1
|
runGAMFusion <- function(i) { # loop over dates
|
2
|
|
3
|
#date<-strptime(dates[i], "%Y%m%d") # interpolation date being processed
|
4
|
date<-strptime(sampling_dat$date[i], "%Y%m%d") # interpolation date being processed
|
5
|
|
6
|
month<-strftime(date, "%m") # current month of the date being processed
|
7
|
LST_month<-paste("mm_",month,sep="") # name of LST month to be matched
|
8
|
|
9
|
#Adding layer LST to the raster stack
|
10
|
|
11
|
pos<-match("LST",layerNames(s_raster)) #Find column with name "LST"
|
12
|
s_raster<-dropLayer(s_raster,pos) # If it exists drop layer
|
13
|
pos<-match(LST_month,layerNames(s_raster)) #Find column with the current month for instance mm12
|
14
|
r1<-raster(s_raster,layer=pos) #Select layer from stack
|
15
|
layerNames(r1)<-"LST"
|
16
|
s_raster<-addLayer(s_raster,r1) #Adding current month as "LST"
|
17
|
|
18
|
###Regression part 1: Creating a validation dataset by creating training and testing datasets
|
19
|
|
20
|
mod_LST <-ghcn.subsets[[i]][,match(LST_month, names(ghcn.subsets[[i]]))] #Match interpolation date and monthly LST average
|
21
|
ghcn.subsets[[i]] = transform(ghcn.subsets[[i]],LST = mod_LST) #Add the variable LST to the subset dataset
|
22
|
dst$LST<-dst[[LST_month]]
|
23
|
#n<-nrow(ghcn.subsets[[i]])
|
24
|
#ns<-n-round(n*prop) #Create a sample from the data frame with 70% of the rows
|
25
|
#nv<-n-ns #create a sample for validation with prop of the rows
|
26
|
#ind.training <- sample(nrow(ghcn.subsets[[i]]), size=ns, replace=FALSE) #This selects the index position for 70% of the rows taken randomly
|
27
|
ind.training<-sampling[[i]]
|
28
|
ind.testing <- setdiff(1:nrow(ghcn.subsets[[i]]), ind.training)
|
29
|
data_s <- ghcn.subsets[[i]][ind.training, ] #Training dataset currently used in the modeling
|
30
|
data_v <- ghcn.subsets[[i]][ind.testing, ] #Testing/validation dataset using input sampling
|
31
|
|
32
|
ns<-nrow(data_s)
|
33
|
nv<-nrow(data_v)
|
34
|
#i=1
|
35
|
date_proc<-sampling_dat$date[i]
|
36
|
date_proc<-strptime(sampling_dat$date[i], "%Y%m%d") # interpolation date being processed
|
37
|
mo<-as.integer(strftime(date_proc, "%m")) # current month of the date being processed
|
38
|
day<-as.integer(strftime(date_proc, "%d"))
|
39
|
year<-as.integer(strftime(date_proc, "%Y"))
|
40
|
|
41
|
datelabel=format(ISOdate(year,mo,day),"%b %d, %Y")
|
42
|
|
43
|
###########
|
44
|
# STEP 1 - LST 10 year monthly averages
|
45
|
###########
|
46
|
|
47
|
themolst<-raster(molst,mo) #current month being processed saved in a raster image
|
48
|
plot(themolst)
|
49
|
|
50
|
###########
|
51
|
# STEP 2 - Weather station means across same days: Monthly mean calculation
|
52
|
###########
|
53
|
|
54
|
modst<-dst[dst$month==mo,] #Subsetting dataset for the relevant month of the date being processed
|
55
|
|
56
|
##########
|
57
|
# STEP 3 - get LST at stations
|
58
|
##########
|
59
|
|
60
|
sta_lola=modst[,c("lon","lat")] #Extracting locations of stations for the current month..
