-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path3_Maps.R
328 lines (276 loc) · 14 KB
/
3_Maps.R
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
################################################################################
## R-Script: 3_Maps.R ##
## author: Javier Lopatin ##
## mail: javierlopatin@gmail.com ##
## ##
## Manuscript: Using aboveground vegetation attributes as proxies for mapping ##
## peatland belowground carbon stocks ##
## ##
## description: Create the prediction maps ##
## ##
################################################################################
library(randomForest)
library(caret)
library(doParallel)
### =============== ######################## =============== ###
### Spatial predicitons of aboveground vegetation attributes ###
### =============== ######################## =============== ###
## leave-one-out cross validation
train_control <- trainControl(method="LOOCV")
saveRasters = "D:/Peatland1/rasters.RData"
#load(saveRasters)
# =======================================================
# Prediction with remote sensing data and Random Forests
# =======================================================
# Predict floristic composition (MNMDS first axis)
m1data <- data.frame(x = data$NNMDS.sp1, H=data$Altura_vegetacion_cm,
hyperData[,2:ncol(hyperData)])
m1 <- train(x ~., data=m1data, trControl=train_control, tuneLength = 10, method="rf")
pred_m1 <- predicted(m1)
plot(m1data$x, pred_m1)
# Load other rasters predictors
hyper <- stack("D:/out_P1/hyper_P1_2m.tif")
names(hyper) <- paste0( rep("B", 41), seq(1,41,1) )
hyper[hyper==0]<- NA
DCM <- stack("D:/out_P1/treesvis/ndsm/DCM_2m.tif")
names(DCM) <- "H"
DCM2 <- resample(DCM2, hypr, resample='bilinear')
r <- stack(DCM2, hyper)
r[r$H > 2] <- NA; r[r$H < 0] <- NA # height mask
plot(r[[2]])
# create NDVI mask
NDVI <- ( hyper[[30]] - hyper[[20]] ) / ( hyper[[30]] + hyper[[20]] )
#NDVI <- mask(NDVI, DCM2) # exclude trees
NDVI[NDVI<0.3] <- NA
plot(NDVI)
r <- mask(r, NDVI); r <- mask(r, lim)
plot(r[[1]])
# Predict floristic composition at the species level
ordi_sp = predict(m1, r, type="response")
ordi_sp <- mask(ordi_sp, r[[1]])
plot(ordi_sp, zlim=c(-2,1))
# At the PFT level (not included in the final verison of the manuscript)
ordi_pft = predict(m2, r, type="response")
ordi_pft <- mask(ordi_pft, r[[1]])
plot(ordi_pft, zlim=c(-2,1))
save.image("peatland.RData")
# =======================================================
# Prediction with remote sensing data and PLS-PM
# =======================================================
## prepare PLS-PM rasters
### create validation data for PLS-PM
#FC
Spec_FC <- nnorm( (RS1$MNF1*PLS_FC$outer_model$weight[1])+
(RS1$MNF2*PLS_FC$outer_model$weight[2])+
(RS1$MNF3*PLS_FC$outer_model$weight[3]) )
Height_FC <- nnorm( (RS1$Elev_P25*PLS_FC$outer_model$weight[4])+
(RS1$Elev_P50*PLS_FC$outer_model$weight[5])+
(RS1$Elev_P75*PLS_FC$outer_model$weight[6])+
(RS1$Elev_P90*PLS_FC$outer_model$weight[7]) )
Structure_FC <- nnorm( (RS1$Elev_stddev*PLS_FC$outer_model$weight[8])+
(RS1$Elev_L1*PLS_FC$outer_model$weight[9])+
(RS1$Elev_SQRT_mean_SQ*PLS_FC$outer_model$weight[10])+
