# 11 Statistical learning for geographic data

## Prerequisites

This chapter assumes proficiency with spatial data, for example gained by studying the contents and working-through the exercises in chapters 2 to 6. A familiarity with generalized linear regression and machine learning is highly recommended (for example from Zuur et al. 2009; James et al. 2013).

The chapter uses the following packages:^{54}

Required data will be attached in due course.

## 11.1 Introduction

Statistical learning is concerned with the use of statistical and computational models for identifying patterns in data and predicting from these patterns.
Due to its origins, statistical learning is one of R’s great strengths (see section 1.3).^{55}

Statistical learning combines and blends methods from both statistics and machine learning that learn from data. Roughly, one can distinguish statistical learning into supervised and unsupervised techniques, both of which are used throughout a vast range of disciplines including economics, physics, medicine, biology, ecology and geography (James et al. 2013).

This chapter focuses on supervised techniques, as opposed to unsupervised techniques such as clustering. Response variables can be binary (such as landslide occurrence), categorical (land use), integer (species richness count) or numeric (soil acidity measured in pH). Supervised techniques model the relationship between such responses — which are known for a sample of observations — and one or more predictors.

The primary aim of machine learning is to make good predictions. It is increasingly appealing in the age of ‘big data’ because it makes few assumptions about input variables and can scale to handle problems that involve large datasets. Machine learning is conducive to tasks such as the prediction of future customer behavior, recommendation services (music, movies, what to buy next), face recognition, autonomous driving, text classification and predictive maintenance (infrastructure, industry).

This chapter is based on a case study: the (spatial) prediction of landslides.
This application links to the applied nature of geocomputation, defined in Chapter 1, and illustrates how machine learning borrows from the field of statistics when the sole aim is prediction.
Therefore, this chapter first introduces modeling and cross-validation concepts with the help of a Generalized Linear Model (GLM; Zuur et al. 2009).
Building on this the chapter implements a more typical machine learning algorithm, namely a Support Vector Machine (SVM).
The models’ **predictive performance** will be assessed using spatial cross-validation (CV), which accounts for the fact that geographic data is special.

CV determines a model’s ability to generalize to new data, by splitting a dataset (repeatedly) into training and test sets. It uses the training data to fit the model, and checks its performance when predicting to the test data. CV helps to detect overfitting since models that predict the training data too closely (noise) will tend to perform poorly on the test data.

Randomly splitting spatial data can lead to training points that are neighbors in space with test points.
Due to spatial autocorrelation, test and training datasets would not be independent in this scenario, with the consequence that CV fails to detect a possible overfitting.
Spatial CV alleviates this problem and is the **central** theme in this chapter.

## 11.2 Case study: Landslide susceptibility

The case study is based on a dataset of landslide locations in Southern Ecuador, illustrated in Figure 11.1 and described in detail in Muenchow, Brenning, and Richter (2012).
A subset of the dataset used in that paper is provided in the **RSAGA** package, which can be loaded as follows:

This should load three objects: a `data.frame`

named `landslides`

, a `list`

named `dem`

, and an `sf`

object named `study_area`

.
`landslides`

contains a factor column `lslpts`

where `TRUE`

corresponds to an observed landslide ‘initiation point’, with the coordinates stored in columns `x`

and `y`

.^{56}

The coordinates for the non-landslide points were sampled randomly from the study area, with the restriction that they must fall outside a small buffer around the landslide polygons.
There are 175 landslide points and 1360 non-landslide, as shown by `summary(landslides)`

.
To make number of landslide and non-landslide points balanced, let us sample 175 from the 1360 non-landslide points.

```
# select non-landslide points
non_pts = filter(landslides, lslpts == FALSE)
# select landslide points
lsl_pts = filter(landslides, lslpts == TRUE)
# randomly select 175 non-landslide points
set.seed(11042018)
non_pts_sub = sample_n(non_pts, size = nrow(lsl_pts))
# create smaller landslide dataset (lsl)
lsl = bind_rows(non_pts_sub, lsl_pts)
```

