##################    MULTI SAMPLING GAM FUSION METHOD ASSESSMENT ####################################

############################ Merging LST and station data ##########################################

#interpolation area. It requires the text file of stations and a shape file of the study area.           

#Note that the projection for both GHCND and study area is lonlat WGS84.       

#Options to run this program are:

#1) Multisampling: vary the porportions of hold out and use random samples for each run

#2)Constant sampling: use the same sample over the runs

#3)over dates: run over for example 365 dates without mulitsampling

#4)use seed number: use seed if random samples must be repeatable

#5)GAM fusion: possibilty of running GAM+FUSION or GAM separately 

#AUTHOR: Benoit Parmentier                                                                        

###Loading R library and packages                                                      

library(gtools)                                         # loading some useful tools 

library(mgcv)                                           # GAM package by Simon Wood

library(sp)                                             # Spatial pacakge with class definition by Bivand et al.

library(spdep)                               # Spatial pacakge with methods and spatial stat. by Bivand et al.

library(rgdal)                               # GDAL wrapper for R, spatial utilities

library(gstat)                               # Kriging and co-kriging by Pebesma et al.

library(rasterVis)

library(parallel)                            # Urbanek S. and Ripley B., package for multi cores & parralel processing

library(plotrix)

infile_daily<-"daily_covariates_ghcn_data_TMAXy2010_2010_VE_02082013.shp"

infile_locs<-"stations_venezuela_region_y2010_2010_VE_02082013.shp"

out_path<-"/home/parmentier/Data/IPLANT_project/Venezuela_interpolation/Venezuela_01142013/output_data"

predval<-1

bias_val<-0            #if value 1 then training data is used in the bias surface rather than the all monthly stations

nb_sample<-1           #number of time random sampling must be repeated for every hold out proportion

prop_min<-0.3          #if prop_min=prop_max and step=0 then predicitons are done for the number of dates...

prop_max<-0.3

step<-0         

###################### START OF THE SCRIPT ########################

ghcn<-readOGR(dsn=in_path,layer=sub(".shp","",infile_daily))

CRS_interp<-proj4string(ghcn)                       #Storing projection information (ellipsoid, datum,etc.)

ghcn$LC3[is.na(ghcn$LC3)]<-0

ghcn$CANHEIGHT[is.na(ghcn$CANHEIGHT)]<-0

ghcn$LC6[is.na(ghcn$LC6)]<-0

##Extracting the variables values from the raster files                                             

lst_names<-c("mm_01","mm_02","mm_03","mm_04","mm_05","mm_06","mm_07","mm_08","mm_09","mm_10","mm_11","mm_12",

                    "nobs_01","nobs_02","nobs_03","nobs_04","nobs_05","nobs_06","nobs_07","nobs_08",

                    "nobs_09","nobs_10","nobs_11","nobs_12")

covar_names<-c(rnames,lc_names,lst_names)

s_raster<-stack(infile3)                   #read in the data stack

names(s_raster)<-covar_names               #Assigning names to the raster layers: making sure it is included in the extraction

#Deal with no data value and zero      

#LC1<-raster(s_raster,layer=pos)             #Select layer from stack

#s_raster<-dropLayer(s_raster,pos)

#LC1[is.na(LC1)]<-0

#pos<-match("LC3",layerNames(s_raster)) #Find column with name "value"

#LC3<-raster(s_raster,layer=pos)             #Select layer from stack

#s_raster<-dropLayer(s_raster,pos)

#LC3[is.na(LC3)]<-0

#pos<-match("CANHEIGHT",layerNames(s_raster)) #Find column with name "value"

#CANHEIGHT<-raster(s_raster,layer=pos)             #Select layer from stack

#s_raster<-dropLayer(s_raster,pos)

#CANHEIGHT[is.na(CANHEIGHT)]<-0

#pos<-match("ELEV_SRTM",layerNames(s_raster)) #Find column with name "ELEV_SRTM"

#ELEV_SRTM<-raster(s_raster,layer=pos)             #Select layer from stack on 10/30

#s_raster<-dropLayer(s_raster,pos)

#ELEV_SRTM[ELEV_SRTM <0]<-NA

#s_sgdf<-as(s_raster,"SpatialGridDataFrame") #Conversion to spatial grid data frame

######  Preparing tables for model assessment: specific diagnostic/metrics

#Model assessment: specific diagnostics/metrics

results_AIC<- matrix(1,1,nmodels+3)  

results_GCV<- matrix(1,1,nmodels+3)

results_DEV<- matrix(1,1,nmodels+3)

#results_RMSE_f<- matrix(1,length(models)+3)

#Model assessment: general diagnostic/metrics 

results_RMSE <- matrix(1,1,nmodels+4)

results_MAE <- matrix(1,1,nmodels+4)

results_ME <- matrix(1,1,nmodels+4)       #There are 8+1 models

results_R2 <- matrix(1,1,nmodels+4)       #Coef. of determination for the validation dataset

results_RMSE_f<- matrix(1,1,nmodels+4)    #RMSE fit, RMSE for the training dataset

results_MAE_f <- matrix(1,1,nmodels+4)

#ghcn_test<-subset(ghcn,ghcn$value>-150 & ghcn$value<400)

#ghcn_test<-ghcn

#ghcn_test2<-subset(ghcn_test,ghcn_test$elev_1>0)

#ghcn<-ghcn_test2

#dst_all<-dst

dst_all<-data3

dst<-data3

#dst<-subset(dst,dst$TMax>-15 & dst$TMax<45) #may choose different threshold??

