I.1What is PalynoPro?

It is a program for quantitative pollen morphometry that runs in the browser, with nothing to install. It takes a folder of scanning electron micrographs and returns, for each grain, its size, shape and ornamentation measured with the same criterion, and for each group, whether its pollen can be told apart from that of the others.

It is designed to compare groups—morphotypes, varieties, populations, closely related species—by their pollen, and for teaching palynology. It calibrates each image from the magnification of the microscope and checks it against the scale bar printed on the picture itself; it separates the grain from the background and reads its outline in 720 directions; it removes the spines to measure the body; it counts and measures the spines on the outline; it finds the lumina of the reticulum and measures the texture of the surface; it estimates the proportion of collapsed grains in the overview pictures; and it compares the groups with tests by variable, principal components, PERMANOVA, discriminant analysis and random forest. At the end it writes a report with the methods, the tables and the figures.

Everything is computed with engines written from scratch for this program, without other people’s libraries, and each method cites the work that defines it: on the app’s home page, in the notes of each block and in the methods of the report.

I.2Questions it answers

A good way to find out whether the program suits you is to see which questions it answers, and in which block.

I.3Six ideas behind this manual

Everything that follows rests on six ideas. If you keep them in mind, many of the mistakes made when measuring pollen in photographs are no longer possible.

1. The size of the pixel depends on the width at which the picture was saved. The magnification refers to a reference frame: the field measures that width divided by the magnification, and the pixel, that field divided by the width of the image in pixels. At the same magnification, a 1280 px picture has pixels twice as large as a 2560 px one. That is why Block 2 computes the pixel of each picture and checks it against its scale bar.
2. A scale bar is there to check, not just to decorate. Its length in pixels times the pixel size must give a round length: 10, 20 or 50 µm. If it does not, something is wrong—a mistyped magnification, a different reference width—and the picture is flagged.
3. Size is that of the body, without the ornamentation. Spines change the outline but not the grain. The app reads the outline as the distance from the center to the edge in 720 directions, cuts the peaks narrower than the base of a spine and measures what remains.
4. What you see is a projection. A micrograph shows the silhouette and the face of the grain that looks toward the detector. The Feret diameters are equivalent to the polar and equatorial axes only in equatorial view, and the spines that are counted are those on the outline, not all of those on the surface.
5. The grain is the unit, not the picture. The same grain is often photographed more than once. If every picture counted as a grain, the tests would believe they have more data than there are, and validation would classify a picture with the help of its twin. The app averages the pictures of each grain and validates by leaving out the whole grain.
6. Shape is not viability. A collapsed grain loses solidity, and that makes it possible to compare groups in the overviews. But it is a morphological approximation: to claim that the pollen is viable you need a stain or a germination test.

I.4Data it accepts and what it delivers

The starting point is a folder of micrographs, with or without the files that accompany them. The other blocks work on what has already been read: they do not ask for separate tables.

I.5What you need and how to open it

Only a modern browser, in a version from the last few years. You do not need to install anything or to have internet access; the whole program fits in one folder.

1. Double-click index.html. It is the simplest way and everything works the same: the calculations and the example are inside the program.
2. Or double-click Open PalynoPro.bat. It starts a minimal local server on port 9940 and opens the app at http://localhost:9940. You only need it to open the app from another computer or a tablet on the same network; the file server.ps1 explains how.

Preferences—the language and the light or dark theme—are stored in the browser itself and remembered next time. Data are not: each time you open the app it starts empty. To resume a study there is the .json project, which stores the settings, the list of images with their groups, their calibration and their exclusions, and the results of Blocks 3 to 5 (those of Block 6 are not stored: when the project is opened they are repeated, identical, with one click). If you try to close the tab with unsaved data, the app warns you.

I.6A tour of the interface

All the applications of the suite share the same bar, the same navigation and the same figure tools, so what you learn here also serves you in the others.

1 2 3 4 5 6 7 8 9 10 11 12 13

• 1Name of the app and its tagline. The name leads to the home page (Block 1).
• 2LABG Suite. Returns to the portal with the other applications.
• 3ES · EN. Changes the language of the whole interface, including that of the figures.
• 4Theme and shortcuts. The sun switches between the light and the dark theme; the question mark opens the list of keyboard shortcuts (? key).
• 5The bar of the seven blocks. The number turns into a check mark when the block already has a result, and the count of how many have one is shown at the top (3/7 in the figure).
• 6The path. The app, the block and the section you are in.
• 7Document · Focused. All the sections of the block one after another, or one at a time; the button beside them collapses or expands all the sections.
• 8Search (Ctrl + K). A block, a section, a control or a figure, by its name.
• 9In this block. The sections of the open block; the current one is highlighted.
• 10The title of the block, with its number and what it does in one line.
• 11How it is measured. A drop-down that explains, in plain words, what the block computes.
• 12A card. Each section is a card with its controls at the top and its results below.
• 13Search, Focused mode and Help, also at the foot of the block bar.

Getting around without getting lost

At the foot of each block two buttons lead to the previous and the next one. When you change a control and the results are recomputed further down, a floating button appears that takes you back to the control you touched, and another one to go back to the top. Long calculations—measuring all the pictures, counting in the overviews, comparing the groups—show a waiting window with their progress. Almost everything can also be done from the keyboard:

The check on the image

Blocks 3, 4 and 5 end with a check card: the picture with what the app measured drawn on top, so that you can see in each grain whether the outline, the body and the spines are the ones you would have marked. Each layer is turned on or off with its checkbox, and a badly measured picture is excluded with a button.

The figures

Every figure of the app can be opened in the figure studio, a window shared by the whole suite where the texts, the font size, the legend and the colors are changed without touching the data, and which exports at the widths and resolutions journals ask for. Next to each figure, the ✎ button opens the figure editor. The SVG and PNG buttons of each figure, and the report, always write the colors of the light theme, even if you are working in the dark one. Chapter 7 explains it in depth.

I.7The simulated example

To learn without images of your own, the app comes with a simulated example: three morphotypes of echinate, reticulate pollen that the program itself draws, every time it is opened, from known parameters. They are not photographs of any plant. It is opened with Try the simulated example on the home page or with Simulated example in Block 2, and the app says that it is simulated.

Each morphotype has eight grains. Each grain is drawn in a 1280 × 960 px picture at a magnification of 1000 (0.1 µm per pixel), with its noise, its scale bar and its caption, as a microscope would; grains 3 and 6 of each morphotype are photographed twice, as happens in a real study. In addition, each morphotype has an overview at a magnification of 150 with fourteen grains, some of them collapsed. In all: 30 one-grain pictures, which are 24 grains, and 3 overviews.

Since it is known what each grain was drawn with, the example also shows how well the app measures. Morphotype B, for example, is drawn with a body of about 68 µm in diameter, and the app measures a mean of 67.93 µm over its eight grains. The names of the pictures follow the form 01_A.png: the number of the grain and, after the underscore, the group; repeated pictures carry a letter (03a_A.png) and overviews the number 90 (90_A.png).

I.8How a study is carried out

The blocks are in the order of a real palynological study. You do not have to go through all of them: if you have no overviews, Block 5 is skipped. The How it works section of the home page sums it up in seven steps.

1. Open and calibrate (Block 2). Open the folder and check that the group and the grain of each picture were read correctly, that all are calibrated and that no bar is doubtful; mark which pictures are overviews and exclude those that are of no use.
2. Measure (Blocks 3 and 4). A single button measures all the one-grain pictures: the outline, the body, the spines, the lumina and the texture come out of the same pass. Check each picture on the image.
3. Count (Block 5). In the overviews, the proportion of turgid and collapsed grains of each group.
4. Compare (Block 6). The tests by variable, the principal components, the PERMANOVA and the two classifiers, with the grain as the unit.
5. Publish (Block 7). The report with the methods, the tables, the figures and one measured grain per group.

I.9How it was checked

A measuring program is worth what its numbers are worth. The PalynoPro engines were checked in two independent ways, and neither uses anybody’s pictures: grains drawn by the program itself, whose truth is known, and statistical calculations repeated separately.

• Image. Sixty one-grain pictures—round and elongated, with 14 to 40 spines, spines of 1.2 to 3.2 µm and meshes of 1.8 to 3.4 µm, at two image widths and at any rotation—and twelve overviews with collapsed grains were measured just as in the app. The 60 scale bars gave round lengths; the diameter of the body deviated by 0.20% in the median; the number of spines came out exact in all 60 pictures; the 141 grains of the overviews were found, and no turgid grain was taken for a collapsed one.
• Statistics. Thirty simulated data sets were computed with the app and, separately, with an independent free statistical environment, number by number.

The tests also show where the app falls short, and that was written down. Spine height comes out somewhat lower than the drawn height—about one pixel, 4% in the median—because the finest tip of the cone does not manage to fill a pixel; the diameter of the lumen is proportional to the mesh of the reticulum, but not equal to it, because the lumen is only the dark part of each cell; and the default solidity threshold is conservative: it recognizes most collapsed grains without confusing healthy ones. The chapters say so where it matters.