`dem`

is a digital elevation model consisting of two elements:
`dem$header`

, a `list`

which represents a raster ‘header’ (see section 2.2), and `dem$data`

, a matrix with the altitude of each pixel.
`dem`

can be converted into a `raster`

object with:

```
dem = raster(
dem$data,
crs = dem$header$proj4string,
xmn = dem$header$xllcorner,
xmx = dem$header$xllcorner + dem$header$ncols * dem$header$cellsize,
ymn = dem$header$yllcorner,
ymx = dem$header$yllcorner + dem$header$nrows * dem$header$cellsize
)
```

To model landslide susceptibility, we need some predictors.
Terrain attributes are frequently associated with landsliding (Muenchow, Brenning, and Richter 2012), and these can be computed from the digital elevation model (`dem`

) using R-GIS bridges (see Chapter 9).
We leave it as an exercise to the reader to compute the following terrain attribute rasters and extract the corresponding values to our landslide/non-landslide data frame (see exercises):

`slope`

: slope angle (°).`cplan`

: plan curvature (rad m^{−1}) expressing the convergence or divergence of a slope and thus water flow.`cprof`

: profile curvature (rad m^{-1}) as a measure of flow acceleration, also known as downslope change in slope angle.`elev`

: elevation (m a.s.l.) as the representation of different altitudinal zones of vegetation and precipitation in the study area.`log10_carea`

: the decadic logarithm of the catchment area (log10 m^{2}) representing the amount of water flowing towards a location.

The first three rows of the resulting data frame, still named `lsl`

look like this (rounded to two significant digits):

```
#> x y lslpts slope cplan cprof elev log10_carea
#> 1 715078 9558647 FALSE 37 0.021 0.0087 2500 2.6
#> 2 713748 9558047 FALSE 42 -0.024 0.0068 2500 3.1
#> 3 712508 9558887 FALSE 20 0.039 0.0150 2100 2.3
```

As a convenience to the reader, `lsl`

is also available in the **spDataLarge** package along with the corresponding terrain attributes stored in a raster stack (`data("ta", package = "spDataLarge")`

).

## 11.3 Conventional modeling approach in R

Before introducing the **mlr** package, an umbrella-package providing a unified interface to dozens of learning algorithms (section 11.5), it is worth taking a look at the conventional modeling interface in R.
This introduction to supervised statistical learning provides the basis for doing spatial CV, and contributes to a better grasp on the **mlr** approach presented subsequently.

Supervised learning involves predicting a response variable as a function of predictors (section 11.1).
In R, modeling functions are usually specified using formulas (see `?formula`

and the detailed Formulas in R Tutorial for details of R formulas).
The following command specifies and runs a generalized linear model:

It is worth understanding each of the three input arguments:

- A formula, which specifies landslide occurrence (
`lslpts`

) as a function of the predictors. - A family, which specifies the type of model, in this case
`binomial`

because the response is binary (see`?family`

). - The dataframe which contains the response and the predictors.

The results of this model can be printed as follows (`summary(fit)`

provides a more detailed account of the results):

```
class(fit)
#> [1] "glm" "lm"
fit
#>
#> Call: glm(formula = lslpts ~ slope + cplan + cprof + elev + log10_carea,
#> family = binomial(), data = lsl)
#>
#> Coefficients:
#> (Intercept) slope cplan cprof elev
#> 1.97e+00 9.30e-02 -2.57e+01 -1.43e+01 2.41e-05
#> log10_carea
#> -2.12e+00
#>
#> Degrees of Freedom: 349 Total (i.e. Null); 344 Residual
#> Null Deviance: 485
#> Residual Deviance: 361 AIC: 373
```

The model object `fit`

, of class `glm`

, contains the coefficients defining the fitted relationship between response and predictors.
It can also be used for prediction.
This is done with the generic `predict()`

method, which in this case calls the function `predict.glm()`

.
Setting `type`

to `response`

returns the predicted probabilities (of landslide occurrence) for each observation in `lsl`

, as illustrated below (see `?predict.glm`

):

```
pred_glm = predict(object = fit, type = "response")
head(pred_glm)
#> 1 2 3 4 5 6
#> 0.3327 0.4755 0.0995 0.1480 0.3486 0.6766
```

Spatial predictions can be made by applying the coefficients to the predictor rasters.
This can be done manually or with `raster::predict()`

.
In addition to a model object (`fit`

), this function also expects a raster stack with the predictors named as in the model’s input dataframe (Figure 11.2).

```
# attaching ta, a raster brick containing the predictors
data("ta", package = "spDataLarge")
# making the prediction
pred = raster::predict(object = ta, model = fit,
type = "response")
```

Here, when making predictions we neglect spatial autocorrelation since we assume that on average the predictive accuracy remains the same with or without spatial autocorrelation structures.
However, it is possible to include spatial autocorrelation structures into models (Zuur et al. 2009, 2017; Blangiardo and Cameletti 2015) as well as into predictions (kriging approaches, see e.g., Goovaerts 1997; Hengl 2007; **???**).
This is, however, beyond the scope of this book.

