all your species. You will then see how the marginality et specialisation differ between the common data-set and the species-optimal one. Tell what you get, should you decide to try this.

  The amount of variance explained by the first factor is in fact the amount of specialisation. It generally ranges from 10% to 70%. This value is given in the eigenvalue table.

  To summarise, the marginality factor explains always 100% of the marginality and some part of the specialisation. It is why it has always a great weight (minimum 0.5) for HS computation.

  The HS computing algorithm is not linear but bases itself on the observed distribution. Your problem arises generally on the marginality factor. Let's imagine a species linearly related to the frequency of forest and that this variable is strongly correlated to the marginality factor (but the reasoning hold also for specialisation factors) ; it means that, the more forest there is around a given cell, the more suitable it is for the species, the maximum being a frequency of 100%. This optimum is what we know from our  knowledge of the species, field studies, etc. Now, let's see the Biomapper's point of view: To it, the optimal frequency is the one where the species is the most frequent (More precisely, the median of the species distribution along the marginality factor. As the distribution of this factor is generally unimodal and more or less symmetrical, the median corresponds also to the most frequent.) Therefore, if large forests are rare in the study area, points with 100% forest freq. will be rare too, and the optimum for the species will not be 100% but lower (say 80%). Then, when computing the HS index, 80% freq. will provide the highest partial suitability value and this will decrease when freq. is either increasing or decreasing. The rarer the large forests, the steeper the rate of decrease.

  Sometimes this effect is welcome (when dealing with median optima) and sometimes counterproductive (with extreme optima). I am currently working at new algorithms which will hopefully address this problem. 

There are several possibilities depending on what kind of data you have at hand:
• List of observations coordinates: Put them in an ASCII file, using a structure x-coordinate tabulation y-coordinate (You can do this with Microsoft Excel) and the use the "Convertor" module to create a boolean presence map from this file.
• Observation map in a point-vector-file: Simply rasterise it (with Idrisi function "PointRas"), using the same resolution and window as your ecogeographical maps.
• Population map in a polygon-vector-file: First, rasterise it (with Idrisi function "PolyRas"), using the same resolution and window as your ecogeographical maps. Then make this map boolean (1=inside populations). Finally, use the module "Sampler" to divide this map into a calibration and a validation data sets.

In Idrisi32:

• Menu Reformat/Raster-vector conversion/POINTRAS
• Select your vector map, e.g. "Species.vct"
• Choose a name for the "image file to be updated". It must be a new name (not an existing raster) (e.g. "Species_bl.rst")
• Choose "Change cells to record the presence of 1 or more points"

Idrisi asks you if you want to bring up INITIAL: Answer YES.

• In the INITIAL dialog box:
• Select "Copy spatial parmeters..."
• Select one of your EGV maps in "Image to copy parameters from"
• In "Output data type", select "Byte"
• Click OK

  And this should do the job. Species_bl can now be used as species map. You may want to partition it into calibration and validation data sets. You can do this with the Biomapper module Sampler.

  Rememember that all the EGV maps and the species map must be in the same directory.

In Idrisi32:

Menu Reformat/Raster-vector conversion/POLYRAS

Select your vector map, e.g. "Species.vct"

Choose a name for the "image file to be updated". It must be a new name (not an existing raster) (e.g. "Species.rst") 

Idrisi asks you if you want to bring up INITIAL: Answer YES.

• In the INITIAL dialog box:
• Select "Copy spatial parmeters..."
• Select one of your EGV maps in "Image to copy parameters from"
• In "Output data type", select "Byte"
• Click OK

Then, you must booleanise the rasterised polygons:

• Menu Analysis/Database query/image calculator
• Select "Logical expression"
• Type Species_BL = [Species]>0
• Click OK

  Species_BL can now be used as species map. You may want to partition it into calibration and validation data sets. You can do this with the Biomapper module Sampler.

  Rememember that all the EGV maps and the species map must be in the same directory.

Once the species-presence map has been created, you must insert it in Biomapper in order to use it in the analyses. The species-map must be inserted in the Work maps list (NOT the EGV maps). This can be done in the  Files/Work maps/Add maps menu. Once inserted in this list, you must declare it as the current species map by selecting it, right-clicking on it and selecting "Mark as species map". The current species map is then marked by a red circle in front of its name.

  Yes. You must find the best trade-off between explained variance and smoothness of the HS model. Note that it is generally not useful to select more than 10 classes.

  You can use the validation module of Biomapper. Look at the box-plot it generates. Focus on the species box; it must be as high as possible. The higher and the narrower the better. (The global-box is not useful to assess model quality; it is here to see how different from randomness is your model. If you built your model on an area globally good for your species, you will get a good model simply because the species can live everywhere.)

  Display the map and save it as a BMP file. This file you can insert in a word document or print with any picture software. The rainbow palette cannot be used in Idrisi because it uses more than 256 colors.

  These extensions are made to simplify the browsing and help the user to select among all the maps, those having the right data type. But these extensions are not used by Biomapper to verify the maps; it uses the raster documentation file (*.rdc). Thus you can as well ignore the biomapper extensions.

These options are intended for the Principal Components Analysis. They do not affect ENFA indeed.

  It happens because your computer is not set to use "." as decimal separator. It is necessary to use Biomapper and greatly advised to use Idrisi. You can change this setting in the Start Menu/parameters/Control panel/Regional settings/Numbers.

  This message means that Biomapper could not find a map. This happen generally because the map is not in the same directory as the project file (*.bio). All EGV maps must be in the same directory. This constraint was introduced in autumn 2000 in order to make the transfer of project between different computers transparent. By now, when you create a new project, Biomapper will verify that all your files are in the same directory. However, you could have problem with project created with previous versions.

  There are already a few papers. More are coming. You can get a list of related publications on the Bibliography page. There, you can also download a PDF version of my PhD thesis. Registered users will be informed when new papers are coming out. 

  So far, Biomapper has been applied on a wide range of the life tree: Mammals, Birds, Insects, tropical Trees and Ferns. Most of them are animals but it mainly because I know more zoologists than botanists...

  In fact, absence data tend to be poorer for animals and so Biomapper could be particularly adapted to this case.