I.10What the program does not do

Knowing where a tool ends is as useful as knowing what it does. The app’s home page says so in What this program does not do:

1.1The home page

It is the first thing you see when you open the app, and also what appears every time you click the name PalynoPro on the top bar or the Home block on the bar on the left.

At the top come the label Quantitative palynology and the app’s tagline, From the electron micrograph to the pollen key, the same as on the cover of this manual. Below, a paragraph sums up everything the program does, and on the right a simulated echinate, reticulate grain shows what the app draws on each picture: the outline of the body (dashed line), the tips of some spines (circles) and the diameter it measures (the dimension line at the bottom).

1 2 3 4 5 6 7

• 1The tagline. With the label Quantitative palynology above it and the introductory paragraph below.
• 2Open my images →. Leads to Block 2, where the folder is chosen. It does not open any file dialog yet.
• 3Try the simulated example. Opens Block 2 and draws the example there (I.7): 33 images of three simulated morphotypes.
• 4See how it measures. Scrolls down, on this same page, to the lab of Section 1.3.
• 5Open a project. Asks for a .json saved earlier by the app (the same as Ctrl + O) and, once it is open, leads to Block 2. If the project already had measured grains, the notice reminds you that the results of Block 6 are not saved and are repeated, identical, with Compare the groups (Chapter 7).
• 6The four badges. Runs entirely in your browser, Your images never leave your computer, Engines checked independently and Spanish and English.
• 7The grain on the home page. A drawing made by the app, not a photograph; what it carries on top is the same as you will see in the check of each picture in Block 3.

As you scroll down the home page, the sections that follow in this chapter appear in this order. The In this block list on the right names them all, and one click takes you to each.

1.2How it works: the seven steps

How it works sets out in a strip the seven steps of a palynological study, in the order in which they are done, each with its label and what it produces (Figure I.6 of the introduction shows it in full):

The first two steps take place in the same block, and Steps 3 and 4 come out of a single measuring pass (the one launched in Block 3). Step 5 is optional: it only makes sense if you photographed overviews, at low magnification, with many grains in each picture.

1.3Try it: this is how it measures a grain

Before trusting a program that measures, it is worth watching it measure something whose answer is known. The lab draws a simulated grain with the values you choose and measures it at once with the same engines the app uses on your pictures.

Five sliders set the grain that is drawn; each time you move one, the grain is drawn again and measured again:

The six tiles

Each tile shows at the top what the app measured and below what was drawn. With the default values the app gives:

The body of the default grain is not a circle: with an elongation of 1.06, its drawn major diameter is 68.0 µm and its minor diameter 64.1 µm, and its body diameter—that of the circle of equal area, the geometric mean of the two axes—is 66.0 µm. That is what the app calls Body diameter in every block: a single number that sums up the size even if the grain is somewhat elongated.

Below the two figures, a status line says it in words. With the default values: “The measured body differs from the drawn one by 0.08% and 26 of the 26 drawn spines were found,” and it goes on with a note on the height of the spines: “about one pixel of height is lost (here 0.161 µm), some 3 to 6% on spines many pixels high and more if the spine is short or the spines touch at the base. It is reported in the methods.” (below).

The two figures

The simulated grain and what was measured is the simulated micrograph, in grayscale, with its scale bar and its caption at the bottom as a microscope would put them, and on top what the app found. Radial profile: outline, body and spines is the same information unrolled: on the horizontal axis, the direction from the center of the grain (from 0 to 360°); on the vertical one, the radius in micrometers. The orange line is the outline with spines, each peak is a spine (red dot) and the dashed blue line is the body that remains when the peaks are cut. The two broad waves of the body are the elongation: in an elongated grain the radius grows and shrinks twice per turn.

A small, elongated and very spiny grain

Move the sliders to see where the measurement is put to the test. Figure 1.3 shows a grain of 40 µm, with an elongation of 1.30, 40 spines of 3.5 µm and a mesh of 4.0 µm.

There are three things to learn from the two figures, and all three come back with your pictures:

1. Size and number of spines come out almost exact. In both grains the body differs by less than 1% from what was drawn, and all the spines are found.
2. Spine height comes out somewhat lower than the drawn one. The finest tip of the cone does not manage to fill a pixel, so the edge of the image ends a little before the spine does. The app measures each spine from its tip to the chord joining its two feet, and what is lost is about one pixel of height: with the default values, the measured height is 5.7% lower (2.08 against 2.20 µm, with pixels of 0.161 µm). In the grain of Figure 1.3 the loss is 6.8% (3.28 against 3.52 µm): its 40 spines almost touch at the base, which is precisely the case in which the status line warns that the difference grows. The program’s validation tests give a median of 3.9% below, between 2.9 and 5.7% in nine grains out of ten. The bias is similar in all grains measured with the same pixel, so it does not get in the way of comparing groups, but do not take the height as absolute.
3. The lumen is not the mesh. The mesh is the distance from one wall to the next; the lumen is only the dark part of each cell, without the wall. That is why the median lumen comes out at about half the mesh (1.49 µm with a mesh of 2.6 µm; 2.30 µm with a mesh of 4.0 µm). What matters is that it grows with it: it serves to compare reticula, not to report the mesh.

The grain of the lab is always drawn with the same seed, in an 800 × 600 px image, and with the magnification that makes the grain fill somewhat more than half of the field. That is why the same values always give the same numbers, and the scale bar comes out round (10.01 µm in Figure 1.2, 9.98 µm in Figure 1.3).

1.4The blocks

The blocks shows one card for each of Blocks 2 to 7, with its number, a small drawing, its label, its name and what it does. A click on the card opens the block, just like its button on the bar on the left. Blocks that need images or earlier results say so when they are opened and lead to the missing block.

1.5Pollen under the electron microscope

Pollen under the electron microscope, in plain words gathers, in six collapsible topics, the ideas behind the measurements. Each topic shows on the right the block where it is used; the first one comes open and the others open with a click on their title.

The introduction (I.3) turns these ideas into six rules, and each chapter applies them where they belong. If you are going to read only one topic before opening your pictures, read the one on calibration: it is the mistake that most often shrinks or enlarges grains without anybody noticing.

1.6The methods and where they come from

Fourteen cards name the methods the program uses, each with its family and the work that defines it. All are written from scratch for the app; the citation says where the idea comes from, not where the code comes from.

Each chapter explains the method where it is used, and the report of Block 7 writes out those that took part in your analysis, with their parameters and their citations.

1.7What this program does not do

Six cards say where the tool ends. The introduction (I.10) already gathers them in a table; here is what each limit means when planning a study.

• It does not rebuild the grain in 3D. Everything comes from the silhouette and from the face that looks toward the detector. The spines of Block 4 are those on the outline: a grain with 26 spines on the silhouette has many more on its surface. For comparing groups that is enough, as long as all are measured the same way.
• It does not guess the view of the grain. The major and minor Feret diameters are the polar (P) and equatorial (E) axes only if the grain is in equatorial view. If your description needs P/E, photograph oriented views and treat them as different groups, or record the view in the note column of the index.
• It does not read the number on the bar. The app finds the bar and measures it, but it does not read the “10 µm” written next to it: it checks that its length gives a round length (Chapter 2). If the bar of a picture is not round by design, measure it by hand in the viewer of Block 2.
• It does not replace a viability test. The proportion of collapsed grains in Block 5 compares groups by their shape; to claim that the pollen is viable you need a stain or a germination test.
• It does not measure the thickness of the wall. The exine, tectum and nexine are measured in sections or with transmission microscopy, not in an image of the surface.
• It does not upload your images. Everything is computed on your computer; the project stores references to the files, not the files (I.5).

1.8References and how to cite

References lists, in alphabetical order, the 22 works cited by the methods and the texts of the app: those of the methods in Section 1.6 and, in addition, Borgefors (1986), for the distance transform with which Block 4 counts the density of lumina away from the edge of the grain; Cohen (1960), for the κ coefficient with which Block 6 sums up the accuracy of the classifiers; Erdtman (1952), Halbritter et al. (2018) and Punt et al. (2007), for the palynological terminology; and McArdle and Anderson (2001), for PERMANOVA. Appendix F gives them in full, grouped by topic.

At the end of the home page, How to cite gives the citation of the program ready to paste, and the Copy the citation button copies it to the clipboard (the app confirms with “Citation copied”). In the Spanish interface the citation is in Spanish. Below it, at the foot of the home page, the User’s manual link opens this manual (in the Spanish interface, Manual de usuario (en español) opens the Spanish edition).

The DOI in the citation is the Zenodo concept DOI, which always resolves to the latest version. Cite the version you worked with—the one given by the citation in your copy—because results may change between versions. Along with the program, cite the methods you used: the methods text of Block 7 already includes their references.

2.1Opening a folder or the example

The block presents itself with its subtitle, Open a folder, check the groups and the calibration of every picture., a drop-down that explains it in two paragraphs, and a single card: the area where the images are dropped.

1 2 3 4 5

• 1What this block does. Sums up the formats it reads, the index, the key of groups (with examples such as A = Morphotype A) and the check against the bar.
• 2Choose a folder. Opens the system dialog to choose a whole folder, with its subfolders.
• 3Choose files. To choose individual files: images, their .txt, a .csv index or a key.
• 4Simulated example. Draws the example of the introduction (I.7). If images of your own are open, it asks before replacing them.
• 5Clear. Removes all the images and results, after you confirm “Remove every image and result?”.

