--- title: "Visualization tools" author: "Elio Campitelli" date: "`r Sys.Date()`" output: markdown::html_format vignette: > %\VignetteIndexEntry{Visualization tools} %\VignetteEngine{knitr::knitr} %\VignetteEncoding{UTF-8} --- ```{r setup, include = FALSE} knitr::opts_chunk$set( dev = "jpeg", collapse = TRUE, message = FALSE, fig.width = 7, comment = "#>") data.table::setDTthreads(1) ``` One of the "conceptual branches" of `metR` is the *visualization tools*. These are a set of functions that interface with [ggplot2](https://ggplot2.tidyverse.org/index.html) for easier and better plotting of meteorological (an other) fields. ```{r, message = FALSE} # Packages and data use throught library(metR) library(ggplot2) library(data.table) temperature <- copy(temperature) temperature[, air.z := Anomaly(air), by = .(lat, lev)] ``` # Geoms and stats ## geom_contour_fill() This geom has a bit of a long story. Before ggplot2 introduced `geom_contour_filled()`, the 'canonical' way to get filled contours was by using `stat_contour()` with a `polygon` geom and mapping `fill` to `level` (see [this](https://github.com/tidyverse/ggplot2/issues/1066) issue). This was ugly and bad and didn't work in most cases, so I threw together some hacks to make it work, and thus `metR::geom_contour_fill()` was born. Then, Claus Wilke created {[isoband](https://github.com/r-lib/isoband)}, which lead to ggplot2 finally getting a proper implementation of filled contours. Since the implementation worked much better than my Frankenstein's monster hacks and workarounds, I decided to scrub it and adopt {isoband} as well. So after all this history the R landscape is left with `metR::geom_contour_fill()` and `ggplot2::geom_contour_filled()` with almost identical implementations. Why should you use one or the other? Why didn't I just deprecate this seemingly redundant function? Because, in the intervening time `metR::geom_contour_fill()` has evolved and it now provides many functionality that is missing from ggplot2's counterpart. ### `breaks` can be a function. The main control parameter of the contour family of functions it their `breaks` argument, which sets the levels at which to draw contour lines. In `geom_contour_fill()` this argument can be a function that takes a binwidth and the range of the data. This allows for great flexibility. For example, is is very common in meteorological science to not draw contours at the zero level. Passing `MakeBreaks()` as the `breaks` argument allows you to control not only the binwidth, but also which levels **not** to draw. ```{r} ggplot(temperature[lon %~% 180], aes(lat, lev, z = air.z)) + geom_contour_fill(breaks = MakeBreaks(binwidth = 2, exclude = 0)) + scale_fill_divergent(breaks = MakeBreaks(binwidth = 2, exclude = 0)) ``` ### Infilling of missing values Both `geom_contour_fill()` and `geom_contour_filled()` will work fine with missing values, but they will leave a somewhat ugly empty patch when they are found. ```{r} data(volcano) volcano <- setDT(reshape2::melt(volcano)) volcano[, value.gap := value] volcano[(Var1 - 40)^2 + (Var2 - 35)^2 < 50, value.gap := NA] ggplot(volcano, aes(Var1, Var2, z = value.gap)) + geom_contour_fill() ``` This is an OK behaviour, but sometimes it's more visually appealing to fill in these missing areas. The `na.fill` argument allows for some control over how this values are imputed. It can take a number, a function function that takes the vector of values and returns one number, or `TRUE`, in which case it will use `metR::Impute2D()` to interpolate the missing values. ```{r} ggplot(volcano, aes(Var1, Var2, z = value.gap)) + geom_contour_fill(na.fill = TRUE) ``` ### On-the-fly projecting Contouring only works on regular grids (although see {[contoureR](https://cran.r-project.org/package=contoureR)} for an implementation that accepts non-regular girds). Sometimes the data is regular in some coordinate system and you need to plot in another. For example, this data is not regular on a a longitude by latitude coordinate system, but it is regular on the x y coordinates which is a lambertian grid. ```{r} data(surface) ggplot(surface, aes(lon, lat)) + geom_point(aes(color = height)) ggplot(surface, aes(x, y)) + geom_point(aes(color = height)) ``` If you know the projection string, you can pass it to the `proj` argument and get filled contours in longitude by latitude space (this needs {[proj4](https://cran.r-project.org/package=proj4)} to be installed). ```{r} proj_string <- "+proj=lcc +lat_1=-30.9659996032715 +lat_2=-30.9659996032715 +lat_0=-30.9660034179688 +lon_0=-63.5670013427734 +a=6370000 +b=6370000" ggplot(surface, aes(x, y)) + geom_contour_fill(aes(z = height)) ggplot(surface, aes(x, y)) + geom_contour_fill(aes(z = height), proj = proj_string) ``` You can actually pass any arbitrary function to `proj`. It should take an iso object as returned by the isoband package with the contour information and return the modified contours. ### Clipping The `clip` argument takes a simple feature polygon and clips the contours to the interior of the polygon. Clipping is done after `proj` is applied. ```{r} argentina <- rnaturalearth::ne_countries(country = "argentina", returnclass = "sf") ggplot(surface, aes(x, y)) + geom_contour_fill(aes(z = height), proj = proj_string, clip = argentina) ``` ### On the fly kriging If everything fails and your data is not on a regular grid on any coordinate system, then you can use kriging to interpolate it to a regular grid and then plot it using contours. Setting the `kriging` argument to`TRUE` uses {[kriging](https://cran.r-project.org/package=kriging)} to *krige* (is that a verb?) the data before computing the contours. Here, as an example, the volcano dataset is sampled and perturbed to create a dataset of irregularly spaced points. Using kriging, `geom_contour_fill()` can recover the general shape. ```{r} set.seed(42) some_volcano <- volcano[sample(.N, .N/7)] # select 70% of the points some_volcano[, Var1 := Var1 + rnorm(.N)] # add some random noise some_volcano[, Var2 := Var2 + rnorm(.N)] ggplot(some_volcano, aes(Var1, Var2)) + geom_point(aes(color = value)) ggplot(some_volcano, aes(Var1, Var2)) + geom_contour_fill(aes(z = value), kriging = TRUE) + geom_point(size = 0.2) ``` ### Unequal `breaks` between facets ggplot2's `geom_contour_filled()` computes breaks using all the data. This is great because it means that breaks are consistent between panels when using `facet_grid()`. But if you want to compute breaks separately for each panel, you can use `geom_contour_fill()`'s `global.breaks` argument. With the default, separation between levels is always the same. Good for comparison, but in this case the panel on the left lacks detail. ```{r} ggplot(temperature[lev %in% c(1000, 300)], aes(lon, lat, z = air.z)) + geom_contour_fill() + scale_fill_divergent() + facet_grid(~lev) ``` With `global.breaks = FALSE`, binwidth is no longer the same in each panel and now you can see more detail in the leftmost panel. ```{r} ggplot(temperature[lev %in% c(1000, 300)], aes(lon, lat, z = air.z)) + geom_contour_fill(global.breaks = FALSE) + scale_fill_divergent() + facet_grid(~lev) ``` All of these features are also available to regular contours in the `geom_contour2()` function. ### Computed variables Like `ggplot::geom_contour_filled()`, `geom_contour_fill()` computes several relevant variables. The two most important ones are `level_mid` (also called `int.level` for backwards compatibility reasons) and `level`. The former (the default) is a numeric value that corresponds to the midpoint of the levels while the latter is an ordered factor that represents the range of the contour. ```{r} ggplot(temperature[lev == 300], aes(lon, lat, z = air.z)) + geom_contour_fill() ggplot(temperature[lev == 300], aes(lon, lat, z = air.z)) + geom_contour_fill(aes(fill = after_stat(level))) ``` A big difference between the two variables is that one is compatible with continuous scales while the other only works with discrete scales. The advantage of using `level_mid` is that it preserves the actual continuous nature of the data, which allows the use of meaningful continuous scales such as `scale_fill_divergent()` and results in proper colours when breaks are not equally spaced. The disadvantage is that the scale function has no information about the discretisation, and thus is very hard to use *synchronise* the breaks that define the levels of the contours and the breaks shown in