(RS1$Elev_CURT_mean_CUBE*PLS_FC$outer_model$weight[11]) )
#BM
Spec_BM <- nnorm( (RS1$MNF1*PLS_BM$outer_model$weight[1])+
(RS1$MNF2*PLS_BM$outer_model$weight[2])+
(RS1$MNF3*PLS_BM$outer_model$weight[3]) )
Height_BM <- nnorm( (RS1$Elev_P25*PLS_BM$outer_model$weight[4])+
(RS1$Elev_P50*PLS_BM$outer_model$weight[5])+
(RS1$Elev_P75*PLS_BM$outer_model$weight[6])+
(RS1$Elev_P90*PLS_BM$outer_model$weight[7]) )
Structure_BM <- nnorm( (RS1$Elev_skewness*PLS_BM$outer_model$weight[8])+
(RS1$Elev_L3*PLS_BM$outer_model$weight[9])+
((RS1$Canopy_relief_ratio*-1)*PLS_BM$outer_model$weight[10]) )
#Rich
Spec_Rich <- nnorm(((RS1$MNF1*-1)*PLS_Rich$outer_model$weight[1])+
(RS1$MNF3*PLS_Rich$outer_model$weight[2]) )
Height_Rich <- nnorm( (RS1$Elev_P25*PLS_Rich$outer_model$weight[3])+
(RS1$Elev_P50*PLS_Rich$outer_model$weight[4])+
(RS1$Elev_P75*PLS_Rich$outer_model$weight[5])+
(RS1$Elev_P90*PLS_Rich$outer_model$weight[6]) )
Structure_Rich<-nnorm( (RS1$Elev_stddev*PLS_Rich$outer_model$weight[7])+
(RS1$Elev_MAD_mode*PLS_Rich$outer_model$weight[8])+
(RS1$Elev_L2*PLS_Rich$outer_model$weight[9])+
(RS1$Elev_CURT_mean_CUBE*PLS_Rich$outer_model$weight[11]) )
#Depth
Spec_Depth <- nnorm( (RS1$MNF1*PLS_depth$outer_model$weight[1])+
(RS1$MNF2*PLS_depth$outer_model$weight[2])+
(RS1$MNF3*PLS_depth$outer_model$weight[3]) )
Height_Depth <- nnorm( (RS1$Elev_P25*PLS_depth$outer_model$weight[4])+
(RS1$Elev_P50*PLS_depth$outer_model$weight[5])+
(RS1$Elev_P75*PLS_depth$outer_model$weight[6])+
(RS1$Elev_P90*PLS_depth$outer_model$weight[7]) )
Structure_Depth<-nnorm((RS1$Elev_stddev*PLS_depth$outer_model$weight[8])+
(RS1$Elev_L1*PLS_depth$outer_model$weight[9])+
(RS1$Elev_SQRT_mean_SQ*PLS_depth$outer_model$weight[10])+
(RS1$Elev_CURT_mean_CUBE*PLS_depth$outer_model$weight[11]))
#C
Spec_C <- nnorm( (RS1$MNF1*PLS_C$outer_model$weight[1])+
(RS1$MNF2*PLS_C$outer_model$weight[2])+
(RS1$MNF3*PLS_C$outer_model$weight[3]) )
Height_C <- nnorm( (RS1$Elev_P25*PLS_C$outer_model$weight[4])+
(RS1$Elev_P50*PLS_C$outer_model$weight[5])+
(RS1$Elev_P75*PLS_C$outer_model$weight[6])+
(RS1$Elev_P90*PLS_C$outer_model$weight[7]) )
Structure_C <- nnorm( (RS1$Elev_stddev*PLS_C$outer_model$weight[8])+
(RS1$Elev_L1*PLS_C$outer_model$weight[9])+
(RS1$Elev_SQRT_mean_SQ*PLS_C$outer_model$weight[10])+
(RS1$Elev_CURT_mean_CUBE*PLS_C$outer_model$weight[11]) )
### =================================================== ###
### ============= Bootstrapping procedure ============= ###
### =================================================== ###
# lists to store maps
PLSPM_FC_map <- list()
PLSPM_BM_map <- list()
PLSPM_Rich_map <- list()
PLSPM_Depth_map <- list()
PLSPM_C_map <- list()
Cnorm <- function(x){
(x*sd(data$Carbono_Subterraneo_kg_m2)) + mean(data$Carbono_Subterraneo_kg_m2)
}
# ====================================================
# Bootstrapping using UAV-based predictors
# ====================================================
# set the bootstrap parameters
N = nrow(data) # N° of observations
B = 500 # N° of bootstrap iterations
RF_FC_map <- list()
RF_BM_map <- list()
RF_Rich_map <- list()
RF_Depth_map <- list()
RF_C_map <- list()