You can also drag the folder, or a handful of files, from your system’s file explorer onto the dotted area of Drop a folder or your images here. Nothing is uploaded anywhere: the files are read on your computer, and each image is read in full only once—to learn its size, look for its scale bar and make a thumbnail—and again only when it is measured.

What the app recognizes in the folder

There is no need to sort anything: the app classifies each file by its extension and its name.

The index

Columns are recognized by their name, in English or in Spanish, regardless of capitals and accents, and the separator can be a comma, a semicolon or a tab. Each row is joined to its image by the file name, with or without the extension.

The key of groups

It gives each group code a readable name, which then appears in tables, figures and the report. For example, a file key.txt with these two lines:

The code is a single word of letters and digits, like those read from file names: A, G1, P12 or 7 will do, and so will one of up to eight characters with a hyphen or an underscore in the middle (north-2). What comes after the = or the colon is the name. If you prefer not to make a separate file, type the names in the Groups card (2.4) or in the name column of the index.

The reading message

When it finishes, a message below the buttons says what was read: how many images, how many microscope files, the name of the index and how many group names the key brought; and, if some image could not be read, its name and the reason. With the example, the message warns that it is simulated: “three morphotypes drawn by the app with known parameters (they are not photographs of real plants). Some grains were photographed twice, as in a real study.” The five cards that follow appear below it.

2.2The summary and its notices

Summary is the card worth looking at first: six tiles and, below them, the notices of whatever the app found. Figure I.5 of the introduction shows it with the example just opened.

The notices

Below the tiles the app writes one notice for each thing it found. With the example only the green one appears: “The 33 bars found give round lengths: the calibration is consistent.”

A doubtful bar does not take the picture out of the analysis: the app goes on measuring it with the calibration it has, because it cannot know which of the two—the magnification or the bar—is wrong. That is for you to decide. When you type 1000 again, the notice disappears.

2.3Calibration and reading of names

Three settings that apply to all the images, and that are not applied until you press Apply.

Width of the reference frame (mm)

The magnification an electron microscope writes refers to a reference frame of fixed width: the photographed field measures that width divided by the magnification. The pixel size is that field divided by the width in pixels of the image as it was saved:

The default value, 128 mm, is that of many instruments (those that leave text files with $CM_MAG lines). In the example: 128 × 1000 ÷ (1000 × 1280) = 0.1 µm per pixel in the one-grain pictures, and 128 × 1000 ÷ (150 × 1280) = 0.66667 µm in the overviews. If your instrument uses a different frame, the doubtful-bar notice will come up in almost all the pictures at once and in the same proportion: that is the sign that this number must be changed. An empty or non-positive value goes back to 128.

The check against the scale bar

When it reads each image, the app looks for its bar in the bottom fifth: a strip of almost white pixels, at least 2% and at most 45% of the width of the picture, that repeats with the same start and the same length (±2 px) over at least three consecutive rows, and that is almost completely filled. If there are several, it takes the longest. Its length in pixels times the pixel size must give a round length: 1, 2 or 5 times a power of ten (1, 2, 5, 10, 20, 50, 100 µm…), with a tolerance of 4%.

The bar also tells where the microscope’s caption begins—the strip at the bottom with the bar, the voltage and the other data—: the app cuts the picture a few rows above the bar, so that neither the bar nor the letters are taken for part of the grain. If there is no bar, it cuts at a dark strip at the foot, if there is one. The viewer (2.6) draws that cut.

Pixel of the analysis (µm)

Before measuring, all the one-grain pictures are brought to the same pixel size by area resampling (Wolberg, 1990). That way texture, lumina and spines are measured with the same fineness in pictures taken at different resolutions. Empty—the normal case—means automatic: the 90th percentile of the pixel size of the included one-grain pictures, rounded to two significant figures, so that almost none of them has to be enlarged. The help line below says what it came to: empty = automatic: 0.1 µm in the example, where all one-grain pictures have the same pixel. Pictures with less resolution than the analysis pixel are enlarged, and Block 3 flags them. Overviews are not brought to this pixel: Block 5 counts them at their own pixel (reduced, if they are more than 1600 px wide).

Type a value only if you want to fix it—for example, to compare two studies at the same pixel. A pixel finer than that of your pictures adds no detail: it only enlarges.

Pattern of the name (group and grain)

If there is no index, the group and the grain of each picture are read from its name (without the extension) with a regular expression that has two named parts, (?<grain>…) for the grain and (?<group>…) for the group. The one the app comes with is:

It reads like this: at the start, one or more digits (the grain); then optional letters (the one that tells a second picture of the same grain apart); an underscore; and letters followed by optional digits (the group). With it, 01_A is grain 01 of group A, and 03a_A is another picture of the same grain 03. The line This is how the first names are read checks it live on the first three images while you type: in the example, “01_A → group A, grain 01 · 02_A → group A, grain 02 · 03_A → group A, grain 03”.

When you press Apply, the app checks that the pattern is a valid expression (if not, it says so and changes nothing); if the pattern changed, it reads the group and the grain of the pictures again with it; it recomputes the calibration of all of them with the reference width, classifies the kinds of picture again and confirms with “Calibration applied”. Everything that had been measured in Blocks 3 to 6 is cleared, because it depended on the previous settings.

2.4Groups

One row per group, with its code and its color—the same it will have in every figure—, a Name field and two counts without the excluded pictures: Images (all) and Grains (in the one-grain pictures).

The name is used in the tables, the figures and the report instead of the code; if you leave it empty, the code is used. It comes from the key of groups, from the name column of the index or, in the example, from the app itself (“Morphotype A (simulated)”). Changing it clears no result: it only changes how things are labeled.

The code, on the other hand, is changed in the table of images, picture by picture (2.5). A group exists as long as some picture has it.

2.5The table of images

One row per image, with what the app read and computed from it, and the fields you can correct. At the top, seven filters choose which rows are shown: All, One grain, Overviews, Detail, Doubtful bar, Not calibrated and Excluded.

The three kinds of picture

The app proposes the kind of each picture by comparing its magnification with the median of all of them: if it is 0.6 times the median or less, it is an overview (many grains, for Block 5); if it is 1.8 times or more, it is a detail (a close-up of the surface); anything else is one grain (for Blocks 3 and 4). In the example the median is 1000×, and the three pictures at 150× are overviews. Block 3 measures only the one-grain pictures and Block 5 counts only the overviews; detail pictures are kept and can be seen in the viewer, but they are not measured. If the proposal is not right, change the kind with the menu: your choice is respected even if you press Apply afterward.

Correcting and excluding

Any change in the group, the grain, the magnification, the kind or the exclusion of a picture clears whatever had been measured in Blocks 3 to 6, because it no longer applies; their check mark on the bar goes out until you measure again. That is why it is worth leaving this block in order before moving on to the next. Exclude a picture when it should not count—out of focus, with the grain cut by the edge, with two grains on top of each other—; a picture that is not calibrated stays out by itself until it has a pixel size.

Download the index (CSV)

Saves the whole table as palynopro_index.csv, with the columns archivo, grupo, nombre, grano, magnificacion, ancho_px, um_por_px, fuente, barra_px, barra_um, nivel, excluir and nota (the column names are always these, in either interface language). It is a good record of how each picture was calibrated, for the supplementary material of a paper, and also an index that the app reads again: correct it in a spreadsheet, put it in the folder of the pictures and open the folder again. The fuente (source) and nivel (kind) columns are written with the words of the interface (with the interface in English, magnification, one grain, overview, detail), and when the index is read again the app understands the kind in either language.

2.6The viewer and the bar measured by hand

The viewer shows the chosen picture, in grayscale and at most 1200 px wide, with two marks on top, and on its right the data the app knows about it. When the block is opened it shows the first image.

1 2 3 4 5 6 7

• 1← Previous. The previous picture among those the table filter lets through; from the first it jumps to the last.
• 2Position. The place of the picture among the visible rows (1 / 33).
• 3Next →. The next picture; from the last it goes back to the first.
• 4Measure the bar by hand. Calibrates the picture with two clicks on its bar (below).
• 5The legend. detected bar: the yellow rectangle; start of the caption (what lies below is not analyzed): the dashed red line.
• 6The data of the picture. File, Group · grain (with the name of the group), Size, Magnification, Pixel size (with its source) and Field.
• 7Scale bar. Its length in pixels, what it measures and the verdict: round or expected.

For 01_A.png the data are: 1280 × 960 px, 1000×, 0.1 µm (magnification), a field of 128 × 96 µm and a bar of “100 px = 10 µm (round: 10 µm)”. If the microscope file has them, the Voltage, the Working distance and the Date are added, as well as the Note from the index. When the app replaced the pixel size from an index or from the instrument (2.2), it says here what it was and why it was not used.

Look at the two marks on any picture that seems odd to you. The yellow rectangle must surround the scale bar, not some other white thing in the caption; the red line must lie above the whole caption and below the grain. Whatever lies below the line is never analyzed.