|
61
|
|
62
|
proj_str="+proj=lcc +lat_1=43 +lat_2=45.5 +lat_0=41.75 +lon_0=-120.5 +x_0=400000 +y_0=0 +ellps=GRS80 +units=m +no_defs";
|
63
|
lookup<-function(r,lat,lon) {
|
64
|
xy<-project(cbind(lon,lat),proj_str);
|
65
|
cidx<-cellFromXY(r,xy);
|
66
|
return(r[cidx])
|
67
|
}
|
68
|
sta_tmax_from_lst=lookup(themolst,sta_lola$lat,sta_lola$lon) #Extracted values of LST for the stations
|
69
|
|
70
|
#########
|
71
|
# STEP 4 - bias at stations
|
72
|
#########
|
73
|
|
74
|
sta_bias=sta_tmax_from_lst-modst$TMax; #That is the difference between the monthly LST mean and monthly station mean
|
75
|
#Added by Benoit
|
76
|
modst$LSTD_bias<-sta_bias #Adding bias to data frame modst containning the monthly average for 10 years
|
77
|
|
78
|
bias_xy=project(as.matrix(sta_lola),proj_str)
|
79
|
png(paste("LST_TMax_scatterplot_",sampling_dat$date[i],"_",sampling_dat$prop[i],"_",sampling_dat$run_samp[i], out_prefix,".png", sep=""))
|
80
|
plot(modst$TMax,sta_tmax_from_lst,xlab="Station mo Tmax",ylab="LST mo Tmax",main=paste("LST vs TMax for",datelabel,sep=" "))
|
81
|
abline(0,1)
|
82
|
dev.off()
|
83
|
|
84
|
#added by Benoit
|
85
|
#x<-ghcn.subsets[[i]] #Holds both training and testing for instance 161 rows for Jan 1
|
86
|
x<-data_v
|
87
|
d<-data_s
|
88
|
|
89
|
pos<-match("value",names(d)) #Find column with name "value"
|
90
|
#names(d)[pos]<-c("dailyTmax")
|
91
|
names(d)[pos]<-y_var_name
|
92
|
names(x)[pos]<-y_var_name
|
93
|
#names(x)[pos]<-c("dailyTmax")
|
94
|
d$dailyTmax=(as.numeric(d$dailyTmax))/10 #stored as 1/10 degree C to allow integer storage
|
95
|
x$dailyTmax=(as.numeric(x$dailyTmax))/10 #stored as 1/10 degree C to allow integer storage
|
96
|
pos<-match("station",names(d)) #Find column with name "value"
|
97
|
names(d)[pos]<-c("id")
|
98
|
names(x)[pos]<-c("id")
|
99
|
names(modst)[1]<-c("id") #modst contains the average tmax per month for every stations...
|
100
|
|
101
|
dmoday=merge(modst,d,by="id",suffixes=c("",".y2")) #LOOSING DATA HERE!!! from 113 t0 103
|
102
|
xmoday=merge(modst,x,by="id",suffixes=c("",".y2")) #LOOSING DATA HERE!!! from 48 t0 43
|
103
|
mod_pat<-glob2rx("*.y2")
|
104
|
var_pat<-grep(mod_pat,names(dmoday),value=FALSE) # using grep with "value" extracts the matching names
|
105
|
dmoday<-dmoday[,-var_pat]
|
106
|
mod_pat<-glob2rx("*.y2")
|
107
|
var_pat<-grep(mod_pat,names(xmoday),value=FALSE) # using grep with "value" extracts the matching names
|
108
|
xmoday<-xmoday[,-var_pat] #Removing duplicate columns
|
109
|
|
110
|
data_v<-xmoday
|
111
|
###
|
112
|
|
113
|
#dmoday contains the daily tmax values for training with TMax being the monthly station tmax mean
|
114
|
#xmoday contains the daily tmax values for validation with TMax being the monthly station tmax mean
|
115
|
|
116
|
png(paste("Daily_tmax_monthly_TMax_scatterplot_",sampling_dat$date[i],"_",sampling_dat$prop[i],
|
117
|
"_",sampling_dat$run_samp[i],out_prefix,".png", sep=""))
|
118
|
plot(dailyTmax~TMax,data=dmoday,xlab="Mo Tmax",ylab=paste("Daily for",datelabel),main="across stations in OR")
|
119
|
dev.off()
|
120
|
|
121
|
########
|
122
|
# STEP 5 - interpolate bias
|
123
|
########
|
124
|
|
125
|
# ?? include covariates like elev, distance to coast, cloud frequency, tree heig
|
126
|
#fitbias<-Tps(bias_xy,sta_bias) #use TPS or krige
|
127
|
|
128
|
#Adding options to use only training stations : 07/11/2012
|
129
|
bias_xy=project(as.matrix(sta_lola),proj_str)
|
130
|
#bias_xy2=project(as.matrix(c(dmoday$lon,dmoday$lat),proj_str)
|
131
|
if(bias_val==1){
|
132
|
sta_bias<-dmoday$LSTD_bias
|
133
|
bias_xy<-cbind(dmoday$x_OR83M,dmoday$y_OR83M)
|
134
|
}
|
135
|
|
136
|
fitbias<-Krig(bias_xy,sta_bias,theta=1e5) #use TPS or krige
|
137
|
#The output is a krig object using fields
|
138
|
mod9a<-fitbias
|
139
|
|
140
|
##########
|
141
|
# STEP 7 - interpolate delta across space
|
142
|
##########
|
143
|
|
144
|
daily_sta_lola=dmoday[,c("lon","lat")] #could be same as before but why assume merge does this - assume not
|
145
|
daily_sta_xy=project(as.matrix(daily_sta_lola),proj_str)
|
146
|
daily_delta=dmoday$dailyTmax-dmoday$TMax
|
147
|
|
148
|
#fitdelta<-Tps(daily_sta_xy,daily_delta) #use TPS or krige
|
149
|
fitdelta<-Krig(daily_sta_xy,daily_delta,theta=1e5) #use TPS or krige
|
150
|
#Kriging using fields package
|
151
|
mod9b<-fitdelta
|
152
|
|
153
|
png(paste("Delta_surface_LST_TMax_",sampling_dat$date[i],"_",sampling_dat$prop[i],
|
154
|
"_",sampling_dat$run_samp[i],out_prefix,".png", sep=""))
|
155
|
surface(fitdelta,col=rev(terrain.colors(100)),asp=1,main=paste("Interpolated delta for",datelabel,sep=" "))
|
156
|
dev.off()
|
157
|
#
|
158
|
|
159
|
#### Added by Benoit on 06/19
|
160
|
data_s<-dmoday #put the
|
161
|
data_s$daily_delta<-daily_delta
|
162
|
|
163
|
#data_s$y_var<-daily_delta #y_var is the variable currently being modeled, may be better with BIAS!!