# initialize parallel processing
cl <- makeCluster(detectCores())
registerDoParallel(cl)
for(i in 1:500){
# create random numbers with replacement to select samples from each group
idx = sample(1:N, N, replace=TRUE)
# select subsets of the five groups based on the random numbers
train1 <- RS_test_data[idx,] # for PLS-PM
train2 <- RS_test_data2[idx,] # for the rest
### Run PLSPM
PLSrun_FC = plspm(train1, inner, outer_FC, modes, maxiter= 1000, boot.val = F,
br = 1000, scheme = "factor", scaled = T)
PLSrun_BM = plspm(train1, inner, outer_BM, modes, maxiter= 1000, boot.val = F,
br = 1000, scheme = "factor", scaled = T)
PLSrun_Rich = plspm(train1, inner, outer_Rich, modes, maxiter= 1000, boot.val = F,
br = 1000, scheme = "factor", scaled = T)
PLSrun_Depth = plspm(train1, inner, outer_depth, modes, maxiter= 1000, boot.val = F,
br = 1000, scheme = "factor", scaled = T)
PLSrun_C = plspm(train1, inner, outer_C, modes, maxiter= 1000, boot.val = F,
br = 1000, scheme = "factor", scaled = T)
# plspm LV scores
Scores_FC <- as.data.frame(PLSrun_FC$scores)
Scores_BM <- as.data.frame(PLSrun_BM$scores)
Scores_Rich <- as.data.frame(PLSrun_Rich$scores)
Scores_Depth <- as.data.frame(PLSrun_Depth$scores)
Scores_C <- as.data.frame(PLSrun_Depth$scores)
##################
#### train RF ####
##################
RFrun_FC <- randomForest(FC ~., data=train2[, c(2, 8:ncol(train2))],
mtry=fitRF_FC$bestTune$mtry, mtree=500, verbose=F)
RFrun_BM <- randomForest(BM ~., data=train2[, c(3, 8:ncol(train2))],
mtry=fitRF_BM$bestTune$mtry, mtree=500, verbose=F)
RFrun_Rich <- randomForest(Rich ~., data=train2[, c(4, 8:ncol(train2))],
mtry=fitRF_Rich$bestTune$mtry, mtree=500, verbose=F)
RFrun_Depth <- randomForest(Depth ~., data=train2[, c(5, 8:ncol(train2))],
mtry=fitRF_depth$bestTune$mtry, mtree=500, verbose=F)
RFrun_C <- randomForest(C ~., data=train2[, c(6, 8:ncol(train2))],
mtry=fitRF_C$bestTune$mtry, mtree=500, verbose=F)
### Prediction to validation data PLSPM
predPLSPM_FC <- PLSrun_FC$inner_model$X[1] + Spec_FC * PLSrun_FC$inner_model$X[2] +
Height_FC * PLSrun_FC$inner_model$X[3] + Structure_FC * PLSrun_FC$inner_model$X[4]
predPLSPM_BM <- PLSrun_BM$inner_model$X[1] + Spec_BM * PLSrun_BM$inner_model$X[2] +
Height_BM * PLSrun_BM$inner_model$X[3] + Structure_BM * PLSrun_BM$inner_model$X[4]
predPLSPM_Rich <- PLSrun_Rich$inner_model$X[1] + Spec_Rich * PLSrun_Rich$inner_model$X[2] +
Height_Rich * PLSrun_Rich$inner_model$X[3] + Structure_Rich * PLSrun_Rich$inner_model$X[4]
predPLSPM_Depth <- PLSrun_Depth$inner_model$X[1] + Spec_Depth * PLSrun_Depth$inner_model$X[2] +
Height_Depth * PLSrun_Depth$inner_model$X[3] + Structure_Depth * PLSrun_Depth$inner_model$X[4]
predPLSPM_C <- PLSrun_C$inner_model$X[1] + Spec_C * PLSrun_C$inner_model$X[2] + Height_C *
PLSrun_C$inner_model$X[3] + Structure_C * PLSrun_C$inner_model$X[4]
# RF
predRF_FC <- predict(RS2, RFrun_FC)
predRF_BM <- predict(RS2, RFrun_BM)
predRF_Rich <- predict(RS2, RFrun_Rich)
predRF_Depth <- predict(RS2, RFrun_Depth)
predRF_C <- predict(RS2, RFrun_C)
### store maps
PLSPM_FC_map[[i]] <- Cnorm(predPLSPM_FC)
PLSPM_BM_map[[i]] <- Cnorm(predPLSPM_BM)
PLSPM_Rich_map[[i]] <- Cnorm(predPLSPM_Rich)