Measuring the bar by hand

For a picture without a magnification, with a bar the app does not find, or with a bar you know is correct even though the calibration says otherwise:

1. Choose the picture in the table and press Measure the bar by hand. The app asks: “Click on both ends of the scale bar.”
2. Click on one end and then on the other. Each click leaves a blue dot.
3. Type the length of the bar in micrometers when the app asks for it; the question says how many pixels you marked.

The pixel size becomes length ÷ pixels, with the source by hand, which prevails over all the others; the app confirms with “Calibrated by hand:” and the value, and clears whatever had been measured afterward. Make the clicks carefully and on the largest picture you can: an error of two pixels on a 100 px bar changes every size in that picture by 2%.

A picture that came from a project but whose folder has not been opened again cannot be seen: the viewer says “The image is not open in this session: open its folder again.”

3.1How the body is measured

Below the title of the block, Outline and size, and its subtitle, “The body of every grain without its spines: diameters, elongation and circularity.”, the drop-down How it is measured sums up the method in two paragraphs. This section explains it step by step, in the order in which the app works through each picture.

1. The common pixel. All the one-grain pictures are brought to the same pixel size before measuring, so that a 2 µm spine spans the same number of pixels in all of them. That pixel is the one you typed in Block 2›Calibration and reading of names›Pixel of the analysis (µm) or, if you left it empty, the 90th percentile of the native pixel of the one-grain pictures, rounded to two significant digits: that way hardly any picture has to be enlarged. Resampling is by area (Wolberg 1990). The microscope’s caption is cut off first: whatever lies below the red line of the Block 2 viewer is not analyzed. In the example, the 30 pictures are at 1000× and 1280 px, and the common pixel is 0.1 µm.
2. Grain and background. In a scanning electron micrograph the grain is light on a dark background. Otsu’s (1979) threshold picks the gray level that best separates the two; pixels lighter than it are grain. Holes (dark areas surrounded by grain, such as the lumina) are filled, stray specks are removed and gaps of one or two pixels are closed.
3. The central grain. If several regions remain in the picture, the app takes the largest one whose center falls in the central 70% of the frame and that is at least 30% of the largest; if none qualifies, simply the largest. That way a neighboring grain peeking in at the edge is not mistaken for the one in the picture.
4. The outline in 720 directions. From the center of the grain, at every half degree, the app looks for the farthest edge pixel. The result is the radial profile: the radius of the grain in 720 directions, with the spines as narrow peaks. Chapter 4 shows it as a plot (Figure 4.8).
5. The body. A morphological opening (Serra 1982) is applied to that profile: every peak narrower than the Largest base of a spine (µm) is cut off, and what remains, smoothed, is the outline of the body. An undulation wider than that base is not a spine: it is the shape of the grain, and it stays in the body.
6. The measures. The area and perimeter of the body polygon are computed, and from them the measures of Table 3.1. The Feret diameters are obtained with rotating calipers on the polygon (Shamos 1978).
7. The cut outline. Directions in which the outline reaches the edge of the picture or the caption are cut: there it is unknown where the grain ends. If they exceed the Acceptable cut outline (%), the picture is flagged and left out of the size comparisons. Spines in cut directions are not counted either.

3.2Measure

The Measure card has three fields and a button. The default values suit most echinate grains of 20 to 100 µm photographed at 1000× or more; change them only if the check in Section 3.4 shows you that something is not being recognized well.

Next to the button Measure every one-grain picture a line tells how many pictures will be measured and at what pixel: with the example, 30 one-grain picture(s), at a pixel of 0.1 µm. Only the included pictures that Block 2 has as One grain enter; the overviews go to Block 5. If there are no images yet, the line says “First open your images in Block 2.” and the button is disabled.

When you press the button the waiting window appears, telling which picture is being measured and how many are left. At the end, a green message gives the total and the time—“30 of 30 pictures measured in 26.0 s.” in Figure 3.1; the time varies from one computer to another—and the block bar puts the check mark on Blocks 3 and 4. If no grain is found in a picture, the message turns amber and names the pictures after No recognizable grain:.

3.3Results by group

The Results card appears as soon as there are measured pictures. At the top, four tiles:

• Pictures measured: the one-grain pictures that were measured, including those with a cut outline.
• Grains, with a complete outline: the distinct grains that enter the comparisons. Two pictures of the same grain count once.
• Cut outline, left out of comparisons: the pictures that exceed the acceptable percentage. The tile has an amber border if there are any and a green one if not.
• Body (mean): the mean body diameter of all grains, all groups together, and below it the mean of its two Feret diameters.

If any picture had to be enlarged to reach the common pixel (its native pixel was larger), an amber notice says so: “… had less resolution than the pixel of the analysis and were enlarged: their fine measures (spines, lumina) are less reliable.” If any has a cut outline, another notice asks you to check them below. With the example neither appears.

The table by group

The By group table puts one group in each column, with its number of grains below the name (n = 8 in the example), and one measure in each row, under the heading Measure: the mean ± standard deviation of the six measures of Table 3.1. That way it fits on the screen even with many groups. Its caption says so: “Mean ± standard deviation among grains (the pictures of a grain are averaged).” The pictures of each grain are averaged first, and then the mean and the deviation are computed among grains: the standard deviation is the variation among grains, not among pictures. The CSV button downloads the table as palynopro_size_by_group.csv, with one row per group and a mean column and a deviation column for each measure.

The two figures

Body diameter by group is a box plot with the grains as points on top: the box runs from the first to the third quartile, the line is the median and the whiskers reach the most extreme data within 1.5 times the interquartile range. Feret diameters of the body places each grain with its major Feret on the horizontal axis and its minor Feret on the vertical one, with an ellipse enclosing 95% of each group. A round grain falls near the diagonal; an elongated one, below it.

3.4Check of every picture

No number is worth more than the outline it comes from. The Check of every picture card draws over each image what the app measured, so you can verify, picture by picture, that the grain, the body and the spines are the ones you would have traced.

1 2 3 4 5

• 1The picture, at the pixel of the analysis, without its caption, with the layers on top.
• 2← Previous, the selector and Next →. They step through the measured pictures; after the last one you go back to the first. In the selector, pictures with a cut outline carry ⚠, and excluded ones stay in the list with “— excluded” after the name.
• 3The layers, each with its check box and its color.
• 4The data of the picture: name, group and grain, the six measures, the cut outline, the threshold and the flags.
• 5Exclude this picture. It takes the picture out of the study; on an excluded picture, the same button reads Include again (see below).

The layers

Directions with a cut outline are always marked with small white squares on the edge, whatever layer is on. The data on the right end with the Cut outline as a percentage, the Threshold (gray) chosen by Otsu (between 89 and 91 in the pictures of the example; 90.0 in Figure 3.4) and the Flags: a check mark if the picture is fine, cut outline or enlarged if not.

Excluding a picture

Exclude this picture takes the picture out of the study: its measure is erased, it disappears from the table and the figures, and the comparison of Block 6 is erased so that you compute it again. Use it with a picture that is truly useless—a broken grain, two grains stuck together, a blurred picture—not with one that gives a number you do not like. With the example, excluding 03_C.png brings the Pictures measured tile down to 29, but Grains stays at 24, because grain 3 of C keeps its other picture.

The excluded picture does not vanish from the viewer: it stays in the selector as “03_C.png — excluded”, and you can view it with its layers to decide; in the data on the right, instead of the measures, it reads Status: “Excluded: it does not enter the analysis.”. The button changes to Include again, which measures it again on the spot, with the same settings as the others, and reports “Picture included and measured”; the others need not be measured again. There is one exception: the common pixel (Section 3.1) depends on the included pictures, and if excluding or including one shifts it, the notice says “The common pixel of the analysis changed: measure every picture again.”; then press Measure every one-grain picture again. You can also exclude or include pictures in Block 2, with the Exclude column of the table of images.

The table of every picture

Below the viewer, the drop-down Table of every picture has one row per picture with Picture, Group, Grain, the six measures, Cut (%) and Flags. The rows of pictures left out of the comparisons are highlighted; a click on a row brings that picture into the viewer. Its CSV button downloads palynopro_measurements_by_picture.csv, with the 19 variables of the app (the ornamentation ones too) and a usable column that is 1 or 0.

If you opened a saved project without opening the folder of images again, the measures are still there but the viewer says “Open the folder again to see the picture.”: the project stores references, not the pictures (Section I.5).

4.1Spines, lumina and texture

Block 4 has no button to measure: the ornamentation is measured in the same pass as the outline, when you press Measure every one-grain picture in Block 3. If you come here before measuring, the block shows only a box that explains this and a button to go and measure.

The subtitle of the block is “Spines on the outline, lumina of the reticulum and texture of the surface.” The drop-down How it is measured explains the three groups of measures in one paragraph each; here they are in more detail.

The spines of the outline

The spines come from the radial profile of Block 3 (Section 3.1): the height of each direction above the body is the difference between the outline and the body. A peak of that difference is a spine if it stands out more than the Least height of a spine (µm) and is narrower than the Largest base of a spine (µm); two peaks closer than half a base count as one, the taller. Spines that fall in directions with a cut outline, or next to them, are not counted.