|
164
|
#data_s$y_var<-data_s$LSTD_bias
|
165
|
#### Added by Benoit ends
|
166
|
|
167
|
#########
|
168
|
# STEP 8 - assemble final answer - T=LST+Bias(interpolated)+delta(interpolated)
|
169
|
#########
|
170
|
|
171
|
bias_rast=interpolate(themolst,fitbias) #interpolation using function from raster package
|
172
|
#themolst is raster layer, fitbias is "Krig" object from bias surface
|
173
|
#plot(bias_rast,main="Raster bias") #This not displaying...
|
174
|
|
175
|
#Saving kriged surface in raster images
|
176
|
data_name<-paste("bias_LST_",sampling_dat$date[i],"_",sampling_dat$prop[i],
|
177
|
"_",sampling_dat$run_samp[i],sep="")
|
178
|
raster_name<-paste("fusion_",data_name,out_prefix,".rst", sep="")
|
179
|
writeRaster(bias_rast, filename=raster_name,overwrite=TRUE) #Writing the data in a raster file format...(IDRISI)
|
180
|
|
181
|
daily_delta_rast=interpolate(themolst,fitdelta) #Interpolation of the bias surface...
|
182
|
|
183
|
#plot(daily_delta_rast,main="Raster Daily Delta")
|
184
|
|
185
|
#Saving kriged surface in raster images
|
186
|
data_name<-paste("daily_delta_",sampling_dat$date[i],"_",sampling_dat$prop[i],
|
187
|
"_",sampling_dat$run_samp[i],sep="")
|
188
|
raster_name<-paste("fusion_",data_name,out_prefix,".rst", sep="")
|
189
|
writeRaster(daily_delta_rast, filename=raster_name,overwrite=TRUE) #Writing the data in a raster file format...(IDRISI)
|
190
|
|
191
|
tmax_predicted=themolst+daily_delta_rast-bias_rast #Final surface as a raster layer...
|
192
|
#tmax_predicted=themolst+daily_delta_rast+bias_rast #Added by Benoit, why is it -bias_rast
|
193
|
#plot(tmax_predicted,main="Predicted daily")
|
194
|
|
195
|
#Saving kriged surface in raster images
|
196
|
data_name<-paste("tmax_predicted_",sampling_dat$date[i],"_",sampling_dat$prop[i],
|
197
|
"_",sampling_dat$run_samp[i],sep="")
|
198
|
raster_name<-paste("fusion_",data_name,out_prefix,".rst", sep="")
|
199
|
writeRaster(tmax_predicted, filename=raster_name,overwrite=TRUE) #Writing the data in a raster file format...(IDRISI)
|
200
|
|
201
|
########
|
202
|
# check: assessment of results: validation
|
203
|
########
|
204
|
RMSE<-function(x,y) {return(mean((x-y)^2)^0.5)}
|
205
|
MAE_fun<-function(x,y) {return(mean(abs(x-y)))}
|
206
|
#ME_fun<-function(x,y){return(mean(abs(y)))}
|
207
|
#FIT ASSESSMENT
|
208
|
sta_pred_data_s=lookup(tmax_predicted,data_s$lat,data_s$lon)
|
209
|
rmse_fit=RMSE(sta_pred_data_s,data_s$dailyTmax)
|
210
|
mae_fit=MAE_fun(sta_pred_data_s,data_s$dailyTmax)
|
211
|
|
212
|
sta_pred=lookup(tmax_predicted,data_v$lat,data_v$lon)
|
213
|
#sta_pred=lookup(tmax_predicted,daily_sta_lola$lat,daily_sta_lola$lon)
|
214
|
#rmse=RMSE(sta_pred,dmoday$dailyTmax)
|
215
|
#pos<-match("value",names(data_v)) #Find column with name "value"
|
216
|
#names(data_v)[pos]<-c("dailyTmax")
|
217
|
tmax<-data_v$dailyTmax
|
218
|
#data_v$dailyTmax<-tmax
|
219
|
rmse=RMSE(sta_pred,tmax)
|
220
|
mae<-MAE_fun(sta_pred,tmax)
|
221
|
r2<-cor(sta_pred,tmax)^2 #R2, coef. of var
|
222
|
me<-mean(sta_pred-tmax)
|
223
|
|
224
|
png(paste("Predicted_tmax_versus_observed_scatterplot_",sampling_dat$date[i],"_",sampling_dat$prop[i],
|
225
|
"_",sampling_dat$run_samp[i],out_prefix,".png", sep=""))
|
226
|
plot(sta_pred~tmax,xlab=paste("Actual daily for",datelabel),ylab="Pred daily",main=paste("RMSE=",rmse))
|
227
|
abline(0,1)
|
228
|
dev.off()
|
229
|
#resid=sta_pred-dmoday$dailyTmax
|
230
|
resid=sta_pred-tmax
|
231
|
|
232
|
###BEFORE GAM prediction the data object must be transformed to SDF
|
233
|
|
234
|
coords<- data_v[,c('x_OR83M','y_OR83M')]
|
235
|
coordinates(data_v)<-coords
|
236
|
proj4string(data_v)<-CRS #Need to assign coordinates...
