In TESTgroup-BNL/PLSR_for_plant_trait_prediction: A simple add-on package to aid in the fitting of leaf or canopy scale spectra-trait PLSR models

knitr::opts_chunk$set(echo = TRUE)

Overview

This is an R Markdown Notebook to illustrate how to retrieve a dataset from the EcoSIS spectral database, choose the "optimal" number of plsr components, and fit a plsr model for leaf-mass area (LMA)

Getting Started

Step 1. Load libraries needed to run example script

list.of.packages <- c("pls","dplyr","here","plotrix","ggplot2","gridExtra","spectratrait")
invisible(lapply(list.of.packages, library, character.only = TRUE))

Step 2. Setup other functions and options

### Setup options

# Script options
pls::pls.options(plsralg = "oscorespls")
pls::pls.options("plsralg")

# Default par options
opar <- par(no.readonly = T)

# What is the target variable?
inVar <- "LMA_g_m2"

# What is the source dataset from EcoSIS?
ecosis_id <- "9db4c5a2-7eac-4e1e-8859-009233648e89"

# Specify output directory, output_dir 
# Options: 
# tempdir - use a OS-specified temporary directory 
# user defined PATH - e.g. "~/scratch/PLSR"
output_dir <- "tempdir"

Step 3. Set working directory (scratch space)

if (output_dir=="tempdir") {
  outdir <- tempdir()
} else {
  if (! file.exists(output_dir)) dir.create(output_dir,recursive=TRUE)
  outdir <- file.path(path.expand(output_dir))
}
setwd(outdir) # set working directory
getwd()  # check wd

Step 4. Pull example dataset from EcoSIS (ecosis.org)

print(paste0("Output directory: ",getwd()))  # check wd
### Get source dataset from EcoSIS
dat_raw <- spectratrait::get_ecosis_data(ecosis_id = ecosis_id)
head(dat_raw)
names(dat_raw)[1:40]

Step 5. Create full plsr dataset

### Create plsr dataset
Start.wave <- 500
End.wave <- 2400
wv <- seq(Start.wave,End.wave,1)
Spectra <- as.matrix(dat_raw[,names(dat_raw) %in% wv])
colnames(Spectra) <- c(paste0("Wave_",wv))
sample_info <- dat_raw[,names(dat_raw) %notin% seq(350,2500,1)]
head(sample_info)

sample_info2 <- sample_info %>%
  select(Plant_Species=`Latin Species`,Species_Code=`species code`,Plot=`plot code`,
         LMA_g_cm2=`Leaf mass per area (g/cm2)`)
sample_info2 <- sample_info2 %>%
  mutate(LMA_g_m2=LMA_g_cm2*10000)
head(sample_info2)

plsr_data <- data.frame(sample_info2,Spectra)
rm(sample_info,sample_info2,Spectra)

Step 6. Example data cleaning.

#### Example data cleaning.  End user needs to do what's appropriate for their 
#### data.  This may be an iterative process.
# Keep only complete rows of inVar and spec data before fitting
plsr_data <- plsr_data[complete.cases(plsr_data[,names(plsr_data) %in% 
                                                  c(inVar,paste0("Wave_",wv))]),]

Step 7. Create cal/val datasets

method <- "dplyr" #base/dplyr
# base R - a bit slow
# dplyr - much faster
split_data <- spectratrait::create_data_split(dataset=plsr_data, approach=method, 
                                              split_seed=7529075, prop=0.8, 
                                              group_variables="Species_Code")
names(split_data)
cal.plsr.data <- split_data$cal_data
head(cal.plsr.data)[1:8]
val.plsr.data <- split_data$val_data
head(val.plsr.data)[1:8]
rm(split_data)