Spatial prediction maps are one very important outcome of a model.
Even more important is how good the underlying model is at making them since a prediction map is useless if the model’s predictive performance is bad.
The most popular measure to assess the predictive performance of a binomial model is the Area Under the Receiver Operator Characteristic Curve (AUROC).
This is a value between 0.5 and 1.0 with 0.5 indicating no and 1.0 indicating a perfect discrimination of the two classes.
Thus, the higher the AUROC the better is our model at making predictions.
In the following we compute the receiver operator characteristic with the help of `roc()`

by providing it with the response variable and the predicted values.
`auc()`

returns the area under the curve.

An AUROC of 0.83 represents a good fit. However, this is an overoptimistic estimation since we have computed it on the complete dataset. To derive a biased-reduced assessment we have to use cross-validation and in the case of spatial data should make use of spatial CV.

## 11.4 Introduction to (spatial) cross-validation

Cross-validation belongs to the family of resampling methods (James et al. 2013). The basic idea is to split (repeatedly) a dataset into training and test sets whereby the training data is used to fit a model which then is applied to the test set. Comparing the predicted values with the known response values from the test set (using a performance measure such as the AUROC in the binomial case) gives a bias-reduced assessment of the model’s capability to generalize the learned relationship to independent data. For example, a 100-repeated 5-fold cross-validation means to randomly split the data into five partitions (folds) with each fold being used once as a test set (see upper row of Figure 11.3). This guarantees that each observation is used once in one of the test sets, and requires the fitting of five models. Subsequently, this procedure is repeated 100 times. Of course, the data splitting will differ in each repetition. Overall, this sums up to 500 models whereas the mean performance measure (AUROC) of all models is the model’s overall predictive power.

However, geographic data is special.
As we saw in Chapter 12, the ‘first law’ of geography states that points close to each other are, generally, more similar than points further away (Miller 2004).
This means these points are not statistically independent because training and test points in conventional CV are often too close to each other (see first row of 11.3).
‘Training’ observations near the ‘test’ observations can provide a kind of ‘sneak preview’:
information that should be unavailable to the training dataset.
To alleviate this problem ‘spatial partitioning’ is used to split the observations into spatially disjointed subsets (using the observations’ coordinates in a *k*-means clustering; A. Brenning (2012b); second row of Figure 11.3).
This partitioning strategy is the **only** difference between spatial and conventional CV.
As a result spatial CV leads to a bias-reduced assessment of a model’s predictive performance, and hence helps to avoid overfitting.

## 11.5 Spatial CV with **mlr**

There are dozens of packages for statistical learning, as described for example in the CRAN machine learning task view.
Getting acquainted with each of these packages, including how to undertake cross-validation and hyperparameter tuning, can be a time-consuming process.
Comparing model results from different packages can be even more laborious.
The **mlr** package was developed to address these issues.
It acts as a ‘meta-package’, providing a unified interface to popular supervised statistical learning techniques including classification, regression, survival analysis and clustering (Bischl et al. 2016)
The standardized **mlr** interface is based on eight ‘building blocks’.
As illustrated in Figure 11.4, these have a clear order.

The **mlr** modeling process consists of three main stages.
First, a **task** specifies the data (including response and predictor variables) and the model type (such as regression or classification).
Second, a **learner** defines the specific learning algorithm that is applied to the created task.
Third, the **resampling** approach assesses the predictive performance of the model, i.e. its ability to generalize to new data (see also section 11.1).