|
237
|
coords<- data_s[,c('x_OR83M','y_OR83M')]
|
238
|
coordinates(data_s)<-coords
|
239
|
proj4string(data_s)<-CRS #Need to assign coordinates..
|
240
|
coords<- modst[,c('x_OR83M','y_OR83M')]
|
241
|
coordinates(modst)<-coords
|
242
|
proj4string(modst)<-CRS #Need to assign coordinates..
|
243
|
|
244
|
ns<-nrow(data_s) #This is added to because some loss of data might have happened because of the averaging...
|
245
|
nv<-nrow(data_v)
|
246
|
|
247
|
###GAM PREDICTION
|
248
|
|
249
|
if (bias_prediction==1){
|
250
|
data_s$y_var<-data_s$LSTD_bias #This shoudl be changed for any variable!!!
|
251
|
data_v$y_var<-data_v$LSTD_bias
|
252
|
data_month<-modst
|
253
|
data_month$y_var<-modst$LSTD_bias
|
254
|
}
|
255
|
|
256
|
if (bias_prediction==0){
|
257
|
data_v$y_var<-data_v[[y_var_name]]
|
258
|
data_s$y_var<-data_s[[y_var_name]]
|
259
|
}
|
260
|
|
261
|
#Model and response variable can be changed without affecting the script
|
262
|
|
263
|
formula1 <- as.formula("y_var ~ s(lat) + s(lon) + s(ELEV_SRTM)", env=.GlobalEnv)
|
264
|
formula2 <- as.formula("y_var~ s(lat,lon)+ s(ELEV_SRTM)", env=.GlobalEnv)
|
265
|
formula3 <- as.formula("y_var~ s(lat) + s (lon) + s (ELEV_SRTM) + s (Northness)+ s (Eastness) + s(DISTOC)", env=.GlobalEnv)
|
266
|
formula4 <- as.formula("y_var~ s(lat) + s (lon) + s(ELEV_SRTM) + s(Northness) + s (Eastness) + s(DISTOC) + s(LST)", env=.GlobalEnv)
|
267
|
formula5 <- as.formula("y_var~ s(lat,lon) +s(ELEV_SRTM) + s(Northness,Eastness) + s(DISTOC) + s(LST)", env=.GlobalEnv)
|
268
|
formula6 <- as.formula("y_var~ s(lat,lon) +s(ELEV_SRTM) + s(Northness,Eastness) + s(DISTOC) + s(LST)+s(LC1)", env=.GlobalEnv)
|
269
|
formula7 <- as.formula("y_var~ s(lat,lon) +s(ELEV_SRTM) + s(Northness,Eastness) + s(DISTOC) + s(LST)+s(LC3)", env=.GlobalEnv)
|
270
|
formula8 <- as.formula("y_var~ s(lat,lon) +s(ELEV_SRTM) + s(Northness,Eastness) + s(DISTOC) + s(LST) + s(LC1,LC3)", env=.GlobalEnv)
|
271
|
|
272
|
if (bias_prediction==1){
|
273
|
mod1<- try(gam(formula1, data=data_month))
|
274
|
mod2<- try(gam(formula2, data=data_month)) #modified nesting....from 3 to 2
|
275
|
mod3<- try(gam(formula3, data=data_month))
|
276
|
mod4<- try(gam(formula4, data=data_month))
|
277
|
mod5<- try(gam(formula5, data=data_month))
|
278
|
mod6<- try(gam(formula6, data=data_month))
|
279
|
mod7<- try(gam(formula7, data=data_month))
|
280
|
mod8<- try(gam(formula8, data=data_month))
|
281
|
|
282
|
} else if (bias_prediction==0){
|
283
|
|
284
|
mod1<- try(gam(formula1, data=data_s))
|
285
|
mod2<- try(gam(formula2, data=data_s)) #modified nesting....from 3 to 2
|
286
|
mod3<- try(gam(formula3, data=data_s))
|
287
|
mod4<- try(gam(formula4, data=data_s))
|
288
|
mod5<- try(gam(formula5, data=data_s))
|
289
|
mod6<- try(gam(formula6, data=data_s))
|
290
|
mod7<- try(gam(formula7, data=data_s))
|
291
|
mod8<- try(gam(formula8, data=data_s))
|
292
|
}
|
293
|
|
294
|
#Added
|
295
|
#tmax_predicted=themolst+daily_delta_rast-bias_rast #Final surface?? but daily_rst
|
296
|
|
297
|
### Added by benoit
|
298
|
#Store results using TPS
|
299
|
j=nmodels+1
|
300
|
results_RMSE[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
301
|
results_RMSE[2]<- ns #number of stations used in the training stage
|
302
|
results_RMSE[3]<- "RMSE"
|
303
|
|
304
|
results_RMSE[j+3]<- rmse #Storing RMSE for the model j
|
305
|
|
306
|
results_RMSE_f[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
307
|
results_RMSE_f[2]<- ns #number of stations used in the training stage
|
308
|
