Read me before proceeding to the analysis!
Pine Beetles are insects that can cause devastation to forests by
eating into tree bark and upsetting the flow of water and nutrients to
infected trees. Pine Beetles cause economic and ecological harm, so it’s
important to figure out how to protect forests from them.
There are five categories of variables shown on the Variable Chart page,
which is accessed through the tabs at the top of this page: Tree
diameter, infestation severity, nearest brood tree, forest density, and
beetle population pressure.
Knowing how these variables
relate to each other can help foresters figure out how to protect
forests. For example, in this analysis, I will make two models for
predicting the minimum linear distance to the nearest brood tree, or
DeadDist. In other words, how far away from an infected tree can a given
healthy tree be to be considered “safe”? This is important for foresters
so that they don’t clear trees unnecessarily and know when to expect
tree death.
The panel to the left is a scatterplot matrix showing the
relationships between variables.
IND_BA_Infest_20th is a variable that shows whether a given
tree is or is not within 1/20th acres of an infested
tree. Variables like DeadDist or Infest_Serv1
also relate to distance and Pine Beetle infestation. As such, I was
interested to see whether dividing the each ggpair correlation into
either Yes (an infested tree is within 1/20th acres) or No (an infested
tree is not within 1/20th acres) to observe whether there was an
observable difference between Yes and No.
It looks like for
single-variable distributions (e.g. TreeDiam by
TreeDiam), the groups of points divided by
IND_BA_Infest_20th varied in whether they were observably
different or not. The TreeDiam plot did not, which makes
sense: why would the presence of a Pine Beetle within 1/20th acres
affect the diameter of an already grown tree? DeadDist,
another variable, has two very distinct groups for Yes and No, which
also makes sense as the closer to an infested tree a given tree is, the
more likely it is within 1/20th acres of that given tree.
One of the two models I used to find how far away an infected tree
can be from a given healthy tree be to be considered “safe” was a ridge
regression model. This kind of model is good for when there are numerous
predictors, like in this model. for this model I use
DeadDist_sqrt (I took the square root of DeadDist to make
the distribution more normal) as the outcome variable and
TreeDiam, Infest_Serv1, SDI_20th,
and IND_BA_Infest_20th as the predictor variables. In a
ridge regression the coefficients of each predictor variable is
optimized to create an accurate model, so some coefficients may be near
zero if they are not very relevant to the outcome variable, or a
coefficient may be much more than another variable’s coefficient if it
is very influential to the outcome variable.
This model’s
RMSE is 1.3785935, on which is a means that the average difference
between this model’s predicted values and the actual values in the
dataset is 1.3785935. To put that into the context of the scale of the
DeadDist_sqrt, the total range of values in
DeadDist_sqrt is 0.3277936 to 11.1783.
The graph
below was made with predicted values from the model and the actual
values from the testing data. There is a strange split at 3 where the
data divides into two groups: one that seems slightly diagonal where the
trees are nearer to an infested tree, and one horizontal, where the
nearest infested tree is not within 1/20th acres. I know that a
predicted vs actual graph for a regression model is supposed to be
diagonal, however I tried using different evaluations and the base r
function to no better of a graph. Additionally, the adjusted r squared
seems good. I cannot see where the model is incorrect with my current
time and skill, though I know it is somewhere. I am sorry about this but
I have the code and know how to evaluate it, so I am proceeding with
what I have.
The actual vs predicted graph here seems more diagonal, but is still
split into two groups.
A random forest model uses a decision
tree to narrow down what parts of the data indicate well for the outcome
variable values.
On the next page of this flexdashboard is a
comparison of the two models’ metrics
Overall, the random forest model had a lower rmse and a higher rsq. Both are indicators of a more functional model, so I will say the random forest model is more accurate.