### 11.5.1 Generalized linear model

To implement a GLM in **mlr** we must create a **task** containing the landslide data.
Since the response is binary (two-category variable) we create a classification task with `makeClassifTask()`

(for regression tasks use `makeRegrTask()`

, see `?makeClassifTask`

for other task types).
The first essential argument of these `make*()`

functions is `data`

.
The `target`

argument expects the name of a response variable and `positive`

determines which of the two factor levels of the response variable indicate the landslide initiation point (in our case this is `TRUE`

).
All other variables of the `lsl`

dataset will serve as predictors except for the coordinates (see the result of `getTaskFormula(task)`

for the model formula).
For spatial CV the `coordinates`

parameter is used (see section 11.1 and Figure 11.3) which expects the coordinates as a xy-dataframe.

```
library(mlr)
# coordinates needed for the spatial partitioning
coords = lsl[, c("x", "y")]
# select response and predictors to use in the modeling
data = dplyr::select(lsl, -x, -y)
coords = lsl[, c("x", "y")]
# create task
task = makeClassifTask(data = data, target = "lslpts",
positive = "TRUE", coordinates = coords)
```

`makeLearner()`

determines the statistical learning method to use.
All classification **learners** start with `classif.`

and all regression learners with `regr.`

(see `?makeLearners`

for details).
`listLearners()`

helps to find out about all available learners and from which package **mlr** imports them (Table 11.1).
For a specific task, we can run:

```
listLearners(task, warn.missing.packages = FALSE) %>%
dplyr::select(class, name, short.name, package) %>%
head
```

class | name | short.name | package |
---|---|---|---|

classif.binomial | Binomial Regression | binomial | stats |

classif.featureless | Featureless classifier | featureless | mlr |

classif.fnn | Fast k-Nearest Neighbour | fnn | FNN |

classif.knn | k-Nearest Neighbor | knn | class |

classif.lda | Linear Discriminant Analysis | lda | MASS |

classif.logreg | Logistic Regression | logreg | stats |

This yields all learners able to model two-class problems (landslide yes or no).
We opt for the binomial classification method used in section 11.3 and implemented as `classif.binomial`

in **mlr**.
Additionally, we must specify the link-function, `logit`

in this case, which is also the default of the `binomial()`

function.
`predict.type`

determines the type of the prediction with `prob`

resulting in the predicted probability for landslide occurrence between 0 and 1 (this corresponds to `type = response`

in `predict.glm`

).

```
lrn = makeLearner(cl = "classif.binomial",
link = "logit",
predict.type = "prob",
fix.factors.prediction = TRUE)
```

To find out from which package the specified learner is taken and how to access the corresponding help pages, we can run:

The set-up steps for modeling with **mlr** may seem tedious.
But remember this single interface provides access to the 150+ learners shown by `listLearners()`

; it would be far more tedious to learn the interface for each learner!
Further advantages are simple parallelization of resampling techniques and the ability to tune machine learning hyperparameters (see section 11.5.2).
Most importantly, (spatial) resampling in **mlr** is straightforward, requiring only two more steps: specifying a resampling method and running it.
We will use a 100-repeated 5-fold spatial CV: five partitions will be chosen based on the provided coordinates in our `task`

and the partitioning will be repeated 100 times:^{57}

To execute the spatial resampling, we run `resample()`

using the specified learner, task, resampling strategy and of course the performance measure, here the AUROC.
This takes some time (around 10 seconds on a modern laptop) because it computes the AUROC for 500 models.
Setting a seed ensures the reprocubility of the obtained result and will ensure the same spatial partitioning when re-running the code.

```
set.seed(012348)
sp_cv = mlr::resample(learner = lrn, task = task,
resampling = resampling,
measures = mlr::auc)
```

The output of the preceding code chunk is a bias-reduced assessment of the model’s predictive performance, as illustrated in the following code chunk (required input data is saved in the file `spatialcv.Rdata`

in the book’s GitHub repo):

```
# summary statistics of the 500 models
summary(sp_cv$measures.test$auc)
#> Min. 1st Qu. Median Mean 3rd Qu. Max.
#> 0.686 0.757 0.789 0.780 0.795 0.861
# mean AUROC of the 500 models
mean(sp_cv$measures.test$auc)
#> [1] 0.78
```

To put these results in perspective let us compare them with AUROC values from a 100-repeated 5-fold non-spatial cross-validation (Figure 11.5; the code for the non-spatial cross-validation is not shown here but will be explored in the exercise section). As expected, the spatially cross-validated result yields lower AUROC values on average than the conventional cross-validation approach, underlining the over-optimistic predictive performance due to spatial autocorrelation of the latter.

