---
title: "Spatiotemporal Visualization"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Spatiotemporal Visualization}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
```

## 3D Visualization (spatiotemporal)

{mlr3spatiotempcv} makes use of {plotly} to create the 3D plots for visualizing spatiotemporal folds.
Arranging multiple 3D plots in {plotly} is done via [3D subplots](https://plotly.com/r/3d-subplots/).

Unfortunately, {plotly}'s subplot implementation is not dynamic.
This means that multiple "scene" objects need to be specified in `plotly::layout()` to determine the coordinates of the respective subplot.
Depending on the number of chosen folds by the user in `autoplot()`, a different number of scenes with different coordinates needs to be given to align the plots properly.

Hence, manual action is needed to create a properly aligned grid of 3D plots.

Below is an example how to create a 2x2 grid showing four folds as 3D subplots.
It makes use of the returned 3D plotly objects which are returned in a list by `autoplot()`:

```{r, fig.align='center', eval=FALSE}
library(mlr3)
library(mlr3spatiotempcv)
task_st = tsk("cookfarm_mlr3")
task_st$set_col_roles("SOURCEID", roles = "space")
task_st$set_col_roles("Date", roles = "time")
resampling = rsmp("sptcv_cstf", folds = 5)

pl = autoplot(resampling, task_st, c(1, 2, 3, 4),
  crs = 4326, point_size = 3, axis_label_fontsize = 10,
  plot3D = TRUE
)

# Warnings can be ignored
pl_subplot = plotly::subplot(pl)

plotly::layout(pl_subplot,
  title = "Individual Folds",
  scene = list(
    domain = list(x = c(0, 0.5), y = c(0.5, 1)),
    aspectmode = "cube",
    camera = list(eye = list(z = 2.5))
  ),
  scene2 = list(
    domain = list(x = c(0.5, 1), y = c(0.5, 1)),
    aspectmode = "cube",
    camera = list(eye = list(z = 2.5))
  ),
  scene3 = list(
    domain = list(x = c(0, 0.5), y = c(0, 0.5)),
    aspectmode = "cube",
    camera = list(eye = list(z = 2.5))
  ),
  scene4 = list(
    domain = list(x = c(0.5, 1), y = c(0, 0.5)),
    aspectmode = "cube",
    camera = list(eye = list(z = 2.5))
  )
)
```

```{r, echo=FALSE, fig.align='center', fig.align='center'}
# plotly::save_image(foo, "man/figures/sptcv_cstf_multiplot.png", width = 1200, height = 1200)
knitr::include_graphics("../man/figures/sptcv_cstf_multiplot.png")
```

Note: The image shown above is a static version created with `plotly::save_image()`.

Unfortunately, titles in subplots cannot be created dynamically.
However, there is a manual workaround via [annotations](https://plotly.com/r/reference/#layout-annotations) show in this [RPubs post](https://rpubs.com/bcd/subplot-titles).

## Examples of spatial partitioning (2D)

The following plots are based on a three fold partitioning of the [ecuador](https://mlr3spatiotempcv.mlr-org.com/reference/mlr_tasks_ecuador.html) example task.
Method `mlr_resampling_sptcv_cstf` is omitted here as the `ecuador` task does not have a factor grouping variable which is required by this method.
The partitions created on this dataset do not claim to necessarily make sense in practice - they should only contrast the individual methods against each other.
The source code can be found [here](https://github.com/mlr-org/mlr3spatiotempcv/blob/main/vignettes/spatiotemp-viz.Rmd).

```{r, echo=FALSE, cache=TRUE, fig.width=10, fig.height=8}
library(patchwork)
library(ggplot2)
library(mlr3)
library(mlr3spatiotempcv)
task_ecuador = tsk("ecuador")
task_cookfarm = tsk("cookfarm_mlr3")

set.seed(42)

# block
resampling_block = rsmp("spcv_block", range = 1000L, folds = 3)
p_block = autoplot(resampling_block, task_ecuador,
  fold_id = 1,
  show_blocks = TRUE, size = 0.6
) +
  scale_y_continuous(breaks = seq(-3.97, -4, -0.02)) +
  scale_x_continuous(breaks = seq(-79.085, -79.055, 0.02)) +
  theme(legend.position = "none")

# buffer
resampling_buffer = rsmp("spcv_buffer", theRange = 1000L)
p_buffer = autoplot(resampling_buffer, task_ecuador, fold_id = 1, size = 0.6) +
  scale_y_continuous(breaks = seq(-3.97, -4, -0.02)) +
  scale_x_continuous(breaks = seq(-79.085, -79.055, 0.02)) +
  theme(legend.position = "none")

# coords
resampling_coords = rsmp("spcv_coords", folds = 3)
p_coords = autoplot(resampling_coords, task_ecuador, fold_id = 1, size = 0.6) +
  scale_y_continuous(breaks = seq(-3.97, -4, -0.02)) +
  scale_x_continuous(breaks = seq(-79.085, -79.055, 0.02)) +
  theme(legend.position = "none")

# env
resampling_env = rsmp("spcv_env", folds = 3, features = "distroad")
p_env = autoplot(resampling_env, task_ecuador, fold_id = 1, size = 0.6) +
  scale_y_continuous(breaks = seq(-3.97, -4, -0.02)) +
  scale_x_continuous(breaks = seq(-79.085, -79.055, 0.02)) +
  theme(legend.position = "none")

# disc
resampling_disc = rsmp("spcv_disc", folds = 3L, radius = 200L, buffer = 200L)
p_disc = autoplot(resampling_disc, task_ecuador, fold_id = 1, size = 0.6) +
  scale_y_continuous(breaks = seq(-3.97, -4, -0.02)) +
  scale_x_continuous(breaks = seq(-79.085, -79.055, 0.02)) +
  theme(legend.position = "none")

# tiles
resampling_tiles = rsmp("spcv_tiles", nsplit = c(4L, 3L), reassign = FALSE)
p_tiles = autoplot(resampling_tiles, task_ecuador, fold_id = 1, size = 0.6) +
  scale_y_continuous(breaks = seq(-3.97, -4, -0.02)) +
  scale_x_continuous(breaks = seq(-79.085, -79.055, 0.02)) +
  theme(legend.position = "none")

# plot
p_block + p_coords + p_env + p_disc + p_tiles + p_buffer +
  plot_layout(ncol = 3) +
  plot_annotation(tag_levels = "A")
```

| Identifier | Method Name                  |
| :--------- | :--------------------------- |
| A          | `mlr_resampling_spcv_block`  |
| B          | `mlr_resampling_spcv_coords` |
| C          | `mlr_resampling_spcv_env`    |
| D          | `mlr_resampling_spcv_disc`   |
| E          | `mlr_resampling_spcv_tiles`  |
| F          | `mlr_resampling_spcv_buffer` |