---
title: "Predicting the Minimum Linear Distance from a Given Pine Tree to the Nearest Tree that is Infected by Pine Beetles"
output:
flexdashboard::flex_dashboard:
orientation: rows
vertical_layout: fill
source_code: embed
theme: yeti
---
```{r setup, include=FALSE}
library(flexdashboard) # base for presentation
library(tidyverse)
library(readxl) # reading in data
library(GGally) # ggpairs
library(tidymodels)
library(DT) # datatable
library(RColorBrewer) # colors for ggplots
library(randomForest) # model type
library(plotly) # guage
```
Read Me
==============================================================================
Row {data-height=700}
-----------------------------------------------------------------------
### What are Pine Beetles? And Why Are You Reading This?
__*Read me before proceeding to the analysis!*__
Pine Beetles are insects that can cause devastation to forests by eating into tree bark and upsetting the flow of water and nutrients to infected trees. Pine Beetles cause economic and ecological harm, so it's important to figure out how to protect forests from them.
<br>
</br>
There are five categories of variables shown on the Variable Chart page, which is accessed through the tabs at the top of this page: Tree diameter, infestation severity, nearest brood tree, forest density, and beetle population pressure.
<br>
</br>
Knowing how these variables relate to each other can help foresters figure out how to protect forests. For example, in this analysis, I will make two models for predicting the minimum linear distance to the nearest brood tree, or DeadDist. In other words, how far away from an infected tree can a given healthy tree be to be considered "safe"? This is important for foresters so that they don't clear trees unnecessarily and know when to expect tree death.
### Pine Beetle characteristics that might be relevant to our analysis
* Female Pine Beetles use pheromones to attract other Pine Beetles nearby when they are ready to lay eggs<sup>1</sup>
+ Pine Beetles will tend to attack as a group and infected trees will cause more infection around them (a reason to include infestation severity and nearest brood tree variables)
<br>
</br>
* Pine Beetles infestations in forests are usually directional<sup>1</sup>
<br>
</br>
* Most Pine Beetles like attacking bigger trees (>10in), but when there is a lot of competition for trees, they will attack smaller trees<sup>2</sup>
+ May be wise to include TreeDiam as a variable!
<br>
</br>
* Pine Beetles usually fly 300 feet or less at a time
+ 300 feet = ~.007 acres, which means that we will probably not want to use any of the beetle population pressure measures of over the smallest variable measure (1/20 acres = .05 acres)<sup>2</sup>
Row {data-height=200}
-----------------------------------------------------------------------
### Sources
1. [Basic Pine Beetle info](https://www.treefarmsystem.org/SPBwhenbeetlesattack)
2. [Pine Beetle FAQ USDA](https://www.fs.usda.gov/Internet/FSE_DOCUMENTS/stelprdb5304956.pdf)
3. [Stand Density Index info](https://fennerschool-associated.anu.edu.au/mensuration/density.html)
<br>
</br>
_Credit for a majority of the code used in these analyses comes from Dr. Timothy York from Virginia Commonwealth University_
Data Exploration
==============================================================================
Row {data-height=300}
-----------------------------------------------------------------------
### Data table, explore how the raw data looks!
```{r readindata}
pine_tbl <- read_excel("../../hgen-612_temp/p1/data/Data_1993.xlsx", sheet = 1)
```
```{r makeshowcasebones}
pine_tbl %>%
dplyr::select(DeadDist, TreeDiam, Infest_Serv1, SDI_20th, BA_Inf_20th, IND_BA_Infest_20th) %>%
DT::datatable()
```
Row {data-height=450}
-----------------------------------------------------------------------
### Factors related to Pine Beetle infestation in forests as they relate to each other when a tree within 1/20th acres is infested vs not
```{r ggpairsIND}
pine_fact_tbl <- pine_tbl %>%
mutate(IND_BA_Infest_20th_fac = factor(IND_BA_Infest_20th,
levels = 0:1,
labels = c("No", "Yes")))
ggpairs(pine_fact_tbl, columns = c("DeadDist",
"TreeDiam",
"Infest_Serv1",
"SDI_20th"), ggplot2::aes(color = IND_BA_Infest_20th_fac)) +
scale_fill_brewer(palette = "Dark2") +
scale_color_brewer(palette = "Dark2")
```
### Factors related to Pine Beetle infestation in forests as they relate to each other in general
```{r ggpairsregular}
ggpairs(pine_fact_tbl, columns = c("DeadDist",
"TreeDiam",
"Infest_Serv1",
"SDI_20th",
"IND_BA_Infest_20th"))
```
### Things to Note
The panel to the left is a scatterplot matrix showing the relationships between variables.