### 11.5.2 Spatial tuning of machine-learning hyperparameters

Section 11.1 introduced machine learning as part of statistical learning. To recap, we adhere to the following definition of machine learning by Jason Brownlee:

Machine learning, more specifically the field of predictive modeling is primarily concerned with minimizing the error of a model or making the most accurate predictions possible, at the expense of explainability. In applied machine learning we will borrow, reuse and steal algorithms from many different fields, including statistics and use them towards these ends.

In section 11.5.1 a GLM was used to predict landslide susceptibility.
This section introduces support vector machines (SVM) for the same purpose.
In short, SVMs search for the best possible ‘hyperplanes’ to separate classes (in a classification case) and estimate ‘kernels’ with specific hyperparameters to allow for non-linear boundaries between classes (James et al. 2013).
Hyperparameters should not be confused with coefficients of parametric models, which are sometimes also referred to as parameters.^{58}
Coefficients can be estimated from the data while hyperparameters are set before the learning begins.
Optimal hyperparameters are usually determined within a defined range with the help of cross-validation methods.
This is called hyperparameter tuning.

Some SVM implementations such as that provided by **kernlab** allow hyperparameters to be tuned automatically, usually based on random sampling (see upper row of Figure 11.3).
This works for non-spatial data but is of less use for spatial data where ‘spatial tuning’ should be undertaken.

Before defining spatial tuning we will set-up the **mlr** building blocks, introduced in section 11.5.1, for the SVM.
The task remains the same as the `task`

object created in section 11.5.1.
Learners implementing SVM can be found using `listLearners()`

as follows:

```
lrns = listLearners(task, warn.missing.packages = FALSE)
filter(lrns, grepl("svm", class)) %>%
dplyr::select(class, name, short.name, package)
#> class name short.name package
#> 6 classif.ksvm Support Vector Machines ksvm kernlab
#> 9 classif.lssvm Least Squares Support Vector Machine lssvm kernlab
#> 17 classif.svm Support Vector Machines (libsvm) svm e1071
```

Of the options illustrated above, we will use `ksvm()`

from the **kernlab** package (Karatzoglou et al. 2004).
To allow for non-linear relationships we use the popular radial basis function (or Gaussian) kernel which is also the default of `ksvm()`

.

The next stage is to specify a resampling strategy. Again we will use a 100-repeated 5-fold spatial CV:

So far, the process is identical to that described in section 11.5.1. The next step is new, however: to tune the hyperparameters. Using the same data for the performance assessment and the tuning would potentially lead to overoptimistic results (Cawley and Talbot 2010). This can be avoided using nested spatial CV.

This means that we split each fold again into five spatially disjoint subfolds which are used to determine the optimal hyperparameters (`tune_level`

object in the code chunk below; see Figure 11.6 for a visual representation).
To find the optimal hyperparameter combination we here fit 50 models in each of these subfolds with randomly selected hyperparameter values (`ctrl`

object in the code chunk below).
Additionally, we restrict the randomly chosen values to a predefined tuning space (`ps`

object).
The latter was chosen with values recommended in the literature (Schratz et al. 2018).

```
# five spatially disjoint partitions
tune_level = makeResampleDesc("SpCV", iters = 5)
# use 50 randomly selected hyperparameters
ctrl = makeTuneControlRandom(maxit = 50)
# define the outer limits of the randomly selected hyperparameters
ps = makeParamSet(
makeNumericParam("C", lower = -12, upper = 15, trafo = function(x) 2^x),
makeNumericParam("sigma", lower = -15, upper = 6, trafo = function(x) 2^x)
)
```

The next stage is to modify the learner `lrn_ksvm`

in accordance with all the characteristics defining the hyperparameter tuning with `makeTuneWrapper()`

.

```
wrapped_lrn_ksvm = makeTuneWrapper(learner = lrn_ksvm,
resampling = tune_level,
par.set = ps,
control = ctrl,
show.info = TRUE,
measures = mlr::auc)
```

The **mlr** is now set-up to fit 250 models to determine optimal hyperparameters for one fold.
Repeating this for each fold, we end up with 1250 (250 * 5) models for each repetition.
Repeated 100 times means fitting a total of 125,000 models to identify optimal hyperparameters (Figure 11.3).
These are used in the performance estimation, which requires the fitting of another 500 models (5 folds * 100 repetitions; see Figure 11.3).

