Calculate and Display dissimilarity between 2 sets of Reflectance spectra

• Agustin.Lobo@geo3bcn.csic.es
• 20231107

library(prospectr)
library(resemble)
library(ggplot2)
library(gridExtra)

1. Load Data

load("IS_Refl.rda") #"observed" spectra
load("G_Refl.rda")  #reference spectra

IS_Refl[1:3, 1:5] #"observed" spectra

                  Reflectance.350 Reflectance.351 Reflectance.352 Reflectance.353 Reflectance.354
IS-03.IS-03_01a.1      0.07266829      0.06986278      0.06873384      0.07570227      0.07288244
IS-03.IS-03_01a.2      0.06574845      0.06434544      0.06420722      0.06560403      0.06362657
IS-03.IS-03_01a.3      0.09143127      0.09247673      0.09320479      0.09260173      0.09185399

G_Refl[1:3, 1:5]  #reference spectra

          Reflectance.350 Reflectance.351 Reflectance.352 Reflectance.353 Reflectance.354
rh01_core      0.06313507      0.06242771      0.06323991      0.06122314      0.06041997
rh01_rim       0.07792549      0.07626121      0.07553060      0.07374593      0.07261300
rh02_core      0.11810216      0.11717834      0.11674443      0.11524004      0.11522657

#Just to get the wavelength values:
load("gsli.lf.rda")
a <- gsli.lf[gsli.lf$Sample=="rh01_core",]
wl <-a$Wavelength
rm(a)

2.Calculate PCA dissimilarity

dsm_pca <- dissimilarity(
  Xr = IS_Refl, Xu = G_Refl,
  diss_method = c("pca"),
  gh = FALSE,
  pc_selection = list("cumvar", 0.99),
  return_projection = TRUE)

Calculate 1st nearest neighbor of each IS_Refl spectrum to each G_Refl spectrum

knn_pc <- search_neighbors(Xr = G_Refl, Xu = IS_Refl,
                           diss_method = c("pca"),
                           k = 1,
                           pc_selection = list("cumvar", 0.99),
                           scale = TRUE)
hist(knn_pc$neighbors_diss)

knn.pc.df <- data.frame(point.index=1:nrow(IS_Refl),
                        point=rownames(IS_Refl),
                        GNN.index=as.vector(knn_pc$neighbors),
                        GNN=rownames(G_Refl)[as.vector(knn_pc$neighbors)],
                        GNN.diss=as.vector(knn_pc$neighbors_diss))

3.Calculate Correlation dissimilarity

knn_c <- search_neighbors(Xr = G_Refl, Xu = IS_Refl,
                          diss_method = c("cor"),
                          k = 1,
                          scale = TRUE)
hist(knn_c$neighbors_diss)

knn.c.df <- data.frame(point.index=1:nrow(IS_Refl),
                       point=rownames(IS_Refl),
                       GNN.index=as.vector(knn_c$neighbors),
                       GNN=rownames(G_Refl)[as.vector(knn_c$neighbors)],
                       GNN.diss=as.vector(knn_c$neighbors_diss))

4.Display couples of most similar spectra

We select the couples with maximum and minimum dissimilarity, and 13 more randomly selected couples for each dissimilarity metric

4.1 PCA dissimilarity

knn.df <- knn.pc.df
metric <- "PCdissim"

knn.sel <- rbind(knn.df[which.min(knn.df$GNN.diss),],
                 knn.df[which.max(knn.df$GNN.diss),],
                 knn.df[sample(1:nrow(knn.df),13),])
knn.sel <- knn.sel[order(knn.sel$GNN.diss),]

l <- list(1:nrow(knn.sel))
for (i in 1:nrow(knn.sel)){
  df= data.frame(Wavelength=wl, 
                 IS.Refl= IS_Refl[knn.sel$point.index[i],],
                 G.Refl= G_Refl[knn.sel$GNN.index[i],])
  g <- ggplot(data=df) +
    geom_line(aes(x=Wavelength,y=IS.Refl),color="red") +
    geom_line(aes(x=Wavelength,y=G.Refl),color="blue") +
    ggtitle(paste(knn.sel$point[i], knn.sel$GNN[i],round(knn.sel$GNN.diss[i],3)))
  l[[i]] <- g
}
grid.arrange(grobs=l,ncol=3,nrow=5)

4.2 Correlation dissimilarity

knn.df <- knn.c.df
metric <- "Correlation"

knn.sel <- rbind(knn.df[which.min(knn.df$GNN.diss),],
                 knn.df[which.max(knn.df$GNN.diss),],
                 knn.df[sample(1:nrow(knn.df),13),])
knn.sel <- knn.sel[order(knn.sel$GNN.diss),]

l <- list(1:nrow(knn.sel))
for (i in 1:nrow(knn.sel)){
  df= data.frame(Wavelength=wl, 
                 IS.Refl= IS_Refl[knn.sel$point.index[i],],
                 G.Refl= G_Refl[knn.sel$GNN.index[i],])
  g <- ggplot(data=df) +
    geom_line(aes(x=Wavelength,y=IS.Refl),color="red") +
    geom_line(aes(x=Wavelength,y=G.Refl),color="blue") +
    ggtitle(paste(knn.sel$point[i], knn.sel$GNN[i],round(knn.sel$GNN.diss[i],3)))
  #print(g)
  l[[i]] <- g
  #readline("Waiting...")
}
grid.arrange(grobs=l,ncol=3,nrow=5)

In both cases, I do not see a clear relationship between dissimiarity values and the visual inspection the couples of spectra.