-
Notifications
You must be signed in to change notification settings - Fork 7
/
mapping_example.R
299 lines (232 loc) · 11.3 KB
/
mapping_example.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
# Example scenario of a resampling fraemwork for making a map -------------
library(raster)
library(MASS)
library(data.table)
library(dplyr)
# note the function masking...
# use dplyr::select or raster::select or MASS:select is there are issues (i've loaded dplyr last)
# load data ---------------------------------------------------------------
# load survey data
survey_points_raw <- read.csv("classification_data/observer_class_ads40.csv",
header = T, stringsAsFactors = F)
survey_points <- survey_points_raw %>%
filter(veg_class %in% c("bt", "ew", "ttt", "wh"),
study_area %in% c(12,10)) %>%
select(study_area, veg_class, blue_mean:nir_mean) %>%
mutate(veg_class = as.factor(veg_class),
id = 1:nrow(.)) %>%
na.omit()
rm(survey_points_raw)
# load image data
blue_band <- raster("classification_data/blue_mapeg.tif") # so we have dimentions/CRS to use
image_data <- data.frame(as.data.frame(raster("classification_data/blue_mapeg.tif")),
as.data.frame(raster("classification_data/green_mapeg.tif")),
as.data.frame(raster("classification_data/red_mapeg.tif")),
as.data.frame(raster("classification_data/nir_mapeg.tif")))
names(image_data) <- c("blue_mean","green_mean","red_mean","nir_mean") # so they match the modelling data names
# fill the few NA's (deal with them better for serious use)
backfill <- function(x) {
x[which(is.na(x))] <- x[which(is.na(x))+1]
x
}
forwardfill <- function(x) {
x[which(is.na(x))] <- x[which(is.na(x))-1]
x
}
image_data <- as.data.frame(lapply(image_data, backfill))
image_data <- as.data.frame(lapply(image_data, forwardfill))
# resampling model and map ------------------------------------------------
# this is a simplified function to demonstrate a resampling framework
# change the paramers directly in this function for different resampling designs
# for more serious use, substitue in functions from "allocation_functions.R" and "accuracy_functions.R"
# for even more serious use, investigate existing packages, such as sperrorest (https://cran.r-project.org/web/packages/sperrorest/index.html)
fit_and_map <- function(niter, data, image_data) {
print(paste0("Fitting iteration ",niter))
# get training sample (simple random sample, Monte Carlo cross-validation with 67:33 split ratio)
train <- (data %>% select(id) %>% sample_frac(0.67))$id
# get test from left overs
test <- data$id[!data$id %in% train]
# fit ML model to training data
fm <- lda(veg_class ~ blue_mean + green_mean + red_mean + nir_mean,
data = inner_join(data, data.frame(id=train), by="id"),
prior = rep(1/length(unique(data$veg_class)), length(unique(data$veg_class))))
# predict classes for test data
test_preds <- predict(fm, newdata = inner_join(data, data.frame(id=test), by="id"))$class
# get percentage agreement accuracy
perc_agr <- sum(as.character(test_preds) == as.character(data$veg_class[test])) / length(test_preds)
# predict classes for image data
# this is possibly a very slow way to do raster predictions - investigate other options for serious use
image_preds <- predict(fm, newdata = image_data)$class
# # calculate conf matrix
# conf_mat <- table(as.character(test_preds),
# as.character(data$veg_class[test]))
# match-up for conf_mat and single-run bootstrap
test_true <- data.frame(test = as.character(test_preds),
true = as.character(data$veg_class[test]),
stringsAsFactors = F)
# return a list of the accuracy and the predicitons and the confusion matrix counts
list(perc_agr, image_preds, test_true)
}