The process of hyperparameter tuning and performance estimation is computationally intensive. Model runtime can be reduced with parallelization, which can be done in a number of ways, depending on the operating system.

Before starting the parallelization, we ensure that the processing continues even if one of the models throws an error by setting `on.learner.error`

to `warn`

.
This avoids the process stopping just because of one failed model, which is desirable on large model runs.
To inspect the failed models once the processing is completed, we dump them:

To start the parallelization, we set the `mode`

to `multicore`

which will use `mclapply()`

in the background on a single machine in the case of a Unix-based operating system^{59}
Equivalenty, `parallelStartSocket()`

enables parallelization under Windows.
`level`

defines the level at which to enable parallelization, with `mlr.tuneParams`

determining that the hyperparameter tuning level should be parallelized (see lower left part of Figure 11.6, `?parallelGetRegisteredLevels`

, and the **mlr** parallelization tutorial for details).
We will use half of the available cores (set with the `cpus`

parameter), a setting that allows possible other users to work on the same high performance computing cluster in case one is used (which was the case when we ran the code).
Setting `mc.set.seed`

to `TRUE`

ensures that the randomly chosen hyperparameters during the tuning can be reproduced when running the code again.
Unfortunately, `mc.set.seed`

is only available under Unix-based systems.

```
library(parallelMap)
if (Sys.info()["sysname"] %in% c("Linux, Darwin")) {
parallelStart(mode = "multicore",
# parallelize the hyperparameter tuning level
level = "mlr.tuneParams",
# just use half of the available cores
cpus = round(parallel::detectCores() / 2),
mc.set.seed = TRUE)
}
if (Sys.info()["sysname"] == "Windows") {
parallelStartSocket(level = "mlr.tuneParams",
cpus = round(parallel::detectCores() / 2))
}
```

Now we are set-up for computing the nested spatial CV.
Using a seed allows to recreate the exact same spatial partitions when re-running the code.
Specifying the `resample()`

parameters follows the exact same procedure as presented when using a GLM, the only difference being the `extract`

argument.
This allows the extraction of the hyperparameter tuning results which is important if we plan follow-up analyses on the tuning.
After the processing, it is good practice to explicitly stop the parallelization with `parallelStop()`

.
Finally, we save the output object (`result`

) to disk in case we would like to use it another R session.
Before running the subsequent code, be aware that it is time-consuming:
the 125,500 models took ~1/2hr on a server using 24 cores (see below).

```
set.seed(12345)
result = mlr::resample(learner = wrapped_lrn_ksvm,
task = task,
resampling = perf_level,
extract = getTuneResult,
measures = mlr::auc)
# stop parallelization
parallelStop()
# save your result, e.g.:
# saveRDS(result, "svm_sp_sp_rbf_50it.rds")
```

In case you do not want to run the code locally, we have saved a subset of the results in the book’s GitHub repo. They can be loaded as follows:

Note that runtime depends on many aspects: CPU speed, the selected algorithm, the selected number of cores and the dataset.

Even more important than the runtime is the final aggregated AUROC: the model’s ability to discriminate the two classes.

```
# final aggregated AUROC
result$aggr
#> auc.test.mean
#> 0.758
# same as
mean(result$measures.test$auc)
#> [1] 0.758
```

It appears that the GLM (aggregated AUROC was 0.78) is slightly better than the SVM in this specific case. However, using more than 50 iterations in the random search would probably yield hyperparameters that result in models with a better AUROC (Schratz et al. 2018). On the other hand, increasing the number of random search iterations would also increase the total number of models and thus runtime

The estimated optimal hyperparameters for each fold at the performance estimation level can also be viewed. The following command shows the best hyperparameter combination of the first fold of the first iteration (recall this results from the first 5 * 50 model runs):

```
# winning hyperparameters of tuning step, i.e. the best combination out of 50 *
# 5 models
result$extract[[1]]$x
#> $C
#> [1] 0.458
#>
#> $sigma
#> [1] 0.023
```

The estimated hyperparameters have been used for the first fold in the first iteration of the performance estimation level which resulted in the following AUROC value:

So far spatial CV has been used to assess the ability of learning algorithms to generalize to unseen data.
For spatial prediction, one would tune the hyperparameters on the complete dataset (see Chapter **??**).