# Datasets:
print(paste("Cal observations: ",dim(cal.plsr.data)[1],sep=""))
print(paste("Val observations: ",dim(val.plsr.data)[1],sep=""))

text_loc <- c(max(hist(cal.plsr.data[,paste0(inVar)], plot=FALSE)$counts),
              max(hist(cal.plsr.data[,paste0(inVar)], plot=FALSE)$mids))
cal_hist_plot <- ggplot(data = cal.plsr.data, 
                        aes(x = cal.plsr.data[,paste0(inVar)])) + 
  geom_histogram(fill=I("grey50"),col=I("black"),alpha=I(.7)) + 
  labs(title=paste0("Calibration Histogram for ",inVar), x = paste0(inVar), 
       y = "Count") + annotate("text", x=text_loc[2], y=text_loc[1], 
                               label= "1.",size=10)
val_hist_plot <- ggplot(data = val.plsr.data, 
                        aes(x = val.plsr.data[,paste0(inVar)])) +
  geom_histogram(fill=I("grey50"),col=I("black"),alpha=I(.7)) + 
  labs(title=paste0("Validation Histogram for ",inVar), x = paste0(inVar), 
       y = "Count")
histograms <- grid.arrange(cal_hist_plot, val_hist_plot, ncol=2)

# Figure S1. The resulting leaf mass area (LMA, g/m2) distribution (histogram) for the 
# calibration (i.e. model training) and validation datasets. The data was split using 
# the spectratrait::create_data_split() function using "Species_Code" as the 
# group_variable and using a data split proportion per group of 80% to calibration 
# and 20% to validation

ggsave(filename = file.path(outdir,paste0(inVar,"_Cal_Val_Histograms.png")), 
       plot = histograms, device="png", width = 30, height = 12, units = "cm", 
       dpi = 300)
# output cal/val data
write.csv(cal.plsr.data,file=file.path(outdir,paste0(inVar,'_Cal_PLSR_Dataset.csv')),
          row.names=FALSE)
write.csv(val.plsr.data,file=file.path(outdir,paste0(inVar,'_Val_PLSR_Dataset.csv')),
          row.names=FALSE)

Step 8. Create calibration and validation PLSR datasets

### Format PLSR data for model fitting 
cal_spec <- as.matrix(cal.plsr.data[, which(names(cal.plsr.data) %in% 
                                              paste0("Wave_",wv))])
cal.plsr.data <- data.frame(cal.plsr.data[, which(names(cal.plsr.data) %notin% 
                                                    paste0("Wave_",wv))], 
                            Spectra=I(cal_spec))
head(cal.plsr.data)[1:5]

val_spec <- as.matrix(val.plsr.data[, which(names(val.plsr.data) %in% 
                                              paste0("Wave_",wv))])
val.plsr.data <- data.frame(val.plsr.data[, which(names(val.plsr.data) %notin% 
                                                    paste0("Wave_",wv))],
                            Spectra=I(val_spec))
head(val.plsr.data)[1:5]

Step 9. Calibration and Validation spectra plot

par(mfrow=c(1,2)) # B, L, T, R
spectratrait::f.plot.spec(Z=cal.plsr.data$Spectra,wv=wv,
            plot_label="Calibration")
text(550,95,labels = "2.",cex=3)
spectratrait::f.plot.spec(Z=val.plsr.data$Spectra,wv=wv,
            plot_label="Validation")

# Figure S2. The resulting calibration and validation spectral reflectance distribution by
# wavelength. The spectra split was done at the same time as LMA, as described in
# Supplemental Figure S1.

dev.copy(png,file.path(outdir,paste0(inVar,'_Cal_Val_Spectra.png')), 
         height=2500,width=4900, res=340)
dev.off();
par(mfrow=c(1,1))

