Using a Self Organizing Map for Image Analysis

1. Loading raw data and getting it ready

library(readr)
larva2 <- read_csv("~/GENER/SOMs/1. Raw Data/PSOMML_2Larvas.csv", 
    col_names = FALSE)
colnames(larva2) <- c("FILA", "COLUMNA", "RED", "GREEN", "BLUE")

larva2Esp <- larva2
larva2 <- larva2[,-1:-2]

The dataframe obtained is made of integer values:

str(larva2)

This will be a problem down the line when feeding this data to the algorithm, values have to be changed to numeric:

larva2[,'RED'] <- as.numeric(as.factor(larva2$RED))
larva2[,'GREEN'] <- as.numeric(as.factor(larva2$GREEN))
larva2[,'BLUE'] <- as.numeric(as.factor(larva2$BLUE))
###larva2Esp
larva2Esp[,'FILA'] <- as.numeric(as.factor(larva2Esp$FILA))
larva2Esp[,'COLUMNA'] <- as.numeric(as.factor(larva2Esp$COLUMNA))
larva2Esp[,'RED'] <- as.numeric(as.factor(larva2Esp$RED))
larva2Esp[,'GREEN'] <- as.numeric(as.factor(larva2Esp$GREEN))
larva2Esp[,'BLUE'] <- as.numeric(as.factor(larva2Esp$BLUE))

Whit this taken care of we have a database ready to be fed to the algorithmL

str(larva2)

Classes ‘tbl_df’, ‘tbl’ and 'data.frame':   262144 obs. of  3 variables:
 $ RED  : num  166 166 165 164 164 164 164 164 164 164 ...
 $ GREEN: num  166 166 167 167 167 167 167 167 167 167 ...
 $ BLUE : num  170 170 170 170 170 170 170 170 170 170 ...

2 Initial exploration of data

A brief summary of the values of the variable can help with the initial understanding of things:

summary(larva2)

      RED            GREEN            BLUE      
 Min.   :  1.0   Min.   :  1.0   Min.   :  1.0  
 1st Qu.:164.0   1st Qu.:166.0   1st Qu.:169.0  
 Median :166.0   Median :167.0   Median :170.0  
 Mean   :163.3   Mean   :164.6   Mean   :165.5  
 3rd Qu.:167.0   3rd Qu.:168.0   3rd Qu.:171.0  
 Max.   :171.0   Max.   :172.0   Max.   :175.0  

3 Self Organizing Map

Loading required package

require(kohonen)

Changing dataframe to a matrix

#I removed the function scale() because it does not need to be scaled.
larva2.matrix <- as.matrix(larva2)
larva2Esp.matrix <- as.matrix(larva2Esp)
#TEST
#larva2.matrixTest <- larva2.matrix[1:20000,]

Creating the SOM grid

som_grid <- somgrid(xdim = 30, ydim = 30, topo = "rectangular")

Training the SOM

With no spatial references

som_model <- som(larva2.matrix,
                 grid = som_grid,
                 rlen = 1500,
                 alpha = c(0.05,0.01),
                 keep.data = TRUE)

With spatial references

som_modelEsp <- som(larva2Esp.matrix,
                 grid = som_grid,
                 rlen = 1500,
                 alpha = c(0.05,0.01),
                 keep.data = TRUE)

Training progress evaluation

plot(som_modelEsp, type = "changes")

Node counts

source('coolBlueHotRed.R')
plot(som_modelEsp, type = "counts", palette.name=coolBlueHotRed)

SOM by RGB value

par(mfrow = c(2,2))
models <- as.data.frame(som_modelEsp$codes)
plot(som_modelEsp,
     type = "property",
     property = models[,3],
     main=names(models)[3],
     palette.name=coolBlueHotRed)
plot(som_modelEsp,
     type = "property",
     property = models[,4],
     main=names(models)[4],
     palette.name=coolBlueHotRed)

plot(som_modelEsp,
     type = "property",
     property = models[,5],
     main=names(models)[5],
     palette.name=coolBlueHotRed)

Neighbor Distance

plot(som_modelEsp, type = "dist.neighbours", palette.name = grey.colors)

Code Spread

plot(som_modelEsp, type = "codes", palette.name = coolBlueHotRed)

Clustering

mydata <- som_modelEsp$codes 
wss <- (nrow(mydata)-1)*sum(apply(data.frame(mydata),2,var)) 
for (i in 2:20) {
  wss[i] <- sum(kmeans(data.frame(mydata), centers=i)$withinss)
}
plot(1:20, wss, type="b", xlab="Number of Clusters",
     ylab="Within groups sum of squares", main="Within cluster sum of squares (WCSS)")

som_cluster <- cutree(hclust(dist(data.frame(som_modelEsp$codes))), 6)

In this graphic we can see that the ideal amount of clusters is not more than 5. Starting with 5 and going down we will choose the smallest number that gives us a smooth distribution in quantity of it’s members:

# 7 Clusters -- the first cluster has considerably less members than the other four ones.
table(cutree(hclust(dist(data.frame(som_modelEsp$codes))), 3))

  1   2   3 
489 172 239 

# 5 Clusters -- the first cluster has considerably less members than the other four ones.
table(cutree(hclust(dist(data.frame(som_modelEsp$codes))), 5))

  1   2   3   4   5 
132 172 134 239 223 

#4 clusters -- we have founde the good one.
table(cutree(hclust(dist(data.frame(som_modelEsp$codes))), 4))

  1   2   3   4 
266 172 239 223 

pretty_palette <- c("#1f77b4", '#ff7f0e', '#2ca02c', '#d62728', '#9467bd', '#8c564b', '#e377c2')
plot(som_modelEsp, type="mapping", bgcol = pretty_palette[som_cluster], main = "Clusters", keepMargins = TRUE) 
add.cluster.boundaries(som_modelEsp, som_cluster, lwd = 7)