PLSPM_Depth_map[[i]] <- Cnorm(predPLSPM_Depth)
PLSPM_C_map[[i]] <- Cnorm(predPLSPM_C)
RF_FC_map[[i]] <- Cnorm(predRF_FC)
RF_BM_map[[i]] <- Cnorm(predRF_BM)
RF_Rich_map[[i]] <- Cnorm(predRF_Rich)
RF_Depth_map[[i]] <- Cnorm(predRF_Depth)
RF_C_map[[i]] <- Cnorm(predRF_C)
print(i)
}
# stop parallel process
stopCluster(cl)
## prepare maps
PLSPM_FC_stack <- stack(PLSPM_FC_map)
PLSPM_BM_stack <- stack( PLSPM_BM_map)
PLSPM_Rich_stack <- stack(PLSPM_Rich_map)
PLSPM_Depth_stack <- stack(PLSPM_Depth_map)
PLSPM_C_stack <- stack(PLSPM_C_map)
RF_FC_stack <- stack(RF_FC_map)
RF_BM_stack <- stack(RF_BM_map)
RF_Rich_stack <- stack(RF_Rich_map)
RF_Depth_stack <- stack(RF_Depth_map)
RF_C_stack <- stack(RF_C_map)
# median maps
PLSPM_FC_stack_median <- calc(PLSPM_FC_stack, median)
PLSPM_BM_stack_median <- calc(PLSPM_BM_stack, median)
PLSPM_Rich_stack_median <- calc(PLSPM_Rich_stack, median)
PLSPM_Depth_stack_median <- calc(PLSPM_Depth_stack, median)
PLSPM_C_stack_median <- calc(PLSPM_C_stack, median)
RF_FC_stack_median <- calc(RF_FC_stack, median)
RF_BM_stack_median <- calc(RF_BM_stack, median)
RF_Rich_stack_median <- calc(RF_Rich_stack, median)
RF_Depth_stack_median <- calc(RF_Depth_stack, median)
RF_C_stack_median <- calc(RF_C_stack, median)
# coefficient of variation maps
PLSPM_FC_stack_CV <- calc(PLSPM_FC_stack, sd)/mean(Scores_all$FC)
PLSPM_BM_stack_CV <- calc(PLSPM_BM_stack, sd)/mean(Scores_all$BM)
PLSPM_Rich_stack_CV <- calc(PLSPM_Rich_stack, sd)/mean(Scores_all$Rich)
PLSPM_Depth_stack_CV <- calc(PLSPM_Depth_stack, sd)/mean(Scores_all$Depth)
PLSPM_C_stack_CV <- calc(PLSPM_C_stack, sd)/mean(data$Carbono_Subterraneo_kg_m2)
RF_FC_stack_CV <- calc(RF_FC_stack, sd)/mean(Scores_all$FC)
RF_BM_stack_CV <- calc(RF_BM_stack, sd)/mean(Scores_all$BM)
RF_Rich_stack_CV <- calc(RF_Rich_stack, sd)/mean(Scores_all$Rich)
RF_Depth_stack_CV <- calc(RF_Depth_stack, sd)/mean(Scores_all$Depth)
RF_C_stack_CV <- calc(RF_C_stack, sd)/mean(data$Carbono_Subterraneo_kg_m2)
save.image(saveRasters)
# ====================================================
# Hybrid model estimation using PLSPM and RF
# ====================================================
r <- list()
for (i in 1:500) {
# create random numbers with replacement to select samples from each group
idx = sample(1:N, N, replace = TRUE)
# select subsets of the five groups based on the random numbers
train1 <- data[idx, ] # for PLSPM
train2 <- data3[idx, ] # for RF
val1 <- data2[-idx, ] # for PLSPM
val2 <- data3[-idx, ] # for RF
### Run PLSPM for aboveground C stock
PLSrun = plspm(train, inner, outer, modes, maxiter = 1000, boot.val = F, br = 1000,
scheme = "factor", scaled = T)
# model scores
Scores <- as.data.frame(PLSrun$scores)
# rescaled.run <- plspm::rescale(PLSrun)
PLSrun$outer_model
p <- PLSrun$inner_mode$C[1] + DCM2 * PLSrun$inner_mode$C[2] + (PLSR_BM_stack_median *
PLSrun$inner_mode$C[3] + PLSR_Rich_stack_median * PLSrun$inner_mode$C[4])
r[[i]] <- Cnorm(p)
print(i)
}
#### get hybrid maps
C_stack_indirect <- stack(r)
C_indirect_median <- calc(C_stack_indirect, median)
C_indirect_cv <- calc(C_stack_indirect, sd)/mean(data$Carbono_Subterraneo_kg_m2)
plot(C_indirect_cv)