Step 10. Use permutation to determine the optimal number of components

### Use permutation to determine the optimal number of components
if(grepl("Windows", sessionInfo()$running)){
  pls.options(parallel = NULL)
} else {
  pls.options(parallel = parallel::detectCores()-1)
}

method <- "firstMin" #pls, firstPlateau, firstMin
random_seed <- 7529075
seg <- 80
maxComps <- 16
iterations <- 50
prop <- 0.70
if (method=="pls") {
  nComps <- spectratrait::find_optimal_components(dataset=cal.plsr.data, targetVariable=inVar,
                                                  method=method, 
                                                  maxComps=maxComps, seg=seg, 
                                                  random_seed=random_seed)
  print(paste0("*** Optimal number of components: ", nComps))
} else {
  nComps <- spectratrait::find_optimal_components(dataset=cal.plsr.data, targetVariable=inVar,
                                                  method=method, 
                                                  maxComps=maxComps, iterations=iterations, 
                                                  seg=seg, prop=prop, 
                                                  random_seed=random_seed)
}
# Figure S3. Selection of the optimal number of components based on the 
# minimization of the PRESS statistic.  In this example we show "firstMin" 
# option that selects the number of components corresponding to the first 
# statistical minimum PRESS value (vertical broken blue line).

dev.copy(png,file.path(outdir,paste0(paste0("Figure_3_",inVar,
                                            "_PLSR_Component_Selection.png"))), 
         height=2800, width=3400,  res=340)
dev.off();

Step 11. Fit final model

### Fit final model - using leave-one-out cross validation
plsr.out <- plsr(as.formula(paste(inVar,"~","Spectra")),scale=FALSE,ncomp=nComps,
                 validation="LOO",trace=FALSE,data=cal.plsr.data)
fit <- plsr.out$fitted.values[,1,nComps]
pls.options(parallel = NULL)

# External validation fit stats
text_loc <- c(max(RMSEP(plsr.out, newdata = val.plsr.data)$comps),
              RMSEP(plsr.out, newdata = val.plsr.data)$val[1])
par(mfrow=c(1,2)) # B, L, T, R
pls::RMSEP(plsr.out, newdata = val.plsr.data)
plot(pls::RMSEP(plsr.out,estimate=c("test"),newdata = val.plsr.data), main="MODEL RMSEP",
     xlab="Number of Components",ylab="Model Validation RMSEP",lty=1,col="black",cex=1.5,lwd=2)
text(text_loc[1],text_loc[2],labels = "4.", cex=2)
box(lwd=2.2)

pls::R2(plsr.out, newdata = val.plsr.data)
plot(pls::R2(plsr.out,estimate=c("test"),newdata = val.plsr.data), main="MODEL R2",
     xlab="Number of Components",ylab="Model Validation R2",lty=1,col="black",cex=1.5,lwd=2)
box(lwd=2.2)
# Figure S4. A plot of the validation root mean square error of prediction (RMSEP, left) 
# and coefficient of determination (right) for the 0 to optimal number of components

dev.copy(png,file.path(outdir,paste0(paste0(inVar,"_Validation_RMSEP_R2_by_Component.png"))), 
         height=2800, width=4800,  res=340)
dev.off();
par(opar)

Step 12. PLSR fit observed vs. predicted plot data

#calibration
cal.plsr.output <- data.frame(cal.plsr.data[, which(names(cal.plsr.data) %notin% 
                                                      "Spectra")],
                              PLSR_Predicted=fit,
                              PLSR_CV_Predicted=as.vector(plsr.out$validation$pred[,,
                                                                                   nComps]))
cal.plsr.output <- cal.plsr.output %>%
  mutate(PLSR_CV_Residuals = PLSR_CV_Predicted-get(inVar))
head(cal.plsr.output)
cal.R2 <- round(pls::R2(plsr.out,intercept=F)[[1]][nComps],2)
cal.RMSEP <- round(sqrt(mean(cal.plsr.output$PLSR_CV_Residuals^2)),2)

val.plsr.output <- data.frame(val.plsr.data[, which(names(val.plsr.data) %notin% 
                                                      "Spectra")],
                              PLSR_Predicted=as.vector(predict(plsr.out, 
                                                               newdata = val.plsr.data, 
                                                               ncomp=nComps, 
                                                               type="response")[,,1]))
val.plsr.output <- val.plsr.output %>%
  mutate(PLSR_Residuals = PLSR_Predicted-get(inVar))
head(val.plsr.output)
val.R2 <- round(pls::R2(plsr.out,newdata=val.plsr.data,intercept=F)[[1]][nComps],2)
val.RMSEP <- round(sqrt(mean(val.plsr.output$PLSR_Residuals^2)),2)