# fit n iterations of the resmapling routine
mapping_runs <- lapply(1:800, fit_and_map, survey_points, image_data)
saveRDS(mapping_runs, file = "A:/1_UNSW/0_data/Dharawal_project/mapping_runs.rds")
mapping_runs <- readRDS("A:/1_UNSW/0_data/Dharawal_project/mapping_runs.rds")
# calculate resampling stats ----------------------------------------------
med_ci <- function(x) {
med <- median(x)
names(med) <- "median"
ci <- quantile(x, c(0.025,0.975))
round(c(med,ci)*100, 0)
}
# mapped area as per Olofsson et al. 2014;
mapped_areas <- function(idx, conf_mat_list, area_tables) {
# util function
dim_check <- function(x, len = 4) { # DANGER - hard coded to 4 classes, use len = if error matrix is different size
dim(x)[1] != len | dim(x)[2] != len
}
# pixel counts
mapped_areas <- as.numeric(area_tables[[idx]])
if (length(mapped_areas) != 4) {return(rep(NA,4))}
mapped_areas_p <- mapped_areas / sum(mapped_areas)
# p_ij matrix
if (dim_check(conf_mat_list[[idx]])) {return(rep(NA,4))}
conf_mat_nij <- matrix(as.vector(prop.table(conf_mat_list[[idx]], 1)), nrow = 4, ncol = 4)
conf_mat_pij <- matrix(mapped_areas_p, nrow = 4, ncol = 4) * conf_mat_nij # DANGER - hard coded to 4 classes
# area estimates
areas_pk <- colSums(conf_mat_pij)
area_estimates <- areas_pk * sum(mapped_areas)
area_estimates / sum(area_estimates)
}
# get vector of accuracy stats and get mean/intervals
accuracy_distribution <- unlist(lapply(mapping_runs, `[[`, 1))
med_ci(accuracy_distribution)
# get class areas and get mean/intervals
class_areas <- lapply(mapping_runs, `[[`, 2)
class_areas <- lapply(class_areas, table)
conf_mat_list <- lapply(mapping_runs, `[[`, 3)
conf_mat_list <- lapply(conf_mat_list, table)
mapped_areas_list <- lapply(1:length(class_areas), mapped_areas, conf_mat_list, class_areas)
class_proportions <- list(
Banksia = unlist(lapply(mapped_areas_list, `[[`, 1)),
Eucalypt = unlist(lapply(mapped_areas_list, `[[`, 2)),
Teatree = unlist(lapply(mapped_areas_list, `[[`, 3)),
Wetheath = unlist(lapply(mapped_areas_list, `[[`, 4)))
lapply(class_proportions, med_ci)
# calculate simultaneous intervals ----------------------------------------
# As per Olofsson et al. 2014
# Olofsson, P., Foody, G. M., Herold, M., Stehman, S. V., Woodcock, C. E., & Wulder, M. A. (2014). Good practices for estimating area and assessing accuracy of land change. Remote Sensing of Environment, 148, 42-57.
### choose an iteration (close to median value?) to do simultaneous intervals on
eg_run <- 17 # run chosen for the simultaneous example
test_true <- mapping_runs[[eg_run]][[3]]
conf_mat <- table(test_true$test, test_true$true)
mapped_areas <- table(mapping_runs[[eg_run]][[2]])
mapped_areas_p <- as.numeric(mapped_areas / sum(mapped_areas))
# p_ij matrix
conf_mat_nij <- prop.table(conf_mat, 1)
conf_mat_pij <- matrix(mapped_areas_p, nrow = 4, ncol = 4) * conf_mat_nij
# overall accuracy
oa <- sum(diag(conf_mat_pij))
# ovarall accuracy variance
users <- diag(conf_mat_pij) / rowSums(conf_mat_pij)
oa_v <- sum(((mapped_areas_p^2 * users) * (1 - users)) / (rowSums(conf_mat) - 1))
# overall accuracy 95% CI
oa_95 <- 1.96 * sqrt(oa_v * oa)
# area estimates
areas_pk <- colSums(conf_mat_pij)
areas_est <- areas_pk * sum(mapped_areas)
# area estimate standard error
areas_std_err <- sqrt(
colSums(mapped_areas_p^2 *
((conf_mat_nij * (1 - conf_mat_nij)) /
(rowSums(conf_mat) - 1)))
)
# areas 95% CI
areas_95 <- 1.96 * (areas_std_err * sum(mapped_areas))
areas_95_p <- areas_95 / sum(mapped_areas) # duh
# calculate single-run bootstrap intervals --------------------------------
# sensu Hess & Bay 1997
# Hess, G. R., & Bay, J. M. (1997). Generating confidence intervals for composition-based landscape indexes. Landscape Ecology, 12(5), 309-320.