The height of each spine is measured from the tip to the chord joining its two feet: the foot on each side is the lowest point of the outline between the tip and the place where the outline meets the body again. Measuring it this way, and not over the trimmed body, matters in elongated grains: there the body slopes under the spine, and the chord follows that slope. Five measures come from the spines:

The lumina of the reticulum

In a reticulate grain, the lumina—the cells of the reticulum—look darker than the muri that surround them. The app looks for lumina in the inner part of the body, away from its edge (it leaves out the outer 18% of the radius, where the surface is seen edge-on), with a local threshold along the lines of Niblack (1986): a pixel is lumen if it is darker than the mean of its neighborhood by more than a certain number of standard deviations of that neighborhood. Each cluster of dark pixels that does not touch the border of that zone, between 0.05 and 12 µm², is a lumen; larger dark patches, such as the apertures, are left out by their size. Three measures come from them:

• Lumen diameter (median) (µm): the median of the equivalent diameters of the lumina in the picture. The median is used because a few merged or split lumina do not move it.
• Lumen density (/100 µm²): lumina per 100 µm² of surface, counted only in the zone lying more than three median diameters from the border, so that cells cut by the border do not lower it.
• Lumen variation (CV): the coefficient of variation of the diameters: low in a regular reticulum, high in one with uneven meshes.

The texture of the surface

Texture is measured in the same inner zone with the gray-level co-occurrence matrix (Haralick et al. 1973): how often one gray level appears next to another at a given distance. The app reduces the grays to 32 levels between the 1st and 99th percentiles of the zone, counts the pairs at 0.2 and 0.4 µm in four directions (0°, 45°, 90° and 135°), only when both pixels lie on the surface, and averages the eight matrices.

The app also computes Texture: energy, which rises on uniform surfaces; it does not appear in the tables of this block, but it is among the variables of Block 6 and in the CSV of every picture.

4.2Lumina: the two settings

The Lumina card holds the two parameters of the local threshold. Those of the spines are in Block 3, because they also decide where the body is.

The values are applied with Measure again with these values, which goes to Block 3 and measures every picture again (with the example, some 20 to 30 s). Like any new measurement, it erases the comparison of Block 6.

The note at the foot of the card tells how it was validated: “Validated on simulated reticula: the median lumen diameter came out proportional to the drawn mesh over the whole range from 1.8 to 4 µm, with less than 3% variation among replicates.” “Proportional” is not “equal”: Section 4.3 explains why.

4.3Results by group

The Results card begins with four tiles, all of them means over the grains with a complete outline, all groups together: Spines on the outline (per grain (mean)), Spine height, Spine density (per 100 µm of outline) and Lumen (median), with the lumen density below.

Next comes the By group table, like that of Block 3: one group per column, with its number of grains below the name, and one measure per row, with the mean ± standard deviation among grains. The twelve measures fall into three sections: Spines (the five of Table 4.1), Lumina (the three lumen measures) and Texture (four of the five texture measures). Its CSV button downloads palynopro_ornamentation_by_group.csv, with one row per group.

The four figures

Three box plots, with the grains as points—Spine density by group, Lumen diameter by group and Spine height by group—and a scatter plot, Texture: contrast against homogeneity, with an ellipse enclosing 95% of each group.

4.4Check of every picture and radial profile

As in Block 3, the Check of every picture card draws over the image what was measured, with ← Previous, the selector and Next → to step through the pictures. It has three layers: Lumina (in green, each measured lumen), Spines (the tips, in red) and Body (in light blue, dashed). On the right are the picture, its twelve ornamentation measures and the Lumina measured. There is no button to exclude here: pictures are excluded in Block 3.

What to look for in the Lumina layer: that the green falls in the cells and not on the muri; that no two cells merge into one large patch; and that no light cells are left unmarked. Missing lumina at the rim of the grain are expected: that zone is not analyzed.

The radial profile

Below the viewer, Radial profile of the chosen picture shows the outline as the app reads it: the Direction from the center (°), from 0 to 360, on the horizontal axis, and the Radius (µm) on the vertical one. The thin line is the Outline, with the spines as peaks; the thick dashed one, the Body left after trimming them; the circles, the counted Spines. It is the most direct way to see why a peak was or was not counted.

5.1How it is estimated

The block works only with the pictures that Block 2 has as Overview: shots at low magnification with many grains. Its subtitle sums it up: “Turgid and collapsed grains in the overviews, by group.” The drop-down How it is estimated explains the method; these are its steps.

1. The whole picture. Each overview is analyzed at its own pixel (in the example, 150× and 1280 px give 0.6667 µm/px), without the caption; if it is more than 1600 px wide, it is reduced to that width.
2. Every grain. Otsu’s threshold separates the grains from the background, as in Block 3, and each region whose diameter falls between the Least diameter of a grain (µm) and the Largest diameter of a grain (µm) is a candidate.
3. Those on the edge, out. A grain that touches the edge of the picture has an incomplete silhouette and its shape cannot be judged: it is flagged and not counted.
4. The body of each grain. Each candidate’s own radial profile is read (in 360 directions) and its spines are removed with the same Largest base of a spine (µm) as in Block 3. Spines lower the convexity of a healthy grain; without them, only a folded grain loses it.
5. The typical grain. It is the median body diameter of the grains that do not touch the edge of that overview. If there are fewer than three, the median body measured in Block 3 is used.
6. Fragments and aggregates, out. A region smaller than the typical grain by the factor of Fragment: smaller than the typical grain (×) is a fragment or debris of the mount; one larger by the factor of Aggregate: larger than the typical grain (×), several touching grains. Neither is classified.
7. Turgid or collapsed. The remaining grains are classified by the solidity of their body: if it reaches the Least solidity of the body of a turgid grain, the grain is turgid; if not, collapsed or misshapen.

5.2Count

The Count card has five fields and the button Count in every overview. Next to it, a line tells how many overviews there are; if there are none, it says “There are no overviews: mark them as “Overview” in Block 2.” and the button is disabled.

Fit the minimum and the maximum to the pollen you study: they must keep all the grains in and the debris of the mount out. When you press the button, the waiting window tells which overview is being counted; with the three of the example, it takes a few seconds. If an overview cannot be read, an amber notice names it.

5.3Results: proportion, interval and χ²

The Results card begins with four tiles:

• Overviews: how many were counted.
• Grains counted: the turgid plus the collapsed; below, how many were left out (edge, aggregates, fragments).
• Collapsed: the percentage over all groups together and, below, the count.
• χ² test: the statistic, its degrees of freedom (df) and its p value.

The verdicts

Below the tiles come up to three notices. The first gives the conclusion of the test: “The proportion of collapsed grains differs between groups (χ²).” if p < 0.05, or “No difference between groups was detected in the proportion of collapsed grains (χ²).” if not. The second appears when any expected frequency is below 5—with the example, “3 cell(s) with expected frequency below 5: the χ² test is approximate; count more overviews.”. The third is always there: “Remember: solidity is a morphological proxy for fertility, not a viability test.”

The table by group

One row per group with Overviews, Grains (those classified), Collapsed, the %, the 95% CI (Wilson) and the Diameter (µm) of the classified grains, mean ± standard deviation. The CSV button downloads palynopro_fertility.csv, with the proportion and the limits of the interval as fractions. With the example, the table is the one in Figure 5.2: B has 5 collapsed out of 13 (38.5%, with an interval from 17.7 to 64.5%), and A and C, 0 out of 13 (from 0.0 to 22.8%).

The diameter in this table does not replace that of Block 3: it is measured at low magnification (0.67 µm/px in the example) and includes the collapsed grains, which are smaller; that is why the deviation of B is larger. Besides, in the example the grains of the overviews are drawn almost round, without the elongation of each morphotype, and that of C comes out larger than in the one-grain pictures. Use it only to check that pollen of the expected size was counted.

The two figures

Collapsed grains by group gives the percentage of collapsed grains in each group as bars, with the Wilson interval as a whisker. Solidity of the body of the grains is a histogram of the solidity of all classified grains, between 0.75 and 1, with a vertical line at the threshold: the grains to its left are the collapsed ones.

5.4Check of every overview

The Check of every overview card draws a circle around each region found, in the color of its class, with its solidity written to three decimals above the grain or, if the grain is next to the top edge, below it; always inside the picture. The overviews are chosen in the selector. The color key goes below the image:

On the right are the Overview, the Group, the Complete grains (those classified), how many are Turgid and how many Collapsed, how many were On the edge, how many Aggregates and Fragments, and the Typical grain in µm against which they were judged. In the overview of morphotype A, for instance: 13 complete grains, all 13 turgid, 1 on the edge and a typical grain of 60.8 µm.

Block 6 works with what Block 3 measured—the outline, the size and, in the same pass, the ornamentation of Block 4; it does not read the images again. Under the title, the app says what it does: “Can the groups be told apart by their pollen? With the grain as the unit.” It has a drop-down, What is computed, the Variables card with the button that runs the analysis and, once computed, five result cards: In short, By variable, Principal components, PERMANOVA and Pollen key: classifying a grain. Every figure in this chapter comes from the simulated example measured with the default values and compared with the default settings.