rng_quant <- quantile(cal.plsr.output[,inVar], probs = c(0.001, 0.999))
cal_scatter_plot <- ggplot(cal.plsr.output, aes(x=PLSR_CV_Predicted, y=get(inVar))) + 
  theme_bw() + geom_point() + geom_abline(intercept = 0, slope = 1, color="dark grey", 
                                          linetype="dashed", linewidth=1.5) + 
  xlim(rng_quant[1], rng_quant[2]) + 
  ylim(rng_quant[1], rng_quant[2]) +
  labs(x=paste0("Predicted ", paste(inVar), " (units)"),
       y=paste0("Observed ", paste(inVar), " (units)"),
       title=paste0("Calibration: ", paste0("Rsq = ", cal.R2), "; ", 
                    paste0("RMSEP = ", cal.RMSEP))) +
  theme(axis.text=element_text(size=18), legend.position="none",
        axis.title=element_text(size=20, face="bold"), 
        axis.text.x = element_text(angle = 0,vjust = 0.5),
        panel.border = element_rect(linetype = "solid", fill = NA, linewidth=1.5)) + 
  annotate("text", x=rng_quant[1], y=rng_quant[2], label= "5.",size=10)

cal_resid_histogram <- ggplot(cal.plsr.output, aes(x=PLSR_CV_Residuals)) +
  geom_histogram(alpha=.5, position="identity") + 
  geom_vline(xintercept = 0, color="black", 
             linetype="dashed", linewidth=1) + theme_bw() + 
  theme(axis.text=element_text(size=18), legend.position="none",
        axis.title=element_text(size=20, face="bold"), 
        axis.text.x = element_text(angle = 0,vjust = 0.5),
        panel.border = element_rect(linetype = "solid", fill = NA, linewidth=1.5))

rng_quant <- quantile(val.plsr.output[,inVar], probs = c(0.001, 0.999))
val_scatter_plot <- ggplot(val.plsr.output, aes(x=PLSR_Predicted, y=get(inVar))) + 
  theme_bw() + geom_point() + geom_abline(intercept = 0, slope = 1, color="dark grey", 
                                          linetype="dashed", linewidth=1.5) + 
  xlim(rng_quant[1], rng_quant[2]) + 
  ylim(rng_quant[1], rng_quant[2]) +
  labs(x=paste0("Predicted ", paste(inVar), " (units)"),
       y=paste0("Observed ", paste(inVar), " (units)"),
       title=paste0("Validation: ", paste0("Rsq = ", val.R2), "; ", 
                    paste0("RMSEP = ", val.RMSEP))) +
  theme(axis.text=element_text(size=18), legend.position="none",
        axis.title=element_text(size=20, face="bold"), 
        axis.text.x = element_text(angle = 0,vjust = 0.5),
        panel.border = element_rect(linetype = "solid", fill = NA, linewidth=1.5))

val_resid_histogram <- ggplot(val.plsr.output, aes(x=PLSR_Residuals)) +
  geom_histogram(alpha=.5, position="identity") + 
  geom_vline(xintercept = 0, color="black", 
             linetype="dashed", linewidth=1) + theme_bw() + 
  theme(axis.text=element_text(size=18), legend.position="none",
        axis.title=element_text(size=20, face="bold"), 
        axis.text.x = element_text(angle = 0,vjust = 0.5),
        panel.border = element_rect(linetype = "solid", fill = NA, linewidth=1.5))

# plot cal/val side-by-side
scatterplots <- grid.arrange(cal_scatter_plot, val_scatter_plot, cal_resid_histogram, 
                             val_resid_histogram, nrow=2, ncol=2)
# Figure S5. The calibration model and independent validation scatter plot results for 
# the example LMA PLSR model (top row). Also shown are the calibration model and 
# validation PLSR residuals, where the calibration results are based on the internal 
# model cross-validation and the validation residuals are the predicted minus observed 
# values of LMA.