<br>
</br>
`IND_BA_Infest_20th` is a variable that shows whether a given tree _is_ or _is not_ within 1/20th acres of an infested tree. Variables like `DeadDist` or `Infest_Serv1` also relate to distance and Pine Beetle infestation. As such, I was interested to see whether dividing the each ggpair correlation into either Yes (an infested tree is within 1/20th acres) or No (an infested tree is not within 1/20th acres) to observe whether there was an observable difference between Yes and No.
<br>
</br>
It looks like for single-variable distributions (e.g. `TreeDiam` by `TreeDiam`), the groups of points divided by `IND_BA_Infest_20th` varied in whether they were observably different or not. The `TreeDiam` plot did not, which makes sense: why would the presence of a Pine Beetle within 1/20th acres affect the diameter of an already grown tree? `DeadDist`, another variable, has two very distinct groups for Yes and No, which also makes sense as the closer to an infested tree a given tree is, the more likely it is within 1/20th acres of that given tree.
Ridge Model
==============================================================================
Row {data-height=350}
-----------------------------------------------------------------------
### Ridge Model Evaluation
One of the two models I used to find how far away an infected tree can be from a given healthy tree be to be considered "safe" was a ridge regression model. This kind of model is good for when there are numerous predictors, like in this model. for this model I use `DeadDist_sqrt` (I took the square root of DeadDist to make the distribution more normal) as the outcome variable and `TreeDiam`, `Infest_Serv1`, `SDI_20th`, and `IND_BA_Infest_20th` as the predictor variables. In a ridge regression the coefficients of each predictor variable is optimized to create an accurate model, so some coefficients may be near zero if they are not very relevant to the outcome variable, or a coefficient may be much more than another variable's coefficient if it is very influential to the outcome variable.
<br>
</br>
This model's RMSE is 1.3785935, on which is a means that the average difference between this model's predicted values and the actual values in the dataset is 1.3785935. To put that into the context of the scale of the `DeadDist_sqrt`, the total range of values in `DeadDist_sqrt` is 0.3277936 to 11.1783.
<br>
</br>
The graph below was made with predicted values from the model and the actual values from the testing data. There is a strange split at 3 where the data divides into two groups: one that seems slightly diagonal where the trees are nearer to an infested tree, and one horizontal, where the nearest infested tree is not within 1/20th acres. I know that a predicted vs actual graph for a regression model is supposed to be diagonal, however I tried using different evaluations and the base r function to no better of a graph. Additionally, the adjusted r squared seems good. I cannot see where the model is incorrect with my current time and skill, though I know it is somewhere. I am sorry about this but I have the code and know how to evaluate it, so I am proceeding with what I have.
```{r sqrtdeaddist}
# I tried using step_sqrt(all_outcomes()) to normalize curve, but it kept giving me an error because predict() does not take models that have outcomes() used in the recipe. As such, I am transforming the data ahead of time outside of the recipe to avoid this error
pine_tbl_sqrt <- pine_fact_tbl %>%
mutate(DeadDist_sqrt = sqrt(DeadDist))
# source: https://stackoverflow.com/questions/75762005/error-in-step-log-when-trying-to-make-predictions-with-my-model
```
```{r datasplit}
# training/testing data
pine_split <- initial_split(pine_tbl_sqrt)
pine_train <- training(pine_split)
pine_test <- testing(pine_split)
```
```{r ridgerecipe, include=FALSE}
ridge_recipe <- pine_train %>%
recipe(DeadDist_sqrt ~ TreeDiam + Infest_Serv1 + SDI_20th + IND_BA_Infest_20th) %>%
step_corr(all_predictors()) %>% # gets rid of variables to get rid of multicollinearity
step_zv(-all_outcomes()) %>%
step_scale(all_predictors(), -all_outcomes()) # gets rid of variables of all one value