the color guide. A third important computed variable is `level_d`, which is identical to `level` but forces the scale to match the breaks in the data automatically. ```{r} ggplot(temperature[lev == 300], aes(lon, lat, z = air.z)) + geom_contour_fill(aes(fill = after_stat(level_d))) ``` See the [discretised_scale](#discretised_scale) section for a way of using continuous sales for discretised data. ## geom_text_contour and geom_label_contour Labelling contours is also a problematic aspect of `ggplot2`. `geom_text_contour()` and `geom_label_contour()` can be use to automatically add text or labels to the flattest part of a contour. ```{r} ggplot(temperature[lev == 300], aes(lon, lat, z = air.z)) + geom_contour_fill() + geom_contour(color = "black") + geom_text_contour() + scale_fill_divergent() ``` By default it labels every 2nd contour (this can be changed by the `skip` parameter) and it rotates to follow the angle of the contour (this is not available on `geom_label_contour()`). Since in some datasets there can be very small contours that should not be labelled for clarity, the `min.size` argument specifies the minimum points a contour has to have in order to be labelled. Notice how labels are drawn on top of contours? The problem is that `geom_contour()` doesn't know it's being labelled; `geom_text_contour()` addresses this issue by allowing you to draw a stroke around the text. ```{r} ggplot(temperature[lev == 300], aes(lon, lat, z = air.z)) + geom_contour_fill() + geom_contour2(color = "black") + geom_text_contour(stroke = 0.2) + scale_fill_divergent() ``` You can have full control over where labels are placed with the `label.placement` argument. It takes a function that takes a vector of x and y values and returns a logical vector of the same length that indicate the points where labels should be. There are 5 build in placement functions: * `label_placement_flattest()`: The default. It computes the curvature and the angle at each points and then selects the point which minimizes the sum of both numbers. Therefore, it tries to label the flattest and most horizontal part of the contour. * `label_placement_fraction()`: It takes a vector of numbers ranging from 0 to 1 representing where tu put the label in relation to the length of the contour. That is, using 0.5 means to put the label at the midpoint. * `label_placement_n()`: Places `n` labels approximately equally spaced along the contour. * `label_placement_random()`: It puts the label in random places. ¯\\\_(ツ)\_/¯ . * `label_placement_all()`: It puts the label at every point. Probably is best not to use it. ```{r} ggplot(temperature[lev == 300], aes(lon, lat, z = air.z)) + geom_contour_fill() + geom_contour2(color = "black") + geom_text_contour(stroke = 0.2, label.placement = label_placement_random()) + scale_fill_divergent() ``` ## geom_contour_tanaka Illuminated contours (aka Tanaka contours) use varying brightness and width to create an illusion of relief. This can help distinguishing between concave and convex areas (local minimums and maximums), specially in black and white plots. It also allows for photocopy safe plots with divergent colour palettes, and it just looks cool. ```{r} ggplot(temperature[lev == 300], aes(lon, lat, z = air.z)) + geom_contour_fill() + geom_contour_tanaka() + scale_fill_divergent() ``` ## stat_subset `metR` also has `stat_subset()` which makes a subset of the data according to the `subset` aesthetic. This makes it possible to show only part of the data in one geom without needing to specify a `data` argument (specially useful if the data being plotted is the result of a long pipe and not actually assigned to a `data.frame`). It has a somewhat limited use in that it cannot perform further statistical transformations of the data. For example, it can be used if you have a correlation field and want to mark only the points with significant correlations: ```{r} data(geopotential) # geopotential height at 700hPa for the Southern Hemisphere. ggplot(geopotential[, gh.base := gh[lon == 120 & lat == -50], by = date][ , .(correlation = cor(gh.base, gh)), by = .