results_RMSE_f[3]<- "RMSE_f"
|
309
|
results_RMSE_f[j+3]<- rmse_fit #Storing RMSE for the model j
|
310
|
|
311
|
results_MAE_f[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
312
|
results_MAE_f[2]<- ns #number of stations used in the training stage
|
313
|
results_MAE_f[3]<- "RMSE_f"
|
314
|
results_MAE_f[j+3]<- mae_fit #Storing RMSE for the model j
|
315
|
|
316
|
results_MAE[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
317
|
results_MAE[2]<- ns #number of stations used in the training stage
|
318
|
results_MAE[3]<- "MAE"
|
319
|
results_MAE[j+3]<- mae #Storing RMSE for the model j
|
320
|
|
321
|
results_ME[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
322
|
results_ME[2]<- ns #number of stations used in the training stage
|
323
|
results_ME[3]<- "ME"
|
324
|
results_ME[j+3]<- me #Storing RMSE for the model j
|
325
|
|
326
|
results_R2[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
327
|
results_R2[2]<- ns #number of stations used in the training stage
|
328
|
results_R2[3]<- "R2"
|
329
|
results_R2[j+3]<- r2 #Storing RMSE for the model j
|
330
|
|
331
|
#ns<-nrow(data_s) #This is added to because some loss of data might have happened because of the averaging...
|
332
|
#nv<-nrow(data_v)
|
333
|
|
334
|
pred_mod<-paste("pred_mod",j,sep="")
|
335
|
#Adding the results back into the original dataframes.
|
336
|
data_s[[pred_mod]]<-sta_pred_data_s
|
337
|
data_v[[pred_mod]]<-sta_pred
|
338
|
|
339
|
#Model assessment: RMSE and then krig the residuals....!
|
340
|
|
341
|
res_mod_s<- data_s$dailyTmax - data_s[[pred_mod]] #Residuals from kriging training
|
342
|
res_mod_v<- data_v$dailyTmax - data_v[[pred_mod]] #Residuals from kriging validation
|
343
|
|
344
|
name2<-paste("res_mod",j,sep="")
|
345
|
data_v[[name2]]<-as.numeric(res_mod_v)
|
346
|
data_s[[name2]]<-as.numeric(res_mod_s)
|
347
|
|
348
|
|
349
|
for (j in 1:nmodels){
|
350
|
|
351
|
##Model assessment: specific diagnostic/metrics for GAM
|
352
|
|
353
|
name<-paste("mod",j,sep="") #modj is the name of The "j" model (mod1 if j=1)
|
354
|
mod<-get(name) #accessing GAM model ojbect "j"
|
355
|
|
356
|
#If mod "j" is not a model object
|
357
|
if (inherits(mod,"try-error")) {
|
358
|
results_AIC[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
359
|
results_AIC[2]<- ns #number of stations used in the training stage
|
360
|
results_AIC[3]<- "AIC"
|
361
|
results_AIC[j+3]<- NA
|
362
|
|
363
|
results_GCV[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
364
|
results_GCV[2]<- ns #number of stations used in the training
|
365
|
results_GCV[3]<- "GCV"
|
366
|
results_GCV[j+3]<- NA
|
367
|
|
368
|
results_DEV[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
369
|
results_DEV[2]<- ns #number of stations used in the training stage
|
370
|
results_DEV[3]<- "DEV"
|
371
|
results_DEV[j+3]<- NA
|
372
|
|
373
|
results_RMSE_f[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
374
|
results_RMSE_f[2]<- ns #number of stations used in the training stage
|
375
|
results_RMSE_f[3]<- "RSME_f"
|
376
|
results_RMSE_f[j+3]<- NA
|
377
|
|
378
|
results_MAE_f[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
379
|
results_MAE_f[2]<- ns #number of stations used in the training stage
|
380
|
results_MAE_f[3]<- "MAE_f"
|
381
|
results_MAE_f[j+3]<-NA
|
382
|
|
383
|
results_RMSE[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
384
|
results_RMSE[2]<- ns #number of stations used in the training stage
|
385
|
results_RMSE[3]<- "RMSE"
|
386
|
results_RMSE[j+3]<- NA #Storing RMSE for the model j
