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 develop pixel-scale spectra-trait PLSR models. This example uses image data from NEON AOP and associated field measurements of leaf nitrogen content collected across a range of CONUS NEON sites. For more information refer to the dataset EcoSIS page: https://ecosis.org/package/canopy-spectra-to-map-foliar-functional-traits-over-neon-domains-in-eastern-united-states

Getting Started

Load libraries

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

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? What is the variable name in the input dataset?
inVar <- "Nitrogen"

# What is the source dataset from EcoSIS?
ecosis_id <- "b9dbf3db-5b9c-4ab2-88c2-26c8b39d0903"

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

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
print(getwd())  # check wd

Grab data from EcoSIS

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

Create full plsr dataset

# identify the trait data and other metadata
sample_info <- dat_raw[,names(dat_raw) %notin% seq(300,2600,1)]
head(sample_info)

# spectra matrix
Spectra <- as.matrix(dat_raw[,names(dat_raw) %notin% names(sample_info)])

# set the desired spectra wavelength range to include
Start.wave <- 500
End.wave <- 2400
wv <- seq(Start.wave,End.wave,1)
final_spec <- Spectra[,round(as.numeric(colnames(Spectra))) %in% wv]
colnames(final_spec) <- c(paste0("Wave_",colnames(final_spec)))

## Drop bad spectra data - for canopy-scale reflectance, often the "water band" wavelengths
## are too noisy to use for trait estimation.  Its possible to remove these wavelengths
## prior to model fitting. Its best to first identify which wavelengths to drop
## before attempting PLSR, as these ranges may need to be considered on a case-by-case 
## basis or generalized for multiple datasets
dropwaves <- c(1350:1440, 1826:1946)
final_spec <- final_spec[,colnames(final_spec) %notin% paste0("Wave_",dropwaves)]
wv <- as.numeric(gsub(pattern = "Wave_",replacement = "", x = colnames(final_spec)))

## Drop bad spectra data - for canopy-scale reflectance, often the "water band" wavelengths
## are too noisy to use for trait estimation.  Its possible to remove these wavelengths
## prior to model fitting. Its best to first identify which wavelengths to drop
## before attempting PLSR, as these ranges may need to be considered on a case-by-case 
## basis or generalized for multiple datasets
dropwaves <- c(1350:1440, 1826:1946)
final_spec <- final_spec[,colnames(final_spec) %notin% paste0("Wave_",dropwaves)]
wv <- as.numeric(gsub(pattern = "Wave_",replacement = "", x = colnames(final_spec)))

# assemble example dataset
sample_info2 <- sample_info %>%
  select(Plot_ID,Sample_Year,SLA,Nitrogen)
site_plot <- data.frame(matrix(unlist(strsplit(sample_info2$Plot_ID, "_")), 
                               ncol=2, byrow=TRUE))
colnames(site_plot) <- c("Plot_Num","SampleID")
sample_info3 <- data.frame(site_plot,sample_info2)

plsr_data <- data.frame(sample_info3,final_spec*0.01)
rm(sample_info,sample_info2,sample_info3,Spectra, site_plot)

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 %>%  # remove erroneously high values, or "bad spectra"
  filter(Nitrogen<50) %>%
  filter(Wave_859<80) %>%
  filter(Wave_859>15)
plsr_data <- plsr_data[complete.cases(plsr_data[,names(plsr_data) %in% 
                                                  c(inVar,paste0("Wave_",wv))]),]

Create cal/val datasets

## Make a stratified random sampling in the strata USDA_Species_Code and Domain

method <- "base" #base/dplyr
# base R - a bit slow
# dplyr - much faster
split_data <- spectratrait::create_data_split(dataset=plsr_data, approach=method, split_seed=2356326,
                                              prop=0.8, group_variables="Plot_Num")
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=""))

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")
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)
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)

Create calibration and validation PLSR datasets

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]

plot cal and val spectra

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

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))

Use permutation to determine optimal number of components

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

method <- "pls" #pls, firstPlateau, firstMin
random_seed <- 1245565
seg <- 50
maxComps <- 16
iterations <- 80
prop <- 0.70
if (method=="pls") {
  # pls package approach - faster but estimates more components....
  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)
}
dev.copy(png,file.path(outdir,paste0(paste0(inVar,"_PLSR_Component_Selection.png"))), 
         height=2800, width=3400,  res=340)
dev.off();

Fit final model

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
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)
box(lwd=2.2)

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)
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)

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))

