Random Forest - Climate Data Analysis
Contributors: Joey Hulbert,…
2024-03-01
| Redcedar | Data | Analyses | Instructions |
Please note this analysis and R Markdown document are in still in development :)
View a previous, archived, version of this analysis here.
Approach
The overall approach is to model empirical data collected by community scientists with ancillary climate data to identify important predictors of western redcedar dieback.
Data Wrangling
Import iNat Data - Empirical Tree Points (Response variables)
The steps for wrangling the iNat data are described here.
Format and export for collecting climateNA data
Data were subset to include only gps information to use in collecting ancillary data.
Remove iNaturalist columns and explanatory variables not needed for random forest models
Import Normals Data
Climate data then extracted with ClimateNA tool following the below process. Data were downloaded for the iNat GPS locations using the ClimateNA Tool.
ClimateNA version 7.42 -
- Climate data extraction process with ClimateNA
- Convert data into format for climateNA use (see above)
- In ClimateNA
- Normal Data
- Select input file (browse to gps2566 file)
- Choose ‘More Normal Data’
- Select ‘Normal_1991_2020.nrm’
- Choose ‘All variables(265)’
- Specify output file
- Normal Data
- Grouping explored
- data averaged over 30 year normals (1991-2020)
Variables
Note the below analysis uses the iNat data with 1510 observations. Amazing!
- Response variables included in this analysis
- Tree canopy symptoms (binary)
- Explanatory variables included
- Climate data
- 30yr normals 1991-2020 (265 variables - annual, seasonal, monthly)
- Climate data
Remove specific climate variables not useful as explanatory variables (e.g. norm_Latitutde)
Remove Outliers
For some reason there is one observation with a super neg cmi value (-10000 ish)
Seperate climate variable groupings
Normals data for 265 variables were downloaded for each point Monthly - 180 variables represented data averaged over months for the 30 year period Seasonal - 60 variables respresented data averaged over 3 month seasons (4 seasons) for 30 year period Annual - 20 variables represented data averaged for all years during 30 year period
Remove variables with variables that have near zero standard deviations (entire column is same value)
Full
There were length(normals)-length(normals.nearzerovar
monthly variables with zero standard deviation is Dropping columns with
near zero standard deviation removed
length(normals)-length(normals.nearzerovar monthly climate
variables.
Monthly
There were
length(normals.monthly)-length(normals.monthly.nearzerovar
monthly variables with zero standard deviation is Dropping columns with
near zero standard deviation removed
length(normals.monthly)-length(normals.monthly.nearzerovar
monthly climate variables.
Seasonal
There were
length(normals.monthly)-length(normals.seasonal.nearzerovar
monthly variables with zero standard deviation is
Annual
There were
length(normals.monthly)-length(normals.annual.nearzerovar
monthly variables with zero standard deviation.
Join iNat and Climate Data
Remove Dead Trees
Note we chose to remove dead trees after joining with climate data in case we change our mind about that.
Generally, removing dead trees may make the most sense biologically, because we’re not sure about the cause of the dead tree. Later we could test if there is a good climate variable for classifying trees as alive or dead.
Prepare data for random forest models
Remove other explanatory variable categories (binary only)
Compare model errors
Binary Normal Model
Full Normal Model
##
## Call:
## randomForest(formula = binary.tree.canopy.symptoms ~ ., data = binary.full, ntree = 1200, importance = TRUE, proximity = TRUE, na.action = na.omit)
## Type of random forest: classification
## Number of trees: 1200
## No. of variables tried at each split: 14
##
## OOB estimate of error rate: 31.3%
## Confusion matrix:
## Healthy Unhealthy class.error
## Healthy 1143 329 0.2235054
## Unhealthy 430 523 0.4512067Monthly Normal Model
##
## Call:
## randomForest(formula = binary.tree.canopy.symptoms ~ ., data = binary.monthly, ntree = 1200, importance = TRUE, proximity = TRUE, na.action = na.omit)
## Type of random forest: classification
## Number of trees: 1200
## No. of variables tried at each split: 12
##
## OOB estimate of error rate: 31.92%
## Confusion matrix:
## Healthy Unhealthy class.error
## Healthy 1132 340 0.2309783
## Unhealthy 434 519 0.4554040Seasonal Normal Model
##
## Call:
## randomForest(formula = binary.tree.canopy.symptoms ~ ., data = binary.seasonal, ntree = 1200, importance = TRUE, proximity = TRUE, na.action = na.omit)
## Type of random forest: classification
## Number of trees: 1200
## No. of variables tried at each split: 7
##
## OOB estimate of error rate: 31.88%
## Confusion matrix:
## Healthy Unhealthy class.error
## Healthy 1134 338 0.2296196
## Unhealthy 435 518 0.4564533Annual Normal Model
##
## Call:
## randomForest(formula = binary.tree.canopy.symptoms ~ ., data = binary.annual, ntree = 1200, importance = TRUE, proximity = TRUE, na.action = na.omit)
## Type of random forest: classification
## Number of trees: 1200
## No. of variables tried at each split: 4
##
## OOB estimate of error rate: 31.88%
## Confusion matrix:
## Healthy Unhealthy class.error
## Healthy 1138 334 0.2269022
## Unhealthy 439 514 0.4606506Summary of model performance
| Response | Grouping | Num Variables | Vars tried split | OOB Error |
| Binary | Full | 225 | 14 | 31.3 |
| Binary | Monthly | 148 | 12 | 31.92 |
| Binary | Seasonal | 54 | 7 | 31.88 |
| Binary | Annual | 25 | 31.88 | |
Identify important variables
Binary Response, Annual Explanatory Variable
The error rate above may stabilize enough by 600-800 trees. May not be necessary to run 1200 trees.