|
387
|
results_MAE[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
388
|
results_MAE[2]<- ns #number of stations used in the training stage
|
389
|
results_MAE[3]<- "MAE"
|
390
|
results_MAE[j+3]<- NA #Storing MAE for the model j
|
391
|
results_ME[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
392
|
results_ME[2]<- ns #number of stations used in the training stage
|
393
|
results_ME[3]<- "ME"
|
394
|
results_ME[j+3]<- NA #Storing ME for the model j
|
395
|
results_R2[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
396
|
results_R2[2]<- ns #number of stations used in the training stage
|
397
|
results_R2[3]<- "R2"
|
398
|
|
399
|
|
400
|
results_R2[j+3]<- NA #Storing R2 for the model j
|
401
|
|
402
|
}
|
403
|
|
404
|
#If mod is a modelobject
|
405
|
|
406
|
#If mod "j" is not a model object
|
407
|
if (inherits(mod,"gam")) {
|
408
|
|
409
|
results_AIC[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
410
|
results_AIC[2]<- ns #number of stations used in the training stage
|
411
|
results_AIC[3]<- "AIC"
|
412
|
results_AIC[j+3]<- AIC (mod)
|
413
|
|
414
|
results_GCV[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
415
|
results_GCV[2]<- ns #number of stations used in the training
|
416
|
results_GCV[3]<- "GCV"
|
417
|
results_GCV[j+3]<- mod$gcv.ubre
|
418
|
|
419
|
results_DEV[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
420
|
results_DEV[2]<- ns #number of stations used in the training stage
|
421
|
results_DEV[3]<- "DEV"
|
422
|
results_DEV[j+3]<- mod$deviance
|
423
|
|
424
|
y_var_fit= mod$fit
|
425
|
|
426
|
results_RMSE_f[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
427
|
results_RMSE_f[2]<- ns #number of stations used in the training stage
|
428
|
results_RMSE_f[3]<- "RSME_f"
|
429
|
#results_RMSE_f[j+3]<- sqrt(sum((y_var_fit-data_s$y_var)^2)/ns)
|
430
|
results_RMSE_f[j+3]<-sqrt(mean(mod$residuals^2,na.rm=TRUE))
|
431
|
|
432
|
results_MAE_f[1]<- sampling_dat$date[i] #storing the interpolation sampling_dat$date in the first column
|
433
|
results_MAE_f[2]<- ns #number of stations used in the training stage
|
434
|
results_MAE_f[3]<- "MAE_f"
|
435
|
#results_MAE_f[j+3]<-sum(abs(y_var_fit-data_s$y_var))/ns
|
436
|
results_MAE_f[j+3]<-mean(abs(mod$residuals),na.rm=TRUE)
|
437
|
|
438
|
##Model assessment: general diagnostic/metrics
|
439
|
##validation: using the testing data
|
440
|
if (predval==1) {
|
441
|
|
442
|
##Model assessment: specific diagnostic/metrics for GAM
|
443
|
|
444
|
name<-paste("mod",j,sep="") #modj is the name of The "j" model (mod1 if j=1)
|
445
|
mod<-get(name) #accessing GAM model ojbect "j"
|
446
|
|
447
|
s_sgdf<-as(s_raster,"SpatialGridDataFrame") #Conversion to spatial grid data frame
|
448
|
|
449
|
rpred<- predict(mod, newdata=s_sgdf, se.fit = TRUE) #Using the coeff to predict new values.
|
450
|
y_pred<-rpred$fit #rpred is a list with fit being and array
|
451
|
raster_pred<-r1
|
452
|
layerNames(raster_pred)<-"y_pred"
|
453
|
values(raster_pred)<-as.numeric(y_pred)
|
454
|
|
455
|
if (bias_prediction==1){
|
456
|
data_name<-paste("predicted_mod",j,"_",sampling_dat$date[i],"_",sampling_dat$prop[i],
|
457
|
"_",sampling_dat$run_samp[i],sep="")
|
458
|
raster_name<-paste("GAM_bias_",data_name,out_prefix,".rst", sep="")
|
459
|
writeRaster(raster_pred, filename=raster_name,overwrite=TRUE) #Writing the data in a raster file format...(IDRISI)
|
460
|
bias_rast<-raster_pred
|
461
|
|
462
|
raster_pred=themolst+daily_delta_rast-bias_rast #Final surface as a raster layer...