```
```{r ridgelambda, include=FALSE}
# credit for code used to tune lambda goes to Professor York!
# bootstrap samples for re-sampling and tuning the penalty parameter
set.seed(1234)
ridge_boot <- bootstraps(pine_train) # re-sampling
# a grid of tuning parameters
lambda_grid <- grid_regular(penalty(), levels = 50)
```
```{r ridgemodelandengine, include = FALSE}
ridge_model <- linear_reg(mixture = 0, penalty = tune()) %>%
set_engine("glmnet")
ridge_model %>%
translate()
```
```{r ridgeworkflow, include=FALSE}
# workflow
ridge_wflow <-
workflow() %>%
add_model(ridge_model) %>%
add_recipe(ridge_recipe)
```
```{r ridgetuning, include=FALSE}
set.seed(9870) # why do I need to set a seed here?
ridge_grid <- tune_grid(
ridge_wflow,
resamples = ridge_boot, # rset object (re-sample object)
grid = lambda_grid # the lambda list we are tuning to
)
# bootstrap results
ridge_grid %>%
collect_metrics()
```
```{r rmsechart, include=FALSE}
ridge_grid %>%
collect_metrics() %>%
ggplot(aes(penalty, mean, color = .metric)) +
geom_errorbar(aes(
ymin = mean - std_err,
ymax = mean + std_err
),
alpha = 0.5
) +
geom_line(linewidth = 1.5) +
facet_wrap(~.metric, scales = "free", nrow = 2) +
scale_x_log10() +
theme(legend.position = "none")
# rmse low and rsq high is more desireable
lowest_rmse <- ridge_grid %>% # the algorithm selects the best lambda
select_best("rmse")
# update final model with lowest rmse
final_ridge <- finalize_workflow(
ridge_wflow,
lowest_rmse
)
```
Row {data-height=450}
-----------------------------------------------------------------------
```{r ridgefit, include=FALSE}
final_ridge %>%
fit(data = pine_train)
final_ridge %>%
fit(pine_train) %>%
extract_fit_parsnip() %>%
tidy()
final_ridge %>%
extract_spec_parsnip()
last_fit(
final_ridge,
pine_split
) %>%
collect_metrics()
```
### RMSE
```{r ridgermse, include = FALSE}
ridge_predicted <- final_ridge %>%
fit(data = pine_train) %>%
predict(pine_test) %>%
bind_cols(pine_test)
ridge_metrics <- ridge_predicted %>%
metrics(truth = DeadDist_sqrt, estimate = .pred)
```
```{r ridgeguage}
plot_ly(
domain = list(x = c(0, 1), y = c(0, 1)),
value = 1.3785935,
title = list(text = "RMSE"),
type = "indicator",
mode = "gauge+number",
gauge = list(
axis = list(range = list(0, max(pine_tbl_sqrt$DeadDist_sqrt) - min(pine_tbl_sqrt$DeadDist_sqrt)), tickwidth = 1, tickcolor = "yellow"),
bar = list(color = "yellow")))
# https://plotly.com/r/gauge-charts/
```
### Actual vs Predicted Values
```{r actualvspredictedridge}
ridge_predicted %>%
ggplot(aes(x = DeadDist_sqrt, y = .pred, color = factor(IND_BA_Infest_20th,
levels = 0:1,
labels = c("No", "Yes")))) +
geom_point(alpha = 0.5) +
labs(y = "Predicted Minimum Distance to an Infected Tree (sqrt)", x = "Actual Minimum Distance to an Infected Tree", color = "Infested tree within 1/20th acres?") +
coord_obs_pred() +
scale_fill_brewer(palette = "Dark2") +
scale_color_brewer(palette = "Dark2")
```
Random Forest Model
==============================================================================
Row {data-height=350}
-----------------------------------------------------------------------
### Random Forest Evaluation
The actual vs predicted graph here seems more diagonal, but is still split into two groups.