## 11.6 Conclusions

Resampling methods are an important part of a data scientist’s toolbox (James et al. 2013). This chapter used cross-validation to assess predictive performance of various models. As described in section 11.1, observations with spatial coordinates may not be statistically independent due to spatial autocorrelation, violating a fundamental assumption of cross-validation. Spatial CV addresses this issue by reducing bias introduced by spatial autocorrelation.

The **mlr** package facilitates (spatial) resampling techniques in combination with the most popular statistical learning techniques including linear regression, semi-parametric models such as generalized additive models and machine learning techniques such as random forests, SVMs, and boosted regression trees (Bischl et al. 2016; Schratz et al. 2018).
Machine learning algorithms often require hyperparameter inputs, the optimal ‘tuning’ of which can require thousands of model runs which require large computational resources, consuming much time, RAM and/or cores.
**mlr** tackles this issue by enabling parallelization.

Machine learning overall, and its use to understand spatial data, is a large field and this chapter has provided the basics, but there is more to learn. We recommend the following resources in this direction:

- The
**mlr**tutorials on Machine Learning in R and Handling of spatial Data. - An academic paper on hyperparameter tuning (Schratz et al. 2018).
- In case of spatio-temporal data, one should account for spatial and temporal autocorrelation when doing CV (Meyer et al. 2018).

## 11.7 Exercises

- Compute the following terrain attributes from the
`dem`

datasets loaded with`data("landslides", package = "RSAGA")`

with the help of R-GIS bridges (see Chapter 9):- slope
- plan curvature
- profile curvature
- catchment area

- Extract the values from the corresponding output rasters to the
`landslides`

data frame (`data(landslides, package = "RSAGA"`

) by adding new variables called`slope`

,`cplan`

,`cprof`

,`elev`

and`log_carea`

. Keep all landslide initiation points and 175 randomly selected non-landslide points (see section 11.2 for details). - Use the derived terrain attribute rasters in combination with a GLM to make a spatial prediction map similar to that shown in Figure 11.2.
Running
`data("study_mask", package = "spDataLarge")`

attaches a mask of the study area. - Compute a 100-repeated 5-fold non-spatial cross-validation and spatial CV based on the GLM learner and compare the AUROC values from both resampling strategies with the help of boxplots (see Figure 11.5). Hint: You need to specify a non-spatial task and a non-spatial resampling strategy.
- Model landslide susceptibility using a quadratic discriminant analysis (QDA, James et al. 2013). Assess the predictive performance (AUROC) of the QDA. What is the difference between the spatially cross-validated mean AUROC value of the QDA and the GLM? Hint: Before running the spatial cross-validation for both learners set a seed to make sure that both use the same spatial partitions which in turn guarantees comparability.
- Run the SVM without tuning the hyperparameters.
Use the
`rbfdot`

kernel with \(\sigma\) = 1 and*C*= 1. Leaving the hyperparameters unspecified in**kernlab**’s`ksvm()`

would otherwise initialize an automatic non-spatial hyperparameter tuning. For a discussion on the need for (spatial) tuning of hyperparameters please refer to Schratz et al. (2018). - Model landslide susceptibility with the help of
**mlr**using a random forest model as implemented by the**ranger**package. Apply a nested spatial CV. Parallelize the tuning level. Use a random search with 50 iterations to find the optimal hyperparameter combination (here:`mtry`

and`num.trees`

). The tuning space limits are 1 and 4 for`mtry`

, and 1 and 10,000 for`num.trees`

. (warning: this might take a long time).

### References

Zuur, Alain, Elena N. Ieno, Neil Walker, Anatoly A. Saveliev, and Graham M. Smith. 2009. *Mixed Effects Models and Extensions in Ecology with R*. Statistics for Biology and Health. New York: Springer-Verlag.

James, Gareth, Daniela Witten, Trevor Hastie, and Robert Tibshirani, eds. 2013. *An Introduction to Statistical Learning: With Applications in R*. Springer Texts in Statistics 103. New York: Springer.