ggsave(filename = file.path(outdir,paste0(inVar,"_Cal_Val_Scatterplots.png")), 
       plot = scatterplots, device="png", width = 32, height = 30, units = "cm",
       dpi = 300)

Step 13. Generate Coefficient and VIP plots

vips <- spectratrait::VIP(plsr.out)[nComps,]

par(mfrow=c(2,1))
plot(plsr.out, plottype = "coef",xlab="Wavelength (nm)",
     ylab="Regression coefficients",legendpos = "bottomright",
     ncomp=nComps,lwd=2)
legend("topleft",legend = "6.", cex=2, bty="n")
box(lwd=2.2)
plot(seq(Start.wave,End.wave,1),vips,xlab="Wavelength (nm)",ylab="VIP",cex=0.01)
lines(seq(Start.wave,End.wave,1),vips,lwd=3)
abline(h=0.8,lty=2,col="dark grey")
box(lwd=2.2)
# Figure S6. The calibration model PLSR regression coefficient (top) and variable 
# importance of projection (bottom) plots

dev.copy(png,file.path(outdir,paste0(inVar,'_Coefficient_VIP_plot.png')), 
         height=3100, width=4100, res=340)
dev.off();

Step 14. Permutation analysis to derive uncertainty estimates

if(grepl("Windows", sessionInfo()$running)){
  pls.options(parallel =NULL)
} else {
  pls.options(parallel = parallel::detectCores()-1)
}

jk.plsr.out <- pls::plsr(as.formula(paste(inVar,"~","Spectra")), scale=FALSE, 
                         center=TRUE, ncomp=nComps, validation="LOO", trace=FALSE, 
                         jackknife=TRUE, 
                         data=cal.plsr.data)
pls.options(parallel = NULL)

Jackknife_coef <- spectratrait::f.coef.valid(plsr.out = jk.plsr.out, 
                                             data_plsr = cal.plsr.data, 
                                             ncomp = nComps, inVar=inVar)
Jackknife_intercept <- Jackknife_coef[1,,,]
Jackknife_coef <- Jackknife_coef[2:dim(Jackknife_coef)[1],,,]

interval <- c(0.025,0.975)
Jackknife_Pred <- val.plsr.data$Spectra %*% Jackknife_coef + 
  matrix(rep(Jackknife_intercept, length(val.plsr.data[,inVar])), byrow=TRUE, 
         ncol=length(Jackknife_intercept))
Interval_Conf <- apply(X = Jackknife_Pred, MARGIN = 1, FUN = quantile, 
                       probs=c(interval[1], interval[2]))
sd_mean <- apply(X = Jackknife_Pred, MARGIN = 1, FUN =sd)
sd_res <- sd(val.plsr.output$PLSR_Residuals)
sd_tot <- sqrt(sd_mean^2+sd_res^2)
val.plsr.output$LCI <- Interval_Conf[1,]
val.plsr.output$UCI <- Interval_Conf[2,]
val.plsr.output$LPI <- val.plsr.output$PLSR_Predicted-1.96*sd_tot
val.plsr.output$UPI <- val.plsr.output$PLSR_Predicted+1.96*sd_tot
head(val.plsr.output)

### Permutation coefficient plot
spectratrait::f.plot.coef(Z = t(Jackknife_coef), wv = wv, 
            plot_label="Jackknife regression coefficients",position = 'bottomleft')
abline(h=0,lty=2,col="grey50")
legend("topleft",legend = "7.", cex=2, bty="n")
box(lwd=2.2)
# Figure S7. The calibration model jackknife PLSR regression coefficients 

dev.copy(png,file.path(outdir,paste0(inVar,'_Jackknife_Regression_Coefficients.png')), 
         height=2100, width=3800, res=340)
dev.off();