(lon, lat)], aes(lon, lat, z = correlation)) + geom_contour_fill(breaks = MakeBreaks(0.1)) + stat_subset(aes(subset = correlation > 0.5), geom = "point", size = 0.1) + scale_fill_divergent() ``` Another possible use it to quickly mask missing values. ```{r} ggplot(volcano, aes(Var1, Var2, z = value.gap)) + geom_contour_fill(na.fill = TRUE) + stat_subset(aes(subset = is.na(value.gap)), geom = "raster", fill = "#a56de2") ``` ## geom_vector and geom_arrow Plotting arrows can be a pain. Again, the 'canonical' way of plotting vectors is to use `geom_segment()` and specify `x`, `y`, `xend` and `yend` aesthetics which can be a lot of typing when one has the data on location and displacement (or velocity). Instead, `metR`'s `geom_vector()` and `geom_arrow()` draw vectors defined by their lateral displacements (dx, dy) or their magnitude and angle. It also has some useful parameters like `min.mag`, which controls the minimum magnitude for an arrow to be drawn (useful for highlighting only areas of strong 'flow') and `skip`, which draws only the nth arrow in the x and y directions. Both geoms are essentially the same, except that `geom_arrow()` defaults to preserving direction in coordinate transformations and regardless of the plot aspect ratio. This is recommended if the arrows are meant to convey the direction of the flow in each point instead of being a description of the shape of the flow related to the plot coordinates. So, as an example, we can plot the temperature gradient like this: ```{r} temperature[, c("t.dx", "t.dy") := Derivate(air.z ~ lon + lat, cyclical = c(TRUE, FALSE)), by = lev] (g <- ggplot(temperature[lev == 500], aes(lon, lat)) + geom_contour_fill(aes(z = air.z)) + geom_vector(aes(dx = t.dx, dy = t.dy), skip.x = 2, skip.y = 1) + scale_mag()) ``` In the above example, t.dx and t.dy represent displacements in the same units as the x and y dimension, so using `geom_vector()` each arrow points perpendicular to the contours even if we change the coordinate system. ```{r} g + coord_polar() ``` But if we want to plot how the (horizontal) direction changes with height it makes more sense to use `geom_arrow()` (or `geom_vector(preserve.dir = TRUE)`) ```{r} ggplot(temperature[lon %between% c(100, 200) & lat == -50], aes(lon, lev)) + geom_arrow(aes(dx = dx(t.dx, lat), dy = dy(t.dy)), skip = 1) + scale_y_level() + scale_mag() ``` In this case, an arrow at a 45° angle represents temperature gradient from the Southwest regardless of the `x` and `y` scales. The `start` and `direction` arguments adjust the behaviour of the arrows. This is useful for working with winds in the meteorological standard (in which 0° means wind from the North and 90° means wind from the East). ## geom_streamline Streamlines are paths tangential to a vector field and provide an intuitive way of visualizing vector fields. `geom_streamline()` computes streamlines via Euler integration. ```{r} (g <- ggplot(temperature[lev == 500], aes(lon, lat)) + geom_contour_fill(aes(z = air.z)) + geom_streamline(aes(dx = t.dy, dy = -t.dx), L = 10, res = 2, arrow.length = 0.3, xwrap = c(0, 360))) ``` There are several parameters that control the feel of the result. `L` and `res` control the length and resolution, while the `skip` parameters change the amount of the streamlines. More in depth control over the simulation can be accomplished by the `S` and `dt` parameters. Since streamlines imply a tight relationship between the vector field and the spatial dimensions, it's important that their units are compatible. A wind field in m/s in a longitude-latitude grid should first be converted into degrees/s. The functions `dlon()` and `dlat()` are an easy way of doing this. This tight coupling also means that they are robust to coordinate transformations. ```{r} g + coord_polar() ``` `stat_streamline()` computes three useful variables. `step` is the integration step, and `dx` and `dy` are the components of the vector field at each point. They can be used to visualize information about the field in each point, and to give a sense of direction without arrows. ```{r} ggplot(temperature[lev == 500], aes(lon, lat)) + geom_streamline(aes(dx = t.dy, dy = -t.dx, size = after_stat(step), alpha = after_stat(step), color = after_stat(sqrt(dx^2 + dy^2))), arrow = NULL, L = 10, res = 2, xwrap = c(0, 360), lineend = "round") + scale_color_viridis_c(guide = "none") + scale_size(range = c(0, 1), guide = "none") + scale_alpha(guide = "none") ``` # Scales ## discretised_scale() The use of `geom_contour()` and `geom_contour_fill()` generates a problem because those geoms implicitly discretise a continuous variable, but it's hard to communicate that discretisation to the colour scale. In particular, it would be handy to show a discretised version of the colour scale in which the values labelled are in sync with the contour levels. One way of achieving this is to use the `level` stat, which --as discussed previously-- is a discrete variable. With that, one can use `ggplot2::guide_colorsteps()` to get a nice looking guide. ```{r} ggplot(temperature[lev == 300], aes(lon, lat, z = air.z)) + geom_contour_fill(aes(fill = after_stat(level)), breaks = c(-10, -8, -6, -2, -1, 0, 6, 8, 10)) + guides(fill = guide_colorsteps()) ``` This can work, but this example shows its limitations. Since the scale has no real understanding of the underlying numbers, breaks and colours are always distributed evenly even when they shouldn't. For the same reason it's also not possible to use diverging colour scales with a set midpoint. For that, the solution is to use `scale_fill_discretised()`: ```{r} ggplot(temperature[lev == 300], aes(lon, lat, z = air.z)) + geom_contour_fill(aes(fill = after_stat(level)), breaks = c(-10, -8, -6, -2, -1, 0, 6, 8, 10)) + scale_fill_discretised() ``` This type of scale understand discrete values that have an underlying continuous nature. This means that the scale can be modified by referencing the continuous values. For example, using a divergent scale and setting the midpoint. ```{r} ggplot(temperature[lev == 300], aes(lon, lat, z = air.z)) + geom_contour_fill(aes(fill = after_stat(level)), breaks = c(-10, -8, -6, -2, -1, 0, 6, 8, 10)) + scale_fill_divergent_discretised(midpoint = 3) ``` ## Other scales Many meteorological fields are defined in a longitude×latitude×level grid, so `metR` includes scales for each dimension. These are glorified wrappers around `scale_*_continuous()` with sensible defaults and, in the case of `scale_*_level()`, the implementation of `reverselog_trans()`. There are also `scale_color_divergent()` and `scale_fill_divergent()` which are wrappers around `scale_*_gradient2()` but with sane default colours for positive and negative values --particularly useful for plotting anomaly values. To see how this scales work, let's visualize the vertical distribution of temperature anomalies from the zonal mean. ```{r, message = FALSE} # Plot made with base ggplot (g <- ggplot(temperature[lon %~% 180], aes(lat, lev, z = air.z)) + geom_contour2(aes(color = after_stat(level)))) ``` While this is fine, since pressure levels are roughly proportional to $\mathrm{e}^{-z}$ in meteorology we usually plot the vertical coordinate as $-\log(p)$. However, while ggplot2 has `scale_y_log10()` and `scale_y_reverse()`, they don't work together. `metR` defines the new transformation `reverselog_trans()` that can be used with any scale but that is the default for `scale_*_level()`. On the other hand, `scale_*_latitude()` (and `scale_*_longitude()`) not only defaults `expand` to `c(0, 0)`, but also has a `ticks` argument that specifies the spacing of breaks between -90 and 90 in the case of `scale_*_latitude()`, and between 0 and 360 in the case of `scale_*_longitude()`. These scales default to printing no label, since usually the dimensions are understood by the shape of the plot. ```{r} g + scale_y_level() + scale_x_latitude(ticks = 15, limits = c(-90, 90)) + scale_color_divergent() ``` Note: `scale_*_longitude()` (currently) assumes the data goes from 0° to 360° but puts labels between -180° and 180°. This very idiosyncratic choice stems from the fact that model output is usually in the [0; 360) range but it's easier to read maps in the (-180; 180] range. This may change in the future. # Other goodies ## geom_relief This geom produces a shaded relief of the data. It doesn't really map the data per se, but it provides an intuitive representation of hills and valleys that can be used as a background for plotting actual data. ```{r} ggplot(volcano, aes(Var1, Var2)) + geom_relief(aes(z = value)) ```