## num [1:2425, 1:2] -0.02559 0.00402 -0.01364 -0.01961 -0.01828 ...
## - attr(*, "dimnames")=List of 2
## ..$ : chr [1:2425] "1" "2" "3" "4" ...
## ..$ : NULL
Binary Response, Seasonal Explanatory Variable
Clearly all of the climate variables are highly correlated.
Lets pick the top performing metric in our random forests analyses, CMI and then any less correlated variables
Below we can check the correlation of CMI, MAP, and DD_18
Now we can check how the model performs with only these three climate variables
##
## Call:
## randomForest(formula = binary.tree.canopy.symptoms ~ ., data = binary.annual, ntree = 1200, importance = TRUE, proximity = TRUE, na.action = na.omit)
## Type of random forest: classification
## Number of trees: 1200
## No. of variables tried at each split: 4
##
## OOB estimate of error rate: 31.88%
## Confusion matrix:
## Healthy Unhealthy class.error
## Healthy 1138 334 0.2269022
## Unhealthy 439 514 0.4606506It’s hard to give up the seasonality data, but they are all highly correlated (data not shown) and if we look at the above importance plot for the seasonality data, the winter variables (norm_CMI_wt,norm_DD_18_wt, and norm_PPT_wt) all had the highest MeanDecrease Accuracy and Gini. Therefore, even if we chose to build the model on seasonal data, we would likely want to choose to use the winter values for each variable.
GLMMs
Probability of a tree classified as unhealthy
Response variable: category: healthy, unhealthy
Note dead trees were removed
Healthy / Unhealthy
## Family: binomial ( logit )
## Formula: binary.tree.canopy.symptoms ~ norm_CMI
## Data: annual
##
## AIC BIC logLik deviance df.resid
## 3252.8 3264.4 -1624.4 3248.8 2423
##
##
## Conditional model:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -0.4890830 0.0675426 -7.241 4.45e-13 ***
## norm_CMI 0.0008836 0.0008637 1.023 0.306
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1## Warning: Removed 12 rows containing non-finite values (`stat_density()`).
Top Dieback / No Topdieback
## Family: binomial ( logit )
## Formula: top.dieback ~ norm_CMI
## Data: annual
##
## AIC BIC logLik deviance df.resid
## 1808.4 1820.0 -902.2 1804.4 2423
##
##
## Conditional model:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -2.095608 0.098672 -21.238 <2e-16 ***
## norm_CMI 0.002063 0.001182 1.746 0.0808 .
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1## Warning: Removed 12 rows containing non-finite values (`stat_density()`).
Thinning / no-thinning
## Warning: Removed 12 rows containing non-finite values (`stat_density()`).
Dead / Alive
## Family: binomial ( logit )
## Formula: dead.tree ~ norm_CMI
## Data: full.with.dead
##
## AIC BIC logLik deviance df.resid
## 1019.3 1031.0 -507.7 1015.3 2551
##
##
## Conditional model:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -3.011139 0.146278 -20.585 <2e-16 ***
## norm_CMI 0.001112 0.001798 0.619 0.536
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Discussion
Explore if monthly, seasonal, or annual data are best fit for binomial distributed glmm.
Identify which climate variable grouping is best fit then run random forests for determining which climate variable is best for predicting top dieback
then run random forests for predicting which climate variable is best for predicting thinning