##################    Interpolation of Tmax for 10 dates.  #######################################

###########################  TWO-STAGE REGRESSION  ###############################################

#This script interpolates station values for the Oregon case study using a two-stage regression. #

#For each input dates, it performs 1) Step 1: General Additive Model (GAM)                       #

#                                  2) Step 2: Kriging on residuals from step 1                   #

#                                                                                                #

#The script uses LST monthly averages as input variables and  loads the station data             # 

#from a shape file with projection information.                                                  #

#Note that this program:                                                                         #

#1)assumes that the shape file is in the current working.                                        # 

#2)extract relevant variables from raster images before performing the regressions.              #

#This scripts predicts tmax using ing GAM and LST derived from MOD11A1.                          #

#Interactions terms are also included and assessed using the RMSE from validation dataset.       #

#There are 10 dates used for the GAM interpolation. The dates must be provided as a textfile.    #

#AUTHOR: Benoit Parmentier                                                                       #

#DATE: 05/09/212                                                                                 #

#PROJECT: NCEAS INPLAN: Environment and Organisms --TASK#364--                                   #

##################################################################################################

#infile2<-"dates_interpolation_03052012_2dates_test.txt"

infile2<-"dates_interpolation_03052012.txt"                                          #List of 10 dates for the regression

infile4<-"models_interpolation_04032012.txt"

infile5<-"mean_day244_rescaled.rst" #Raster or grid for the locations of predictions

#path<-"/data/computer/parmentier/Data/IPLANT_project/data_Oregon_stations"         #Jupiter LOCATION on EOS

path<-"H:/Data/IPLANT_project/data_Oregon_stations"                                 #Jupiter Location on XANDERS

setwd(path) 

prop<-0.3                                                                           #Proportion of testing retained for validation   

seed_number<-100

out_prefix<-"_05062012m_Kr_LST"

filename<-sub(".shp","",infile1)            #Removing the extension from file.

mean_LST<- readGDAL(infile5)  #This reads the whole raster in memory and provide a grid for kriging

results_AIC<- matrix(1,length(dates),length(models)+2)  #Storing diagnostic statistics

  mod_LST <-ghcn.subsets[[i]][,match(LST_month, names(ghcn.subsets[[i]]))]

  ghcn.subsets[[i]] = transform(ghcn.subsets[[i]],LST = mod_LST)

  ####Regression part 2: GAM models (REGRESSION STEP1)

  #listmod can be created and looped over. In this case we loop around the objects..

  for (j in 1:length(models)){

    name<-paste("mod",j,sep="") #modj is the name of he "j" model (mod1 if j=1) 

    mod<-get(name)                   #accessing GAM model ojbect "j"

    results_AIC[i,1]<- dates[i]  #storing the interpolation dates in the first column

    results_AIC[i,2]<- ns        #number of stations used in the training stage

    results_AIC[i,j+2]<- AIC (mod)

    results_GCV[i,1]<- dates[i]  #storing the interpolation dates in the first column

    results_GCV[i,2]<- ns        #number of stations used in the training stage

    results_GCV[i,j+2]<- mod$gcv.ubre

    results_DEV[i,1]<- dates[i]  #storing the interpolation dates in the first column

    results_DEV[i,2]<- ns        #number of stations used in the training stage

    results_DEV[i,j+2]<- mod$deviance

    #####VALIDATION: Prediction checking the results using the testing data########

    #Automate this using a data frame of size??

    y_mod<- predict(mod, newdata=data_v, se.fit = TRUE) #Using the coeff to predict new values.

    res_mod<- data_v$tmax - y_mod$fit #Residuals for GMA model that resembles the ANUSPLIN interpolation

    RMSE_mod <- sqrt(sum(res_mod^2)/nv) #RMSE FOR REGRESSION STEP 1: GAM     

    results_RMSE[i,1]<- dates[i]  #storing the interpolation dates in the first column

    results_RMSE[i,2]<- ns        #number of stations used in the training stage

    results_RMSE[i,j+2]<- RMSE_mod

    #Saving residuals and prediction in the dataframes: tmax predicted from GAM

    pred<-paste("pred_mod",j,sep="")

    data_v[[pred]]<-as.numeric(y_mod$fit)

    data_s[[pred]]<-as.numeric(mod$fit) #Storing model fit values (predicted on training sample)

    name2<-paste("res_mod",j,sep="")

    data_v[[name2]]<-as.numeric(res_mod)

    data_s[[name2]]<-as.numeric(mod$residuals)

    #end of loop calculating RMSE

    #NEED TO ADD BIAS AND MAE

    }

  #KRIGING ON GAM RESIDUALS: REGRESSION STEP2

    v<-variogram(residuals~1, data_s)   

    plot(v)                               # This plot may be saved at a later stage...

    dev.off()

    #NEED TO ADD IT BACK TO THE PREDICTION FROM GAM

    gam_kr<-paste("pred_gam_kr",j,sep="")

    pred_gam<-paste("pred_mod",j,sep="")

    data_s[[gam_kr]]<-data_s[[pred_gam]]+ data_s[[name2]]

    data_v[[gam_kr]]<-data_v[[pred_gam]]+ data_v[[name2]]

    res_mod_kr_s<- data_s$tmax - data_s[[gam_kr]]           #Residuals from kriging.

    res_mod_kr_v<- data_v$tmax - data_v[[gam_kr]]           #Residuals from cokriging.

    #as.numeric(res_mod)

    #data_s[[name3]]<-res_mod_kr_s

    data_s[[name3]]<-as.numeric(res_mod_kr_s)

    #data_v[[name3]]<-res_mod_kr_v 

    data_v[[name3]]<-as.numeric(res_mod_kr_v)

    #Writing residuals from kriging

  ###SAVING THE DATA FRAME IN SHAPEFILES AND TEXTFILES

  #outfile<-sub(".shp","",data_name)   #Removing extension if it is present

  #writeOGR(data_v,".", outfile, driver ="ESRI Shapefile")

  #outfile<-sub(".shp","",data_name)   #Removing extension if it is present

  #writeOGR(data_s,".", outfile, driver ="ESRI Shapefile")

  # end of the for loop1

  barplot(height,ylim=c(14,36),ylab="RMSE in tenth deg C",beside=TRUE,

  dev.off()

r1<-(results_table_RMSE[,3:10]) #selecting only the columns related to models and method 1

r2<-(results_table_RMSE[,3:10]) #selecting only the columns related to models and method 1

mean_r1<-mean(r1)

mean_r2<-mean(r2)

median_r1<-sapply(r1, median)   #Calulcating the mean for every model (median of columns)

median_r2<-sapply(r2, median)

sd_r1<-sapply(r1, sd)

sd_r2<-sapply(r2, sd)

barplot(mean_r,ylim=c(23,26),ylab="RMSE in tenth deg C")

barplot(median_r,ylim=c(23,26),ylab="RMSE in tenth deg C",add=TRUE,inside=FALSE,beside=TRUE) # put both on the same plot

barplot(sd_r,ylim=c(6,8),ylab="RMSE in tenth deg C") # put both on the same plot

height<-rbind(mean_r,median_r)

barplot(height,ylim=c(23,26),ylab="RMSE in tenth deg C",beside=TRUE,legend=rownames(height))

barplot(height,ylim=c(23,26),ylab="RMSE in tenth deg C",beside=TRUE, col=c("darkblue","red"),legend=rownames(height)) # put both on the same plot

barplot2(mean_r,median_r,ylim=c(23,26),ylab="RMSE in tenth deg C") # put both on the same plot

#Collect var explained and p values for each var...

### End of script  ##########