6.1What is computed: the grain as the unit

Before any test, the app merges the pictures of each grain into a single row: the average of its measurements. Everything else—the tests, the components, the PERMANOVA and the classifiers—is done with those grains, not with the pictures.

The What is computed drop-down sums it up in one paragraph. Step by step, this is what the app does when you press Compare the groups:

1. It takes the usable measured pictures. The included and calibrated one-grain pictures that have a result in Block 3, without those with a cut outline (those above the Acceptable cut outline (%), Chapter 3).
2. It averages by grain. Pictures with the same group and the same grain number are a single unit (in the example, 03_A.png and 03a_A.png are grain A·03). If a grain lacks any of the chosen variables, it is left out, and the app says so next to the button (Section 6.2).
3. It drops the groups with fewer than three grains and stops if fewer than two groups remain.
4. It tests each variable: one-way analysis of variance, the Kruskal–Wallis test and Tukey letters, with the p-values adjusted by Holm over all the chosen variables.
5. It standardizes the variables (mean 0, standard deviation 1) so that a measurement in micrometers does not weigh more than a proportion, and computes the principal components and the global and pairwise PERMANOVA.
6. It fits and validates two classifiers, linear discriminant analysis and random forest, leaving out one whole grain at a time.

While it computes, the “Comparing the groups” wait window says which step it is on: Tests by variable, Principal components, PERMANOVA, Discriminant analysis, Random forest (validation by grain) and Importance of the variables. With the example it takes a few seconds.

6.2Variables and settings

The Variables card decides which measurements enter the multivariate analysis and sets the three numbers that govern the random parts. Below it is the Compare the groups button.

Enter the analysis

The Enter the analysis box has one checkbox for each measurement produced by Block 3 or 4. When the app opens, eleven are checked, chosen so that the analysis does not repeat information:

Those left out are, almost all, combinations of others: elongation is the ratio of the two Feret diameters; the diameter with spines is, roughly, the body plus two spine heights; the number of spines and their spacing follow from the density and the perimeter; texture entropy and energy almost always go hand in hand with homogeneity. You can check or uncheck any you like—the app asks for at least two—and the choice is saved in the project. If you measure again in Block 3, the choice is kept, but the result is cleared and you have to compare again.

The three settings

To the right of the button, the app says how many grains and groups it will work with (“24 grains from 3 groups” in the example); if there are no measurements yet, it says “First measure the grains in Block 3.” and the button is disabled. If, on comparing, some grain was left out because it lacked a chosen variable (for example, with no lumina measured), the line adds: “… · 2 without a value in some chosen variable: left out”. This does not happen with the example.

6.3In short

Six tiles and a few verdicts that answer, at a glance, whether the groups differ and whether a single grain is enough to recognize its group.

The verdicts

Below the tiles, the app writes out in words what they mean. The rules are fixed:

• PERMANOVA. If p < 0.05: “The groups differ as a whole (PERMANOVA p = 0.001); group explains 71.1% of the variation among grains.” (the percentage is R² × 100). If not: “PERMANOVA did not detect a difference between groups as a whole.”
• Classification, with the better of the two classifiers. At 80% or more, the green check mark: “A new grain would be assigned to its correct group 100.0% of the time: pollen works as a key for these groups.” Below 80% but at least twice the chance level, a warning: classification clearly beats chance, “but is not enough to identify a single grain with confidence: the groups overlap”, and it suggests checking the confusion matrix. Below that, a note: classification barely beats chance.
• Groups left out. If any had fewer than three grains, “Groups with fewer than three grains, left out:” and their codes.
• Averaged pictures. “6 grain(s) had more than one picture: they were averaged.” in the example.

6.4By variable

A table with one row per variable: the mean and standard deviation of each group with its Tukey letter, and the three tests. Below it, the figure of the variable that best separates the groups.

The note on the card sums it up: “Mean ± standard deviation by group; groups sharing no letter differ (Tukey, α = 0.05). p-values adjusted by Holm over all variables.” The columns are:

The CSV button downloads the table as palynopro_by_variable.csv, with the numbers unrounded: one row per variable, with its internal key in variable (such as corp_d) and its name with the unit in nombre (“Body diameter (µm)”), and, for each group, the columns A_media, A_de and A_letra (mean, standard deviation and letter); then F, p_anova_holm, H, p_kruskal_holm and eta2.

The variable that best separates the groups

The box plot on the card shows the variable with the largest F, with each grain as a point, the box from the first to the third quartile with the median, the whiskers, and the Tukey letter above each group. It changes by itself if you change the chosen variables: it is always the one with the largest F among those that entered.

6.5Principal components

A two-dimensional map of the grains that summarizes all the variables at once, with arrows that show where each one grows. It is for seeing the structure—which groups separate, which overlap, which variables go together—not for testing anything.

Biplot: grains and variables

Each point is a grain, with the shape and color of its group, and each group has its 95% ellipse. The arrows are the variables: the tip sits at the correlation of the variable with PC1 (horizontal) and with PC2 (vertical), so they point where the variable grows and their length tells how well the plane represents it. Two arrows close together are variables that go hand in hand; opposite ones, that one rises when the other falls. The axes carry the percentage of variance of each component: “PC1 (58.8%)” and “PC2 (23.4%)”.

Each arrow carries the name of its variable, in the language of the interface (without the unit). In the example, the arrows of the reticulum and the texture are spread along PC1: Texture: contrast and Lumen density to the right and slightly down, Lumen diameter (median) and Texture: homogeneity to the left, right where the grains of morphotype C are. Those of size (Body major diameter (Feret), Body diameter, Body minor diameter (Feret)) and Mean spine height rise toward the right, toward the grains of B. Those of A lie at the lower right: they are the smallest and those with the densest lumina.

Variance explained

One bar per component with the percentage of the total variation it explains. It is the scree plot: when the bars stop dropping sharply, the components that follow are noise. In the example, PC1 explains 58.8%, PC2 23.4% and PC3 11.8% (eigenvalues 6.47, 2.57 and 1.29): the first three, the only ones with an eigenvalue greater than 1, add up to almost 94%, and from PC6 on the bars can hardly be seen.

The table of correlations

Under the figures, “Correlation of each variable with the components.”: one row per variable with its correlation (r) with PC1, PC2 and PC3. An |r| close to 1 says that the component is, to a large extent, that variable. In the example, PC1 is mostly texture and reticulum (contrast 0.925, homogeneity −0.917, lumen diameter −0.858, lumen density 0.813); PC2, size (major Feret 0.863, body diameter 0.739); and PC3, the spines (spine density 0.668, spine area 0.590).

6.6PERMANOVA

The test of whether the groups differ in the whole set of their measurements, without assuming that the variables follow a normal distribution: the global answer and, with more than two groups, pair by pair.

The Global line gives the pseudo-F, its degrees of freedom (number of groups − 1 and number of grains − number of groups), R², the p-value, the number of permutations and the distance used. In the example: “pseudo-F = 25.832, df 2, 21; R² = 0.711; p = 0.001 (999 permutations, Euclidean distance on standardized variables)”. The p of 0.001 is the smallest that can be obtained with 999 permutations: no shuffle matched the observed value. To claim a smaller p you have to increase the permutations.

The pairwise table repeats the test for each pair of groups—Pair, F, R², p—and adds p (Holm), adjusted for the number of pairs. Pairs with a Holm p below 0.05 carry a green stripe on the left. With only two groups there is no pairwise table: the global test already is the comparison.

6.7Pollen key: classifying a grain

The last card answers the practical question: if you were given a single grain, could you tell which group it belongs to from its measurements alone? To find out, it fits two classifiers and measures their accuracy with grains they did not see.

The tiles

• Linear discriminant: 100.0% accuracy, “leave-one-grain-out validation”.
• Random forest: 100.0%, with the same validation: “leave-one-grain-out validation, 300 trees in each model”.
• Out-of-bag error: 4.2%, of the “whole forest, 500 trees”. Each tree is fitted with a sample of the grains drawn with replacement; those that did not enter that sample (“out of the bag”) serve to measure its error. It is a second estimate of the accuracy of the forest, with no separate validation. In the example, 4.2% is one grain out of 24; that it does not fully match the validation is normal with so few grains.

Discriminant space (LD1 × LD2)

The grains on the first two discriminant axes, with the 95% ellipse of each group. Each axis carries the percentage of the separation between groups that it captures: “LD1 (78.4%)” and “LD2 (21.6%)” in the example. Unlike the biplot, these axes are chosen to separate the groups, so the figure shows the best separation possible with linear combinations; with three groups there are two axes. The grains are drawn with the model fitted to all the grains; the accuracy in the tiles, in contrast, comes from the validation. With two groups there is a single axis, and instead of the figure the app says “With two groups there is a single discriminant axis.”

Importance of the variables (random forest)

One bar per variable with the mean decrease in Gini impurity: how much, on average, each variable helps make the splits of the trees of the whole forest purer. The bars are horizontal, with the name of each variable on the left, from the most important at the top to the least important at the bottom. In the example the most important is Mean spine height (0.110), followed by Lumen diameter (median) (0.081), Lumen density (0.077) and Spine area / body area (0.075); the least, Spine density (0.020). Importance tells what the forest uses, not which variable differs most: two highly correlated variables share the importance between them.