sampled_accuracies <- function(test_true) {
data <- test_true %>%
sample_frac(1, replace = T) # this is the bootstrap
conf_mat <- table(data$test, data$true)
list(perc_agr = sum(diag(conf_mat)) / sum(conf_mat),
user = diag(conf_mat) / rowSums(conf_mat))
}
singlerun_resample <- replicate(n = 1000,
expr = {sampled_accuracies(test_true)},
simplify = F)
singlerun_oa <- unlist(lapply(singlerun_resample, `[[`, 1))
#### not sure about this - needs more thought re theory, but should ~ do
#########################################################################
singlerun_users_list <- lapply(singlerun_resample, `[[`, 2)
singlerun_users_df <- data.frame(
Banksia = unlist(lapply(singlerun_users_list, `[[`, 1)),
Eucalypt = unlist(lapply(singlerun_users_list, `[[`, 2)),
Teatree = unlist(lapply(singlerun_users_list, `[[`, 3)),
Wetheath = unlist(lapply(singlerun_users_list, `[[`, 4)))
singlerun_area_95_sim <- sqrt(
colSums(mapped_areas_p^2 *
cov(singlerun_users_df))) * 1.96
singlerun_error_df <- 1 - singlerun_users_df
singlerun_area_95 <- (unlist(lapply(singlerun_error_df, quantile, 0.975)) -
unlist(lapply(singlerun_error_df, quantile, 0.025))) * mapped_areas / sum(mapped_areas) / 2
########################################################################
# compare estimates -------------------------------------------------------
## accuracy
# full resample
med_ci(accuracy_distribution)
# bootstrap single run
med_ci(singlerun_oa); sd(singlerun_oa) * 1.96
# simultaneous single run
round(c(estimate=oa, `2.5%`=oa-oa_95, `97.5%`=oa+oa_95)*100)
round(c(oa, oa_95)*100)
# count error matrix accuracy
round(sum(diag(conf_mat)) / sum(conf_mat) * 100)
## areas
# full resample
lapply(class_proportions, med_ci)
# bootstrap single run CIs
round(singlerun_area_95*100)
# simultaneous single run
data.frame(areas_est = areas_pk, `CI_5` = areas_pk-areas_95_p, `CI_95` = areas_pk+areas_95_p)
round(areas_pk*100)
round(areas_95_p*100)
# mapped pixel count areas
round(mapped_areas_p*100)
# make maps ---------------------------------------------------------------
# get data frame of map predicitons
map_distribution <- as.data.frame(lapply(mapping_runs, `[[`, 2))
# check levels so we know what the final raster values equate to
levels(map_distribution[,1])
# convert to integer
map_distribution <- as.data.frame(lapply(map_distribution, as.integer))
names(map_distribution) <- NULL
# calculate the mode/count for the output map
# make a function for finding the mode
Mode <- function(x) {
ux <- unique(x)
uxt <- tabulate(match(x, ux))
return(list(mode = ux[which.max(uxt)],
n = max(uxt)))
}
map_mode <- apply(map_distribution, 1, Mode)
# prepare the mode matrix and make into a raster
mode_matrix <- matrix(unlist(lapply(map_mode, `[[`, 1)), nrow = nrow(blue_band), ncol = ncol(blue_band), byrow = T)
final_mode_map <- raster(mode_matrix)
extent(final_mode_map) = extent(blue_band)
crs(final_mode_map) = crs(blue_band)
writeRaster(final_mode_map, filename = "plots/final_mode_map.tif", overwrite = T)
# prepare the confidence matrix and make into a raster
mode_count_matrix <- matrix(unlist(lapply(map_mode, `[[`, 2)), nrow = nrow(blue_band), ncol = ncol(blue_band), byrow = T)
mode_count_matrix <- mode_count_matrix / length(mapping_runs)
final_confidence_map <- raster(mode_count_matrix)
extent(final_confidence_map) = extent(blue_band)
crs(final_confidence_map) = crs(blue_band)
writeRaster(final_confidence_map, filename = "plots/final_confidence_map.tif", overwrite = T)
# prepare a single run matrix and make into a raster
singlerun_matrix <- matrix(map_distribution[[1]], nrow = nrow(blue_band), ncol = ncol(blue_band), byrow = T)
singlerun_map <- raster(singlerun_matrix)
extent(singlerun_map) = extent(blue_band)
crs(singlerun_map) = crs(blue_band)
writeRaster(singlerun_map, filename = "plots/singlerun_map.tif", overwrite = T)