### Permutation validation plot
rmsep_percrmsep <- spectratrait::percent_rmse(plsr_dataset = val.plsr.output, 
                                              inVar = inVar, 
                                              residuals = val.plsr.output$PLSR_Residuals, 
                                              range="full")
RMSEP <- rmsep_percrmsep$rmse
perc_RMSEP <- rmsep_percrmsep$perc_rmse
r2 <- round(pls::R2(plsr.out, newdata = val.plsr.data,intercept=F)$val[nComps],2)
expr <- vector("expression", 3)
expr[[1]] <- bquote(R^2==.(r2))
expr[[2]] <- bquote(RMSEP==.(round(RMSEP,2)))
expr[[3]] <- bquote("%RMSEP"==.(round(perc_RMSEP,2)))
rng_vals <- c(min(val.plsr.output$LPI), max(val.plsr.output$UPI))
par(mfrow=c(1,1), mar=c(4.2,5.3,1,0.4), oma=c(0, 0.1, 0, 0.2))
plotrix::plotCI(val.plsr.output$PLSR_Predicted,val.plsr.output[,inVar], 
       li=val.plsr.output$LPI, ui=val.plsr.output$UPI, gap=0.009,sfrac=0.004, 
       lwd=1.6, xlim=c(rng_vals[1], rng_vals[2]), ylim=c(rng_vals[1], rng_vals[2]), 
       err="x", pch=21, col="black", pt.bg=scales::alpha("grey70",0.7), scol="grey50",
       cex=2, xlab=paste0("Predicted ", paste(inVar), " (units)"),
       ylab=paste0("Observed ", paste(inVar), " (units)"),
       cex.axis=1.5,cex.lab=1.8)
abline(0,1,lty=2,lw=2)
legend("topleft", legend=expr, bty="n", cex=1.5)
legend("bottomright", legend="8.", bty="n", cex=2.2)
box(lwd=2.2)
# Figure S8. Independent validation results for the LMA PLSR model with associated 
# jackknife uncertainty estimate 95% prediction intervals for each estimate LMA 
# value. The %RMSEP is the model prediction performance standardized to the 
# percentage of the response range, in this case the range of LMA values

dev.copy(png,file.path(outdir,paste0(inVar,"_PLSR_Validation_Scatterplot.png")), 
         height=2800, width=3200,  res=340)
dev.off();

Step 15. Output permutation coefficients for later use

out.jk.coefs <- data.frame(Iteration=seq(1,length(Jackknife_intercept),1),
                           Intercept=Jackknife_intercept,t(Jackknife_coef))
head(out.jk.coefs)[1:6]
write.csv(out.jk.coefs,file=file.path(outdir,
                                      paste0(inVar,
                                             '_Jackkife_PLSR_Coefficients.csv')),
          row.names=FALSE)

Step 16. Output remaining core PLSR outputs

print(paste("Output directory: ", outdir))

# Observed versus predicted
write.csv(cal.plsr.output,file=file.path(outdir,
                                         paste0(inVar,'_Observed_PLSR_CV_Pred_',
                                                nComps,'comp.csv')),
          row.names=FALSE)

# Validation data
write.csv(val.plsr.output,file=file.path(outdir,
                                         paste0(inVar,'_Validation_PLSR_Pred_',
                                                nComps,'comp.csv')),
          row.names=FALSE)

# Model coefficients
coefs <- coef(plsr.out,ncomp=nComps,intercept=TRUE)
write.csv(coefs,file=file.path(outdir,
                               paste0(inVar,'_PLSR_Coefficients_',
                                      nComps,'comp.csv')),
          row.names=TRUE)

# PLSR VIP
write.csv(vips,file=file.path(outdir,
                              paste0(inVar,'_PLSR_VIPs_',
                                     nComps,'comp.csv')))

Step 17. Confirm files were written to temp space

print("**** PLSR output files: ")
print(list.files(outdir)[grep(pattern = inVar, list.files(outdir))])

TESTgroup-BNL/PLSR_for_plant_trait_prediction documentation built on Feb. 15, 2025, 2:08 p.m.