|
463
|
layerNames(raster_pred)<-"y_pred"
|
464
|
#=themolst+daily_delta_rast-bias_rast #Final surface as a raster layer...
|
465
|
|
466
|
data_name<-paste("predicted_mod",j,"_",sampling_dat$date[i],"_",sampling_dat$prop[i],
|
467
|
"_",sampling_dat$run_samp[i],sep="")
|
468
|
raster_name<-paste("GAM_bias_tmax_",data_name,out_prefix,".rst", sep="")
|
469
|
writeRaster(raster_pred, filename=raster_name,overwrite=TRUE) #Writing the data in a raster file format...(IDRISI)
|
470
|
|
471
|
}
|
472
|
|
473
|
if (bias_prediction==0){
|
474
|
data_name<-paste("predicted_mod",j,"_",sampling_dat$date[i],"_",sampling_dat$prop[i],
|
475
|
"_",sampling_dat$run_samp[i],sep="")
|
476
|
raster_name<-paste("GAM_",data_name,out_prefix,".rst", sep="")
|
477
|
writeRaster(raster_pred, filename=raster_name,overwrite=TRUE) #Writing the data in a raster file format...(IDRISI)
|
478
|
#writeRaster(r2, filename=raster_name,overwrite=TRUE) #Writing the data in a raster file format...(IDRISI)
|
479
|
|
480
|
}
|
481
|
|
482
|
|
483
|
pred_sgdf<-as(raster_pred,"SpatialGridDataFrame") #Conversion to spatial grid data frame
|
484
|
#rpred_val_s <- overlay(raster_pred,data_s) #This overlays the kriged surface tmax and the location of weather stations
|
485
|
|
486
|
rpred_val_s <- overlay(pred_sgdf,data_s) #This overlays the interpolated surface tmax and the location of weather stations
|
487
|
rpred_val_v <- overlay(pred_sgdf,data_v) #This overlays the interpolated surface tmax and the location of weather stations
|
488
|
|
489
|
pred_mod<-paste("pred_mod",j,sep="")
|
490
|
#Adding the results back into the original dataframes.
|
491
|
data_s[[pred_mod]]<-rpred_val_s$y_pred
|
492
|
|
493
|
data_v[[pred_mod]]<-rpred_val_v$y_pred
|
494
|
|
495
|
#Model assessment: RMSE and then krig the residuals....!
|
496
|
|
497
|
res_mod_s<-data_s[[y_var_name]] - data_s[[pred_mod]] #residuals from modeling training
|
498
|
res_mod_v<-data_v[[y_var_name]] - data_v[[pred_mod]] #residuals from modeling validation
|
499
|
|
500
|
}
|
501
|
|
502
|
if (predval==0) {
|
503
|
|
504
|
y_mod<- predict(mod, newdata=data_v, se.fit = TRUE) #Using the coeff to predict new values.
|
505
|
|
506
|
pred_mod<-paste("pred_mod",j,sep="")
|
507
|
#Adding the results back into the original dataframes.
|
508
|
data_s[[pred_mod]]<-as.numeric(mod$fit)
|
509
|
data_v[[pred_mod]]<-as.numeric(y_mod$fit)
|
510
|
|
511
|
#Model assessment: RMSE and then krig the residuals....!
|
512
|
|
513
|
#res_mod_s<- data_s$y_var - data_s[[pred_mod]] #Residuals from modeling training
|
514
|
#res_mod_v<- data_v$y_var - data_v[[pred_mod]] #Residuals from modeling validation
|
515
|
res_mod_s<-data_s[[y_var_name]] - data_s[[pred_mod]]
|
516
|
res_mod_v<-data_v[[y_var_name]] - data_v[[pred_mod]]
|
517
|
|
518
|
}
|
519
|
|
520
|
####ADDED ON JULY 20th
|
521
|
res_mod<-res_mod_v
|
522
|
|
523
|
#RMSE_mod <- sqrt(sum(res_mod^2)/nv) #RMSE FOR REGRESSION STEP 1: GAM
|
524
|
RMSE_mod<- sqrt(mean(res_mod^2,na.rm=TRUE))
|
525
|
#MAE_mod<- sum(abs(res_mod),na.rm=TRUE)/(nv-sum(is.na(res_mod))) #MAE from kriged surface validation
|
526
|
MAE_mod<- mean(abs(res_mod), na.rm=TRUE)
|
527
|
#ME_mod<- sum(res_mod,na.rm=TRUE)/(nv-sum(is.na(res_mod))) #ME, Mean Error or bias FOR REGRESSION STEP 1: GAM
|
528
|
ME_mod<- mean(res_mod,na.rm=TRUE) #ME, Mean Error or bias FOR REGRESSION STEP 1: GAM
|
529
|
#R2_mod<- cor(data_v$y_var,data_v[[pred_mod]])^2 #R2, coef. of var FOR REGRESSION STEP 1: GAM
|
530
|
R2_mod<- cor(data_v$y_var,data_v[[pred_mod]], use="complete")^2
|
531
|
results_RMSE[1]<- sampling_dat$date[i] #storing the interpolation sampling_dat$date in the first column