Muenchow, Jannes, Alexander Brenning, and Michael Richter. 2012. “Geomorphic Process Rates of Landslides Along a Humidity Gradient in the Tropical Andes.” *Geomorphology* 139-140 (February): 271–84. https://doi.org/10.1016/j.geomorph.2011.10.029.

Zuur, Alain F., Elena N. Ieno, Anatoly A. Saveliev, and Alain F. Zuur. 2017. *Beginner’s Guide to Spatial, Temporal and Spatial-Temporal Ecological Data Analysis with R-INLA*. Vol. Volume 1: Using GLM and GLMM. Newburgh, United Kingdom: Highland Statistics Ltd.

Blangiardo, Marta, and Michela Cameletti. 2015. *Spatial and Spatio-Temporal Bayesian Models with R-INLA*. Chichester, UK: John Wiley & Sons, Ltd. https://doi.org/10.1002/9781118950203.

Goovaerts, Pierre. 1997. *Geostatistics for Natural Resources Evaluation*. Applied Geostatistics Series. New York: Oxford University Press.

Hengl, Tomislav. 2007. *A Practical Guide to Geostatistical Mapping of Environmental Variables*. Luxembourg: Publications Office.

Miller, Harvey J. 2004. “Tobler’s First Law and Spatial Analysis.” *Annals of the Association of American Geographers* 94 (2).

Brenning, Alexander. 2012b. “Spatial Cross-Validation and Bootstrap for the Assessment of Prediction Rules in Remote Sensing: The R Package Sperrorest.” In, 5372–5. IEEE. https://doi.org/10.1109/IGARSS.2012.6352393.

Bischl, Bernd, Michel Lang, Lars Kotthoff, Julia Schiffner, Jakob Richter, Erich Studerus, Giuseppe Casalicchio, and Zachary M. Jones. 2016. “Mlr: Machine Learning in R.” *Journal of Machine Learning Research* 17 (170): 1–5.

Karatzoglou, Alexandros, Alex Smola, Kurt Hornik, and Achim Zeileis. 2004. “Kernlab - an S4 Package for Kernel Methods in R.” *Journal of Statistical Software* 11 (9). https://doi.org/10.18637/jss.v011.i09.

Cawley, Gavin C., and Nicola LC Talbot. 2010. “On over-Fitting in Model Selection and Subsequent Selection Bias in Performance Evaluation.” *Journal of Machine Learning Research* 11 (Jul): 2079–2107.

Schratz, Patrick, J. Muenchow, Eugenia Iturritxa, Jakob Richter, and A. Brenning. 2018. “Performance Evaluation and Hyperparameter Tuning of Statistical and Machine-Learning Models Using Spatial Data.” *Tba*.

Meyer, Hanna, Christoph Reudenbach, Tomislav Hengl, Marwan Katurji, and Thomas Nauss. 2018. “Improving Performance of Spatio-Temporal Machine Learning Models Using Forward Feature Selection and Target-Oriented Validation.” *Environmental Modelling & Software* 101 (March): 1–9. https://doi.org/10.1016/j.envsoft.2017.12.001.

Package

**pROC**,**RSAGA**and**spDataLarge**must also be installed although these do not need to be attached.↩Moreover, applying statistical techniques to geographic data has been an active topic of research for many decades, within the overlapping fields of Geostatistics and Spatial Statistics (Diggle and Ribeiro 2007; Gelfand et al. 2010) and the vibrant sub-field of point pattern analysis (Baddeley, Rubak, and Turner 2015).↩

The landslide initiation point is located in the scarp of a landslide polygon. See Muenchow, Brenning, and Richter (2012) for further details.↩

Note that package

**sperrorest**initially implemented spatial cross-validation in R (A. Brenning 2012b). In the meantime, its functionality was integrated into the**mlr**package which is the reason why we are using**mlr**(Schratz et al. 2018).The**caret**package is another umbrella-package (Kuhn and Johnson 2013) for streamlined modeling in R, however, so far it does not provide spatial CV which is why we refrain from using it for spatial data.↩For a detailed description of the difference between coefficients and hyperparameters, see the ‘machine mastery’ blog post on the subject. ↩

See

`?parallelStart`

for further modes and github.com/berndbischl/parallelMap for more on the unified interface to popular parallelization back-ends.↩