#dst<-subset(dst,dst$ELEV_SRTM>0) #This will drop two stations...or 24 rows

  set.seed(seed_number)                        #Using a seed number allow results based on random number to be compared...

}

dates_list<-vector("list",nel) #list of one row data.frame

sn<-length(dates)*nb_sample*length(prop_range)               #Number of samples to run

for(i in 1:length(dates)){

  d_tmp<-rep(dates[i],nb_sample*length(prop_range)) #repeating same date

  s_nb<-rep(1:nb_sample,length(prop_range))         #number of random sample per proportion

  prop_tmp<-sort(rep(prop_range, nb_sample))

  tab_run_tmp<-cbind(d_tmp,s_nb,prop_tmp)

  dates_list[[i]]<-tab_run_tmp

}

sampling_dat<-as.data.frame(do.call(rbind,dates_list))

names(sampling_dat)<-c("date","run_samp","prop")

for(i in 2:3){            # start of the for loop #1

  sampling_dat[,i]<-as.numeric(as.character(sampling_dat[,i]))  

}

sampling_dat$date<- as.character(sampling_dat[,1])

#ghcn.subsets <-lapply(dates, function(d) subset(ghcn, date==d)) #this creates a list of 10 or 365 subsets dataset based on dates

ghcn.subsets <-lapply(as.character(sampling_dat$date), function(d) subset(ghcn, date==d)) #this creates a list of 10 or 365 subsets dataset based on dates

if (seed_number>0) {

  set.seed(seed_number)                        #Using a seed number allow results based on random number to be compared...

}

sampling<-vector("list",length(ghcn.subsets))

sampling_station_id<-vector("list",length(ghcn.subsets))

  n<-nrow(ghcn.subsets[[i]])

  prop<-(sampling_dat$prop[i])/100

  ns<-n-round(n*prop)   #Create a sample from the data frame with 70% of the rows

  nv<-n-ns              #create a sample for validation with prop of the rows

  ind.training <- sample(nrow(ghcn.subsets[[i]]), size=ns, replace=FALSE) #This selects the index position for 70% of the rows taken randomly

  ind.testing <- setdiff(1:nrow(ghcn.subsets[[i]]), ind.training)

  data_sampled<-ghcn.subsets[[i]][ind.training,] #selected the randomly sampled stations

  station_id.training<-data_sampled$station     #selected id for the randomly sampled stations (115)

  #Save the information

}

## Use same samples across the year...

if (constant==1){

  sampled<-sampling[[1]]

  data_sampled<-ghcn.subsets[[1]][sampled,] #selected the randomly sampled stations

  station_sampled<-data_sampled$station     #selected id for the randomly sampled stations (115)

  list_const_sampling<-vector("list",sn)

  list_const_sampling_station_id<-vector("list",sn)

  for(i in 1:sn){

    station_id.training<-intersect(station_sampled,ghcn.subsets[[i]]$station)

    ind.training<-match(station_id.training,ghcn.subsets[[i]]$station)

    list_const_sampling[[i]]<-ind.training

    list_const_sampling_station_id[[i]]<-station_id.training

  }

  sampling<-list_const_sampling 

  sampling_station_id<-list_const_sampling_station_id

######## Prediction for the range of dates and sampling data

#gam_fus_mod<-mclapply(1:length(dates), runGAMFusion,mc.preschedule=FALSE,mc.cores = 8) #This is the end bracket from mclapply(...) statement

## Plotting and saving diagnostic measures

tb<-gam_fus_mod_s[[1]][[3]][0,]  #empty data frame with metric table structure that can be used in rbinding...

tb_tmp<-gam_fus_mod_s #copy

for (i in 1:length(tb_tmp)){

  tmp<-tb_tmp[[i]][[3]]

  tb<-rbind(tb,tmp)

}

rm(tb_tmp)

for(i in 4:ncol(tb)){            # start of the for loop #1

  tb[,i]<-as.numeric(as.character(tb[,i]))  

}

  tb_metric<-subset(tb, metric==metrics[i])

  tb_metric<-cbind(tb_metric,sampling_dat[,2:3])

  assign(metric_name,tb_metric)

  tb_metric_list[[i]]<-tb_metric

}

tb_diagnostic<-do.call(rbind,tb_metric_list)

mod_pat<-glob2rx("mod*")   

mod_var<-grep(mod_pat,names(tb_diagnostic),value=TRUE) # using grep with "value" extracts the matching names         

t<-melt(tb_diagnostic,

        measure=mod_var, 

        id=c("dates","metric","prop"),

        na.rm=F)

avg_tb<-cast(t,metric+prop~variable,mean)

median_tb<-cast(t,metric+prop~variable,median)

avg_tb[["prop"]]<-as.numeric(as.character(avg_tb[["prop"]]))

avg_RMSE<-subset(avg_tb,metric=="RMSE")

write.table(median_tb, file= paste(path,"/","results2_fusion_Assessment_measure_median_",out_prefix,".txt",sep=""), sep=",")

write.table(tb, file= paste(path,"/","results2_fusion_Assessment_measure_all",out_prefix,".txt",sep=""), sep=",")

#save(gam_fus_mod_s,file= paste(path,"/","results2_fusion_Assessment_measure_all",out_prefix,".RData",sep=""))

gam_fus_mod_obj<-list(gam_fus_mod=gam_fus_mod_s,sampling_obj=sampling_obj)

save(gam_fus_mod_obj,file= paste(path,"/","results_mod_obj_",out_prefix,".RData",sep=""))

#### END OF SCRIPT