The confusion matrices

One per classifier, Linear discriminant and Random forest, with the caption “Rows: true group · columns: assigned group”. The diagonal holds the correct assignments, shaded with the primary color; off it, the errors, in the accent color. The last column, Correct, is the percentage of grains of each group assigned correctly. In the example, both matrices have 8, 8 and 8 on the diagonal and zeros off it (Correct 100% in each group). When a classifier makes mistakes with your data, look at which groups the errors fall between and compare it with the pairwise PERMANOVA: you would expect the most similar pairs to be confused (in the example, A–B).

Block 7 has two cards: Details of the report, with the fields and the buttons, and Preview, which appears when you make the report. Under the title, the app sums it up: “Methods, results, figures and gallery in one document to print or save as PDF.” Nothing is computed here for the first time: the report takes what Blocks 2 to 6 already computed, with the settings you left in each one. The screenshots in this chapter were taken with the simulated example measured, counted and compared, with the title “Pollen morphometry of the simulated example”.

7.1Details of the report

Four text fields for the cover of the report, the language it is written in, and a list that says which parts of the study are already computed.

The four texts and the language of the report are saved in the project and return to their fields when you open it; texts you leave empty do not appear on the cover. Without a project, every time you open the app the language starts at That of the screen.

Below, four badges say what is there: Calibrated images (Block 2), Outline and size (Block 3), Fertility (Block 5) and Multivariate statistics (Block 6). Each carries a green check mark if its part is computed, or a dot if not. If anything is missing, the app adds: “What is missing appears in the report as "not computed"; you can make it anyway.” This is normal when the study has no overviews: the report comes out without fertility.

7.2Make, print and download

Five buttons: one assembles the report, three take it out of the app, and the last one saves the project.

The three output buttons are disabled until you make the report. Under the buttons, a note recalls two things: “Figures come out with the colors of the light theme, whatever that of the screen. In the print dialog choose "Save as PDF" to get the file; the text stays vector.”

7.3What the report contains

A cover and nine numbered sections, always in the same order. Whatever is not computed appears as “Not computed.”; nothing is made up.

The cover

At the top, the heading “Pollen morphometry report” and the title. Below, the material, the author and the institution if you wrote them, the date of the day (such as “November 2, 2026”) and, if the study is the example, the simulated-data note. Four figures close the cover: the images analyzed, the groups, the grains measured and, if there are statistics, the accuracy of the classification by grain (the better of the two classifiers, without decimals). In the example: 33 images, 3 groups, 24 grains measured and 100%.

The methods

Section 2 is the most useful part of the report: a draft of the methods that writes down what was really done, with the numbers you used and each method cited to its source. It has up to six paragraphs:

• Images and calibration. How many micrographs (one-grain and overviews) and groups; the pixel-size rule with the width of the reference frame; in how many images with a bar its length gave a round value with an error below 4%; the cropping of the caption and the area-averaging resampling to the common pixel (Wolberg 1990). In the example: “… in 33 of the 33 images with a bar … to a common pixel of 0.1 µm”.
• Outline and size. Otsu’s threshold (1979), the 720 directions, the opening of the profile (Serra 1982) with the largest base of a spine, the Feret diameters by rotating calipers (Shamos 1978) on the convex hull (Andrew 1979), the cut-outline limit and the grains that resulted from averaging the pictures (24).
• Ornamentation. The least height of a spine, the lumina by local threshold (Niblack 1986) with the radius and the standard deviations you set, and the texture by gray-level co-occurrence (Haralick et al. 1973).
• Fertility, only if you counted overviews: the minimum and maximum diameters, the solidity threshold, the Wilson interval (1927) and the χ² test.
• Statistics, only if you compared the groups: analysis of variance, Kruskal and Wallis (1952), Holm (1979) over the number of variables, Tukey (1949), principal components (Pearson 1901; Hotelling 1933), PERMANOVA (Anderson 2001) with its permutations, linear discriminant analysis (Fisher 1936), random forest (Breiman 2001) with its trees—“500 trees, 300 in each model of the validation” with the default values—, the leave-one-grain-out validation and Cohen’s κ (1960). If some grain was left out for lacking a value in any of the variables, the paragraph says so at the end (“… grains without a value in some of the variables were left out.”).
• A closing sentence with the version of the program: “All computations were done with PalynoPro 1.0.0 (Barrera-Guzmán 2026)…”.

The gallery and the limitations

The gallery takes, from each group, the first one-grain picture that is included, measured, free of warnings (cut outline or enlarged) and open in the session, and draws it with the outline with spines, the body and the spine tips, as in the review of Block 3. In the example they are 01_A.png (body 59.6 µm), 01_B.png (72.0 µm) and 01_C.png (64.2 µm). If you opened a project without opening the folder of the pictures again, it says instead “The images are not open in this session.”

The limitations are fixed, because they belong to the method, not to your data: the body and the spines are measured on the two-dimensional silhouette; the Feret diameters amount to the polar and equatorial axes only in equatorial view; spine height, measured from the tip to the chord joining its two feet, comes out about one pixel short, because the finest tip of the cone does not fill a pixel (with simulated grains at 0.1 µm per pixel, by 3 to 6%, median 4%; more for short spines or spines touching at the base); fertility by solidity is a morphological proxy, not a viability test; and images enlarged to the common pixel give less reliable fine measurements.

7.4The figures, the studio and the editor

Every figure of Blocks 3 to 6 has the same four tools in its header: two quick downloads and two ways of editing it.

The SVG and PNG buttons save files named after the title of the figure, without accents and with underscores (Biplot_grains_and_variables.svg), and they come out with the colors of the light theme even if you work in the dark one: the colors, the lines and the lettering are written inside the file, which does not depend on the app to be displayed. Whatever you changed with the ✎ editor goes inside too, because the editor writes onto the figure itself.

1 2 3 4

• 1SVG. Downloads the figure as a vector: it can be enlarged without losing sharpness and edited in any drawing program.
• 2PNG. Downloads the figure as an image, at three times its size on screen.
• 3✎ editor. Opens the fine-editing panel of that figure.
• 4Figure studio. Opens the figure full screen to adjust and export it.

The figure studio

The button with the ruler and pencil opens the figure studio, the same in every application of the suite. You can also reach it from the search (Ctrl + K): the figures appear among the results and open directly in the studio, or from the In this block index, with the button next to the sections that have figures.

The top bar undoes and redoes, compares before and after with a curtain, zooms the figure in and out (Fit makes it the size of the screen, with the zoom percentage), shows it at actual size, applies one-click Presets (and saves your own), copies the image to the clipboard and, with Export, downloads it at the width of one column, one and a half columns, two columns or a free width, in millimeters or inches, at 300, 600 or 1200 dpi, as PNG, SVG, vector PDF or TIFF, with a white or transparent background. The buttons on the right change the place of the inspector (right, left or floating) and show the shortcuts; Close studio returns to the app with the figure where it was.

The inspector brings together, in collapsible sections, Figure settings (the figures of PalynoPro have no options of their own, and the studio says so), Fine editing of this figure (the app’s ✎), which is the editor below docked, and the studio’s own tools: scientific palettes with color-blindness simulation (Palette and color), text size and line width tuned to the output size, and the legend (inside the plot in eight places, outside, below or hidden; it can also be dragged). Search a control… finds any setting by its name. Double-clicking a text in the figure lets you change it right there. The studio remembers its settings in this browser.

The ✎ editor

The ✎ button opens, without leaving the page, a floating panel that works on that figure, titled “Edit this figure” with the name of the figure. It has six tabs:

• General: title, subtitle and footnote inside the figure, with their size and alignment; the Font family, the Size of every text, the Color of every text, the Line weight, the Point size, the Background and a Border around the whole figure.
• Axes & grid: the axis titles, their numbers, and the color and width of the axes and the grid.
• Series: each color of the figure with the name of its group: color, width, dash, opacity or hiding it.
• Legend: showing or hiding it, its place and its form.
• Texts: each text one by one (wording, size, bold, italic, color or hiding it); this is where you shorten, for example, a long variable name in the biplot.
• Annotations: notes, arrows, reference lines, shaded bands and the panel letter.

At the foot of the panel, Undo everything on this figure, Copy the style to every figure and Export. The edits are saved figure by figure in this browser, not in the project, and are applied again every time the figure is redrawn—on recomputing, on changing the language or the theme—: texts are recognized by their original wording and colors by their original value. If you open the project on another computer, the figures come out without your edits.

7.5Saving and resuming the project

The project is a small .json file with everything needed to see the results again and repeat the analysis, except the pictures.

It is saved with Ctrl + S from any block or with Save the project in Block 7. The file is named after the title of the report followed by .palynopro.json (project.palynopro.json if there is no title); the app confirms “Project saved:” with the name and reminds you that it does not carry the results of Block 6 or the ✎ edits. It is opened with Ctrl + O or with Open a project on the home page; the app confirms “Project opened:”, reminds you that “Compare the groups” has to be repeated, and takes you to Block 2.

