CANYON_pCO2.R

CANYON_pCO2 <- function(date,lat,lon,pres,temp,psal,doxy){
  # function out=CANYON_pCO2(gtime,lat,lon,pres,temp,psal,doxy)
  # 
  # Multi-layer perceptron to predict CO2 partial pressure / uatm 
  #
  # Neural network training by Raphaëlle Sauzède, LOV; 
  # as R function by Henry Bittig, LOV
  #
  #
  # input:
  # gtime - date (UTC) as string ("yyyy-mm-dd HH:MM")
  # lat   - latitude / °N  [-90 90]
  # lon   - longitude / °E [-180 180] or [0 360]
  # pres  - pressure / dbar
  # temp  - in-situ temperature / °C
  # psal  - salinity
  # doxy  - dissolved oxygen / umol kg-1 (!)
  #
  # output:
  # out   - pCO2 / uatm
  #
  # check value: 647.7954 uatm
  # for 09-Dec-2014 08:45, 17.6° N, -24.3° E, 180 dbar, 16 °C, 36.1 psu, 104 umol O2 kg-1
  #
  # reference:
  # Sauzède, R., Bittig, H.C., Claustre, H., Pasqueron de Fommervault, O.,
  # Gattuso, J.-P., Legendre, L., and Johnson, K.S. (2017). 
  # Estimates of water-column nutrient concentrations and carbonate system
  # parameters in the global ocean: A novel approach based on neural networks. 
  # Front. Mar. Sci. 4:128. doi:10.3389/fmars.2017.00128 
  #
  # Henry Bittig, LOV
  # 14.04.2017
  
  # No input checks! Assumes informed use, e.g., same dimensions for all
  # inputs, ...
  
  require(fields)
  require(seacarb)
  
  basedir <- "../CANYON_Training/" # relative or absolute path to CANYON training files
  
  # input preparation
  #day of the year from the date
  #date <- as.POSIXct(date,format="%Y-%m-%d")
  date <- as.POSIXct(date)
  day <- as.numeric(format(date,"%j"))*360/365 # only full yearday used; entire year (365 d) mapped to 360°
  year <- as.numeric(format(date,"%Y"))
  lon[which(lon>180)]=lon[which(lon>180)]-360
  # doy sigmoid scaling
  presgrid <- read.table(paste(basedir,"CY_doy_pres_limit.csv",sep=""))
  pivot_obj <- list(x=presgrid[,1][-1],y=unlist(presgrid[1,][-1]),z=as.matrix(presgrid[-1,-1]))
  prespivot <- interp.surface(pivot_obj,matrix(cbind(lon,lat),ncol=2))
  fsigmoid <- 1/(1+exp((pres-prespivot)/50))
  # input sequence
  #           lat,      sin(lon),    cos(lon),    cos(day),    sin(day),    year,   temp,   sal,    oxygen, P 
  data <- cbind(lat/90,sinpi(lon/180),cospi(lon/180),cospi(day/180)*fsigmoid,sinpi(day/180)*fsigmoid,year,temp,psal,doxy,pres/2e4+1/((1+exp(-pres/300))^3))
  
  temporaire <- read.table(paste(basedir,"Fichier_poids_pCO2_indirect_hidden_ascii18_8.sn",sep=""))
  poids <- temporaire[,3]
  rm(temporaire)
  
  ne=10 # Number of inputs
  nc1=18 # Number of neurons of the first hidden layer
  nc2=8 # Number of neurons of the second hidden layer
  ns=1  # Number of outputs
  
  # WEIGHT AND BIAS PARAMETERS   
  # weight and bias from the input layer to the first hidden layer w1, b1
  b1=matrix(poids[1:nc1],nc1,1);    
  w1=matrix(poids[nc1+nc2+ns+(1:(ne*nc1))],nc1,ne);
  # weight and bias from the first hidden layer to the second hidden layer
  b2=matrix(poids[nc1+(1:nc2)],nc2,1);
  w2=matrix(poids[nc1+nc2+ns+ne*nc1+(1:(nc1*nc2))],nc2,nc1);
  # weight and bias from the second hidden layer to the output layer w3, b3    
  b3=matrix(poids[nc1+nc2+(1:ns)],ns,1);
  w3=matrix(poids[nc1+nc2+ns+ne*nc1+nc1*nc2+(1:(nc2*ns))],ns,nc2);
  
  # Mean and standard deviation of the training dataset
  # These values are used to normalize the inputs parameters
  Moy <- unlist(read.table(paste(basedir,"moy_pCO2_indirect.dat",sep="")),use.names = "FALSE")
  Ecart <- unlist(read.table(paste(basedir,"std_pCO2_indirect.dat",sep="")),use.names = "FALSE")
  
  # NORMALISATION OF THE INPUT PARAMETERS
  rx <-dim(data)[1]
  data_N=(2./3)*(data-t(matrix(Moy[1:ne],ne,1) %*% rep(1,rx)))/t(matrix(Ecart[1:ne],ne,1) %*% rep(1,rx));
  
  # Two hidden layers
  a <- 1.715905*tanh((2./3)*(data_N %*% t(w1)+t(b1 %*% t(rep(1,rx)))))  # input layer to first hidden layer
  b <- 1.715905*tanh((2./3)*(     a %*% t(w2)+t(b2 %*% t(rep(1,rx))))) # first hidden layer to second hidden layer
  y <-                            b %*% t(w3)+t(b3 %*% rep(1,rx)) # second hidden layer to output layer
  # Y is the normalised output value of the neural network, 
  
  # Denormalisation of the output of the NN for getting the true value 
  y_rescaled <- 1.5*y*Ecart[ne+1]+Moy[ne+1]
  
  # recalculate pCO2
  outcalc <- carb(flag=15,var1=0.0023, var2=y_rescaled*1e-6, S=35, T=25, P=0, Patm=1, Pt=0, Sit=0, pHscale="T", kf="pf", k1k2="l", ks="d", b="u74", gas="standard")
  y_rescaled <- outcalc$pCO2

  #% and put into same shape as the input variables
  #out=reshape(y_rescaled,size(pres));
  out <- y_rescaled
  
  return(out)
} # end of function