|
532
|
results_RMSE[2]<- ns #number of stations used in the training stage
|
533
|
results_RMSE[3]<- "RMSE"
|
534
|
results_RMSE[j+3]<- RMSE_mod #Storing RMSE for the model j
|
535
|
results_MAE[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
536
|
results_MAE[2]<- ns #number of stations used in the training stage
|
537
|
results_MAE[3]<- "MAE"
|
538
|
results_MAE[j+3]<- MAE_mod #Storing MAE for the model j
|
539
|
results_ME[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
540
|
results_ME[2]<- ns #number of stations used in the training stage
|
541
|
results_ME[3]<- "ME"
|
542
|
results_ME[j+3]<- ME_mod #Storing ME for the model j
|
543
|
results_R2[1]<- sampling_dat$date[i] #storing the interpolation dates in the first column
|
544
|
results_R2[2]<- ns #number of stations used in the training stage
|
545
|
results_R2[3]<- "R2"
|
546
|
results_R2[j+3]<- R2_mod #Storing R2 for the model j
|
547
|
|
548
|
#Saving residuals and prediction in the dataframes: tmax predicted from GAM
|
549
|
|
550
|
name2<-paste("res_mod",j,sep="")
|
551
|
data_v[[name2]]<-as.numeric(res_mod_v)
|
552
|
data_s[[name2]]<-as.numeric(res_mod_s)
|
553
|
#end of loop calculating RMSE
|
554
|
}
|
555
|
}
|
556
|
|
557
|
#if (i==length(dates)){
|
558
|
|
559
|
#Specific diagnostic measures related to the testing datasets
|
560
|
|
561
|
results_table_RMSE<-as.data.frame(results_RMSE)
|
562
|
results_table_MAE<-as.data.frame(results_MAE)
|
563
|
results_table_ME<-as.data.frame(results_ME)
|
564
|
results_table_R2<-as.data.frame(results_R2)
|
565
|
results_table_RMSE_f<-as.data.frame(results_RMSE_f)
|
566
|
results_table_MAE_f<-as.data.frame(results_MAE_f)
|
567
|
|
568
|
results_table_AIC<-as.data.frame(results_AIC)
|
569
|
results_table_GCV<-as.data.frame(results_GCV)
|
570
|
results_table_DEV<-as.data.frame(results_DEV)
|
571
|
|
572
|
tb_metrics1<-rbind(results_table_RMSE,results_table_MAE, results_table_ME, results_table_R2,results_table_RMSE_f,results_table_MAE_f) #
|
573
|
tb_metrics2<-rbind(results_table_AIC,results_table_GCV, results_table_DEV)
|
574
|
cname<-c("dates","ns","metric","mod1", "mod2","mod3", "mod4", "mod5", "mod6", "mod7","mod8","mod9")
|
575
|
colnames(tb_metrics1)<-cname
|
576
|
cname<-c("dates","ns","metric","mod1", "mod2","mod3", "mod4", "mod5", "mod6", "mod7","mod8")
|
577
|
colnames(tb_metrics2)<-cname
|
578
|
#colnames(results_table_RMSE)<-cname
|
579
|
#colnames(results_table_RMSE_f)<-cname
|
580
|
#tb_diagnostic1<-results_table_RMSE #measures of validation
|
581
|
#tb_diagnostic2<-results_table_RMSE_f #measures of fit
|
582
|
|
583
|
#write.table(tb_diagnostic1, file= paste(path,"/","results_fusion_Assessment_measure1",out_prefix,".txt",sep=""), sep=",")
|
584
|
|
585
|
#}
|
586
|
print(paste(sampling_dat$date[i],"processed"))
|
587
|
# end of the for loop1
|
588
|
mod_obj<-list(mod1,mod2,mod3,mod4,mod5,mod6,mod7,mod8,mod9a,mod9b)
|
589
|
names(mod_obj)<-c("mod1","mod2","mod3","mod4","mod5","mod6","mod7","mod8","mod9a","mod9b")
|
590
|
#results_list<-list(data_s,data_v,tb_metrics1,tb_metrics2)
|
591
|
results_list<-list(data_s,data_v,tb_metrics1,tb_metrics2,mod_obj,sampling_dat[i,],data_month)
|
592
|
names(results_list)<-c("data_s","data_v","tb_metrics1","tb_metrics2","mod_obj","sampling_dat","data_month")
|
593
|
save(results_list,file= paste(path,"/","results_list_metrics_objects_",sampling_dat$date[i],"_",sampling_dat$prop[i],
|
594
|
"_",sampling_dat$run_samp[i],out_prefix,".RData",sep=""))
|
595
|
return(results_list)
|
596
|
#return(tb_diagnostic1)
|
597
|
}
|