On opening a project, the tables and the figures of Blocks 2 to 5 come back. To see the pictures again, review them on the image or have the gallery of the report, open the same folder in Block 2: the app joins each picture to its row by its name and keeps its groups, calibration, exclusions and results (“… of the project opened again”). Only if the folder brings images that were not in the set does it ask first, because adding them clears the results and you have to measure, count and compare again (Appendix D). Without opening the folder you can consult the tables and the figures; the viewers will say “Open the folder again to see the picture.” With the simulated example, which the app draws, the pictures come back by themselves.

AThe files the app reads and writes

PalynoPro runs in the browser and sends nothing to any server: files come in from your computer and what the app writes goes to your downloads folder. The pictures are never copied: the app keeps a reference to each one while the tab is open.

A.1 · What the app reads

Everything comes in through Block 2—Choose a folder, Choose files or by dropping onto the Drop a folder or your images here area—except the project. The app recognizes each file by its extension and its content; there is no need to tell it what each one is.

The TIFF

The reader accepts TIFF in strips or in tiles; uncompressed, with LZW (with or without horizontal predictor), PackBits or Deflate; at 1, 8, 16 or 32 bits; in grayscale, with a palette or in color. It does not accept TIFF with the channels in separate planes or other compressions (JPEG inside the TIFF, for example): those pictures are reported as not read. 16-bit images are brought to 8 bits for analysis.

The microscope file

The app recognizes these ways of writing the data, in the .txt or inside the TIFF:

If the file has none of this, the magnification is typed by hand in the table of Block 2 or given in the index.

The index

The first line is the header. The separator is guessed from it: semicolon if it has more semicolons than commas; otherwise, tab if it has any; otherwise, comma. Quoted fields are respected, the byte order mark at the start is ignored and empty lines are skipped. Headers are compared ignoring case, accents and symbols (Magnificación and magnificacion are the same). Each column is recognized by any of these names:

Each row is joined to its image by the file name, with or without extension. What the index brings prevails over what is read from the name (group and grain), also if you later change the name pattern (B.1); the magnification of the index is used only if the microscope file did not give it, and a pixel size in the index takes precedence over the magnification (B.1). Decimals may use a point or a comma (0,1), but with a decimal comma separate the columns with semicolons or tabs. The index that Block 2 itself downloads comes back in complete:

archivo;grupo;nombre;grano;magnificacion;excluir;nota
01_G1.tif;G1;Group one;01;1000;;
02_G1.tif;G1;Group one;02;1000;1;broken grain
90_G1.tif;G1;Group one;90;150;;overview

The group key

A text file with one line per group, A = Morphotype A, G1 = Group one or G1: Group one. The code is a single word of letters and digits, 1 to 8 characters long, which may contain hyphens or underscores (- or _) inside: A, G1, M12, P3b, Lin-2. A group that is not in the key is named by hand in the Groups card of Block 2 or with the name column in the index.

Group and grain from the name

Without an index, the group and the grain come from the file name with the Pattern of the name (group and grain) of Block 2, a regular expression with two named parts: (?<grain>…) and (?<group>…). The default, ^(?<grain>\d+)[A-Za-z]*_(?<group>[A-Za-z]+\d*), reads 07a_G1.tif as grain 07 of group G1: the letters after the grain number mark a repeated picture of the same grain. If the pattern has no names, the first and second parentheses are used as grain and group.

A.2 · What the app writes

The names that come from a title are written without accents and with underscores (Pollen_morphometry_of_the_simulated_example.html). All the CSV files use the comma as separator and the decimal point, with unrounded numbers, and are in UTF-8 with a byte order mark so that a spreadsheet reads the accents and the µ. The column headers are the same in both languages, and the measurements go with their internal key (corp_d, sp_h, lum_d…), which does not change with the language; the per-variable CSV of Block 6 also carries the name of each one.

BThe rules, gathered

The rules the app applies to calibrate, measure, classify, warn or decide, with the block where they act and their default values. They are the same ones each chapter explains; here they are together to be consulted at a glance.

B.1 · Calibration (Block 2)

B.2 · Outline, size and ornamentation (Blocks 3 and 4)

B.3 · Fertility (Block 5)

B.4 · Multivariate statistics (Block 6)

All of them are guides for reading the results, not laws: what you know about your material, the preparation and how it was photographed always comes first.

CGlossary

The terms of palynology, image analysis and statistics used in the manual, in the sense in which the app uses them. Pollen terminology follows Punt et al. (2007) and Halbritter et al. (2018).

DWhen something does not work

The messages of the app, as they appear, with what they usually mean and what to do. The ellipses are the part of the message that changes (a name, a number).

If the app goes blank or a card does not respond, reload the page: the data are lost, but the saved project is not. If the problem persists, open the app with Open PalynoPro.bat instead of double-clicking index.html; some browsers limit what a page opened as a file can do.

EThe tool: license and how to cite it

E.1 · What it is and what it is not

PalynoPro is written for this program: the TIFF reading with its compressions, the image analysis, the statistics, the multivariate analysis, the figures, the report and the illustrations are in its own code, which can be read. It uses no third-party libraries, ships no fonts (it uses those of the computer), has no minified code or build steps and loads nothing from the network: that is why it opens with a double click and offline. PNG and JPEG images are decoded by the browser itself. The only outbound link, the one to the portal of the suite, is followed only if you click it.

The common modules of the suite—the block navigator, the figure studio and the ✎ editor—are by the same author and come in their variant with original icons, drawn for the suite. The example is simulated: the grains, their spines, their reticulum, the noise and the scale bar are drawn by code with fixed seeds every time it is opened, and they are not pictures of any plant. It implements published methods, each cited to its original source in the code, on the home page of the app and in the methods of the report (Appendix F). The file PROCEDENCIA.md, in the folder of the program, says where each part comes from.

E.2 · License

PalynoPro is distributed under the GNU General Public License, version 3 or later (GPL-3.0-or-later). The full text is in the LICENSE file, in the folder of the program.

E.3 · How to cite it

The How to cite section on the home page of the app gives the citation ready to copy, with Copy the citation. Both citations below carry the Zenodo concept DOI, which always resolves to the latest version (version 1.0.0 has its own: 10.5281/zenodo.23883214).

In your methods, also cite the works of the methods you used. The methods of the report (Section 7.3) already cite them by author and year, and its references section gives the full entries.

FReferences

The 22 works in the catalog of the app, grouped by topic—palynology, image analysis and statistics—and, within each topic, in alphabetical order. They are the same as in the References section of the home page and of the report.

F.1 · Palynology and pollen terminology

• Erdtman G. 1952. Pollen morphology and plant taxonomy. Angiosperms. Almqvist and Wiksell, Stockholm.
• Halbritter H, Ulrich S, Grímsson F, Weber M, Zetter R, Hesse M, Buchner R, Svojtka M, Frosch-Radivo A. 2018. Illustrated pollen terminology. 2nd ed. Springer, Cham.
• Punt W, Hoen PP, Blackmore S, Nilsson S, Le Thomas A. 2007. Glossary of pollen and spore terminology. Review of Palaeobotany and Palynology 143: 1–81.

F.2 · Image analysis

• Andrew AM. 1979. Another efficient algorithm for convex hulls in two dimensions. Information Processing Letters 9: 216–219.
• Borgefors G. 1986. Distance transformations in digital images. Computer Vision, Graphics, and Image Processing 34: 344–371.
• Haralick RM, Shanmugam K, Dinstein I. 1973. Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics 3: 610–621.
• Niblack W. 1986. An introduction to digital image processing. Prentice Hall, Englewood Cliffs.
• Otsu N. 1979. A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics 9: 62–66.
• Serra J. 1982. Image analysis and mathematical morphology. Academic Press, London.
• Shamos MI. 1978. Computational geometry. PhD thesis, Yale University, New Haven.
• Wolberg G. 1990. Digital image warping. IEEE Computer Society Press, Los Alamitos.

F.3 · Statistics, multivariate analysis and classification

• Anderson MJ. 2001. A new method for non-parametric multivariate analysis of variance. Austral Ecology 26: 32–46.
• Breiman L. 2001. Random forests. Machine Learning 45: 5–32.
• Cohen J. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement 20: 37–46.
• Fisher RA. 1936. The use of multiple measurements in taxonomic problems. Annals of Eugenics 7: 179–188.
• Holm S. 1979. A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics 6: 65–70.
• Hotelling H. 1933. Analysis of a complex of statistical variables into principal components. Journal of Educational Psychology 24: 417–441.
• Kruskal WH, Wallis WA. 1952. Use of ranks in one-criterion variance analysis. Journal of the American Statistical Association 47: 583–621.
• McArdle BH, Anderson MJ. 2001. Fitting multivariate models to community data: a comment on distance-based redundancy analysis. Ecology 82: 290–297.
• Pearson K. 1901. On lines and planes of closest fit to systems of points in space. Philosophical Magazine 2: 559–572.
• Tukey JW. 1949. Comparing individual means in the analysis of variance. Biometrics 5: 99–114.
• Wilson EB. 1927. Probable inference, the law of succession, and statistical inference. Journal of the American Statistical Association 22: 209–212.