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

Generate Coefficient and VIP plots

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

par(mfrow=c(2,1))
plot(plsr.out$coefficients[,,nComps], x=wv,xlab="Wavelength (nm)",
     ylab="Regression coefficients",lwd=2,type='l')
box(lwd=2.2)
plot(wv, vips, xlab="Wavelength (nm)",ylab="VIP",cex=0.01)
lines(wv, vips, lwd=3)
abline(h=0.8, lty=2, col="dark grey")
box(lwd=2.2)
dev.copy(png, file.path(outdir, paste0(inVar,'_Coefficient_VIP_plot.png')), 
         height=3100, width=4100, res=340)
dev.off();
par(opar)

Bootstrap validation

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

### PLSR bootstrap permutation uncertainty analysis
iterations <- 500    # how many permutation iterations to run
prop <- 0.70          # fraction of training data to keep for each iteration
plsr_permutation <- spectratrait::pls_permutation(dataset=cal.plsr.data, targetVariable=inVar,
                                                  maxComps=nComps, 
                                                  iterations=iterations, prop=prop,
                                                  verbose = FALSE)
bootstrap_intercept <- plsr_permutation$coef_array[1,,nComps]
bootstrap_coef <- plsr_permutation$coef_array[2:length(plsr_permutation$coef_array[,1,nComps]),
                                              ,nComps]
rm(plsr_permutation)

# apply coefficients to left-out validation data
interval <- c(0.025,0.975)
Bootstrap_Pred <- val.plsr.data$Spectra %*% bootstrap_coef + 
  matrix(rep(bootstrap_intercept, length(val.plsr.data[,inVar])), byrow=TRUE, 
         ncol=length(bootstrap_intercept))
Interval_Conf <- apply(X = Bootstrap_Pred, MARGIN = 1, FUN = quantile, 
                       probs=c(interval[1], interval[2]))
sd_mean <- apply(X = Bootstrap_Pred, MARGIN = 1, FUN = sd)
sd_res <- sd(val.plsr.output$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)

Jackknife coefficient plot

spectratrait::f.plot.coef(Z = t(bootstrap_coef), wv = wv, 
                          plot_label="Bootstrap regression coefficients", 
                          position = 'bottomleft')
abline(h=0,lty=2,col="grey50")
box(lwd=2.2)
dev.copy(png,file.path(outdir,paste0(inVar,'_Bootstrap_Regression_Coefficients.png')), 
         height=2100, width=3800, res=340)
dev.off();

Bootstrap 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$UPI, gap=0.009,sfrac=0.000, 
                lwd=1.6, xlim=c(rng_vals[1], rng_vals[2]), ylim=c(rng_vals[1], rng_vals[2]), 
                err="x", pch=21, col="black", pt.bg=scales::alpha("grey70",0.7), scol="grey80",
                cex=2, xlab=paste0("Predicted ", paste(inVar), " (units)"),
                ylab=paste0("Observed ", paste(inVar), " (units)"),
                cex.axis=1.5,cex.lab=1.8)
abline(0,1,lty=2,lw=2)
plotrix::plotCI(val.plsr.output$PLSR_Predicted,val.plsr.output[,inVar], 
                li=val.plsr.output$LCI, ui=val.plsr.output$UCI, 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="black",
                cex=2, xlab=paste0("Predicted ", paste(inVar), " (units)"),
                ylab=paste0("Observed ", paste(inVar), " (units)"),
                cex.axis=1.5,cex.lab=1.8, add=T)
legend("topleft", legend=expr, bty="n", cex=1.5)
legend("bottomright", legend=c("Prediction Interval","Confidence Interval"), 
       lty=c(1,1), col = c("grey80","black"), lwd=3, bty="n", cex=1.5)
box(lwd=2.2)
dev.copy(png,file.path(outdir,paste0(inVar,"_PLSR_Validation_Scatterplot.png")), 
         height=2800, width=3200,  res=340)
dev.off();

Output bootstrap results

out.jk.coefs <- data.frame(Iteration=seq(1,length(bootstrap_intercept),1),
                           Intercept=bootstrap_intercept,t(bootstrap_coef))
names(out.jk.coefs) <- c("Iteration","Intercept",paste0("Wave_",wv))
head(out.jk.coefs)[1:6]
write.csv(out.jk.coefs,file=file.path(outdir,paste0(inVar,'_Bootstrap_PLSR_Coefficients.csv')),
          row.names=FALSE)

Create core PLSR outputs

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

# 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')))

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.