<br>
</br>
A random forest model uses a decision tree to narrow down what parts of the data indicate well for the outcome variable values.
<br>
</br>
On the next page of this flexdashboard is a comparison of the two models' metrics
```{r forestrecipe}
forest_recipe <- pine_train %>%
recipe(DeadDist_sqrt ~ TreeDiam + Infest_Serv1 + SDI_20th + IND_BA_Infest_20th) %>%
step_corr(all_predictors()) %>% # gets rid of variables to get rid of multicollinearity
step_zv(all_numeric(), -all_outcomes()) %>%
step_scale(all_predictors(), -all_outcomes()) # gets rid of variables of all one value
```
```{r randomforestengine}
forest_model <- rand_forest(trees = 100, mode = "regression") %>%
set_engine("randomForest")
```
```{r forestworkflow}
# workflow
forest_wflow <-
workflow() %>%
add_model(forest_model) %>%
add_recipe(forest_recipe)
```
```{r forestfit, include = FALSE}
forest_wflow %>%
fit(data = pine_train)
lm(DeadDist_sqrt ~ TreeDiam + Infest_Serv1 + SDI_20th + IND_BA_Infest_20th, data = pine_tbl_sqrt) %>%
glance()
```
Row {data-height=450}
-----------------------------------------------------------------------
### Actual vs Predicted Graph for Random Forest Model
```{r forestrmse, include = FALSE}
forest_predicted <- forest_wflow %>%
fit(data = pine_train) %>%
predict(pine_test) %>%
bind_cols(pine_test)
forest_predicted %>%
metrics(truth = DeadDist_sqrt, estimate = .pred)
```
```{r actualvspredicted}
forest_predicted %>%
ggplot(aes(x = DeadDist_sqrt, y = .pred, color = factor(IND_BA_Infest_20th,
levels = 0:1,
labels = c("No", "Yes")))) +
geom_point(alpha = 0.5) +
labs(y = "Predicted Minimum Distance to an Infected Tree (sqrt)", x = "Actual Minimum Distance to an Infected Tree", color = "Infested tree within 1/20th acres?") +
coord_obs_pred() +
scale_fill_brewer(palette = "Dark2") +
scale_color_brewer(palette = "Dark2")
```
Conclusion
==============================================================================
Row {data-height=250}
-----------------------------------------------------------------------
### Explaination
Overall, the random forest model had a lower rmse and a higher rsq. Both are indicators of a more functional model, so I will say the random forest model is more accurate.
Row {data-height=550}
-----------------------------------------------------------------------
### Model Results in Graph Form
```{r compare}
metric_table <- forest_predicted %>%
metrics(truth = DeadDist_sqrt, estimate = .pred) %>%
mutate(forest_estimate = .estimate) %>%
select(-.estimate, -.estimator) %>%
bind_cols(ridge_metrics) %>%
mutate(ridge_estimate = .estimate) %>%
select(-.estimate, -.estimator, -.metric...3)
metric_table %>%
pivot_longer(
cols = !.metric...1,
names_to = "model_type",
values_to = "value") %>%
filter(.metric...1 == "rmse") %>%
ggplot(aes(model_type, value, fill = model_type)) +
geom_bar(stat = "identity") +
ylim(0, 11.1783) +
scale_fill_brewer(palette = "Dark2")
```
### Model Results in Table Form
```{r resultstable}
metric_table %>%
mutate(metric = .metric...1) %>%
select(metric, forest_estimate, ridge_estimate, -.metric...1) %>%
DT::datatable()
```
Variable Chart
==============================================================================
### Picture of variables
<img src="../img/variables.png" width="200" height="100">