In tabrasel/WetlandTools: General utility functions for TerrainWorks

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(TerrainWorksUtils)

library(data.table)
library(tidyverse)
library(mlr3)
library(mlr3spatial)
library(mlr3pipelines)

Training a Model

This package does not include many tools for building a model, since the mlr3 ecosystem is already pretty nimble and easy-to-use. This vignette lays out a few of those steps, and shows a few useful ways of using mlr3 for this specific modeling purpose.

Training data

See vignette("create_training_data.Rmd") for details.

First, we need to find and load our training data.

files <- list.files(path = params$training_data, 
                    pattern = ".Rdata$", 
                    full.names = TRUE)
for (f in files) {load(f)}

This training data happens to come in separate data frames for the positive/negative points, and the different years, so it requires a bit of preprocessing to create the big data frame we want. Your data might not require this.

# fix an issue with the geo input
ls_2011$geo <- ls_2011$geo$GeoClass
ls_2007$geo <- ls_2007$geo$GeoClass

# concatenate into one data frame
data <- rbind(transform(ls_1996, year = "1996"),
                    transform(ls_2007, year = "2007"),
                    transform(ls_2011, year = "2011"),
                    transform(nonls_1996, year = "1996"),
                    transform(nonls_2007, year = "2007"),
                    transform(nonls_2011, year = "2011"))

# remove NA's and factorize some variables
train_data <- as_tibble(data[!(is.na(data$dist_to_road)), 2:length(data)])
train_data$geo <- as.factor(train_data$geo)
train_data$class <- as.factor(train_data$class)

ranges <- sapply(train_data, function(x) {
            if(is.numeric(x)) {return(range(x))}
            else (return(c(0, 0)))
          })

Feature manipulation

Machine learning models often perform slightly better when we augment the data in certain ways. We found that our landslide model generally improves when we added quadratic features (e.g., gradient^2) and interaction terms (e.g., gradient * pca).

The LSutils package includes a helpful function for doing this - make_quadratic_features - which is pipe-friendly and operates pretty fast.

It is important that you carefully track what you do to your data before training the model - all of these steps will need to be implemented on any new data that you use for predicting.

train_data$ln_pca <- log(train_data$pca_k1_48)

training_cols <- c("x", "y", "class", 
                   "gradient", "mean_curv", "ln_pca")

train_data <- train_data %>% 
                select(all_of(training_cols)) %>%
                make_quadratic_features(ignore = c("x", "y", "class"))
train_data

Defining the task

Mlr3 uses tasks to describe the target of a machine learning problem. Since our problem is landslide susceptibility and we are using spatiotemporal data, we use the method as_task_classif_st() to define a spatiotemporal classification problem.

ls_task <- as_task_classif_st(
  x = train_data, 
  id = "landslide_initiation",
  target = "class", 
  positive = "pos", 
  coordinate_names = c("x", "y"), 
  crs = "epsg:26910"
)
ls_task

The last thing to do before training the model is scaling and centering the data. This is a step that is easily built into mlr3, so we can do it nicely before the task.

# seed for repeatability
set.seed(12345)

# create a pipeline for data preprocessing
scaler <- 
  po("scale") %>>% 
  po("encode") 
scaler$train((ls_task))

# save the scaling information for later
scale_vals <- t(tibble(center_vals = scaler$pipeops$scale$state$center, 
                       scale_vals = scaler$pipeops$scale$state$scale))
colnames(scale_vals) <- ls_task$feature_names

# scale the data
scaled <- scaler$predict(ls_task)$encode.output

Train the model

Now that our data is ready, the model can be created and trained in just a few lines of code.

ls_model <- lrn("classif.log_reg", 
            predict_type = "prob")

ls_model$train(scaled)
ls_model$model

Next steps

This is just the beginning of the machine learning workflow. You will likely want to compare different types of models, implement a cross-validation scheme, calculate performance metrics, etc. The mlr3 ecosystem offers many things that could help (check out mlr3spatiotempcv for cross-validation, mlr3resample and mlr3benchmark for comparison frameworks), but we will not go into any more specifics here.

tabrasel/WetlandTools documentation built on Dec. 20, 2024, 8:50 a.m.