1
|
&args ingrid sd prob bbox:rest
|
2
|
|
3
|
/* 9a - limit to 4 steps, see if that has any significant deterioration of smoothing performance. Should fix
|
4
|
/* the problem with islands and headlands - turns out also need to remove water except
|
5
|
/* for one cell adjacent to land, and give that a higher uncertainty. See cluster_multiscalesmooth9a_clean_216
|
6
|
|
7
|
/* Version 9:
|
8
|
/* Focus on implementing identical algorithm to directsmooth2 using multiscale method i.e. aggregating by factor of 3
|
9
|
/* from already aggregated data, rather than from original resolution each time.
|
10
|
|
11
|
/* Version 8:
|
12
|
/* WARNING: have not yet checked that the additional weighting of the gaussian smoothing is not messing with
|
13
|
/* the calculations of variance etc.
|
14
|
/* Replaced simple 3x3 aggregation with gaussian smoothing
|
15
|
/* Kernel is chosen to give appropriate level of smoothing and produce a fairly accurate approximation of the smoothed
|
16
|
/* surface by interpolation of the coarser scale smoothed values
|
17
|
/* Details in gaussian.xls on john's laptop
|
18
|
|
19
|
/* Version 7:
|
20
|
/* Further reworking of how the final values are selected - a mean is a candidate if its associated variance
|
21
|
/* is less than the mean sample uncertainty, and the mean with the lowest variance among the candidates is the chosen one.
|
22
|
/* Implement that in the nested sense by taking the lowest group variance divided by the chi^2 value, and its associated mean variance,
|
23
|
/* and if that is lower than the data point variance the
|
24
|
|
25
|
/* approximate critical value of chi^2/N with N degrees of freedom at 5% level as 1 + 2.45/sqrt(N) + 0.55/N
|
26
|
/* for 1% level use 1 + 3.4/sqrt(N) + 2.4/N
|
27
|
|
28
|
/* Version 6:
|
29
|
/* Done from scratch after careful working through of theory.
|
30
|
/* ingrid is the (potentially sparse) grid of data to be smoothed and
|
31
|
/* interpolated, which can be of different extent to the output
|
32
|
/* (resolution is assumed to be the same, could adapt this to relax that)
|
33
|
/*
|
34
|
/* var can be a constant or a grid
|
35
|
/*
|
36
|
/* bbox can be either a grid name or the 'xmin ymin xmax ymax' parameters
|
37
|
/* for setwindow
|
38
|
|
39
|
&type NB - using standard deviation as noise specification now, not variance!
|
40
|
|
41
|
|
42
|
/* set up chisq parameters
|
43
|
&sv chisqa = [calc 2.807 - 0.6422 * [log10 %prob% ] - 3.410 * %prob% ** 0.3411 ]
|
44
|
&sv chisqb = [calc -5.871 - 3.675 * [log10 %prob% ] + 4.690 * %prob% ** 0.3377 ]
|
45
|
&type chisq parameters %chisqa% %chisqb%
|
46
|
|
47
|
|
48
|
setcell %ingrid%
|
49
|
|
50
|
/* work out maximum of ingrid and bbox extents
|
51
|
setwindow [unquote %bbox%] %ingrid%
|
52
|
bboxgrid = 1
|
53
|
setwindow maxof
|
54
|
workgrid = %ingrid% + bboxgrid
|
55
|
setwindow workgrid
|
56
|
kill workgrid
|
57
|
|
58
|
/* naming:
|
59
|
/* h - the value being smoothed/interpolated
|
60
|
/* vg - total variance of group of data, or of individual measurement
|
61
|
/* v_bg - variance between groups
|
62
|
/* v_wg - variance within groups
|
63
|
/* wa - weighting for aggregation, based on total variance
|
64
|
/* vm - variance of the calculated mean
|
65
|
/* mv - mean of finer scale variances
|
66
|
/* n - effective number of measurements
|
67
|
|
68
|
|
69
|
/* NB - only calculating sample variances here, not variances of estimated means.
|
70
|
/* Also note that v0_bg is an uncertainty, not a sample variance
|
71
|
/* and v1_bg is total variances, but both are labelled as "between-group" to simplify the smoothing
|
72
|
|
73
|
h0 = %ingrid%
|
74
|
v0 = con(^ isnull(h0), sqr(%sd%))
|
75
|
vg0 = v0
|
76
|
w0 = con(isnull(v0), 0, 1.0 / v0)
|
77
|
wsq0 = sqr(w0)
|
78
|
n0 = con(^ isnull(h0), 1, 0)
|
79
|
|
80
|
&describe v0
|
81
|
&sv bigvar %grd$zmax%
|
82
|
|
83
|
setcell minof
|
84
|
|
85
|
/* aggregate to broader scales
|
86
|
&sv i 1
|
87
|
&sv done .false.
|
88
|
&describe h0
|
89
|
|
90
|
&do &until %done%
|
91
|
&sv j [calc %i% - 1]
|
92
|
|
93
|
&type Aggregate from %j% to %i%
|
94
|
|
95
|
&describe h%j%
|
96
|
&sv cell3 [calc %grd$dx% * 3]
|
97
|
&describe h0
|
98
|
&sv nx0 [round [calc %grd$xmin% / %cell3% - 0.5]]
|
99
|
&sv ny0 [round [calc %grd$ymin% / %cell3% - 0.5]]
|
100
|
&sv nx1 [round [calc %grd$xmax% / %cell3% + 0.5]]
|
101
|
&sv ny1 [round [calc %grd$ymax% / %cell3% + 0.5]]
|
102
|
&sv x0 [calc ( %nx0% - 0.5 ) * %cell3%]
|
103
|
&sv y0 [calc ( %ny0% - 0.5 ) * %cell3%]
|
104
|
&sv x1 [calc ( %nx1% + 0.5 ) * %cell3%]
|
105
|
&sv y1 [calc ( %ny1% + 0.5 ) * %cell3%]
|
106
|
setwindow %x0% %y0% %x1% %y1%
|
107
|
|
108
|
w%i% = aggregate(w%j%, 3, sum)
|
109
|
wsq%i% = aggregate(wsq%j%, 3, sum)
|
110
|
n%i% = aggregate(n%j%, 3, sum)
|
111
|
neff%i% = w%i% * w%i% / wsq%i%
|
112
|
h%i% = aggregate(w%j% * h%j%, 3, sum) / w%i%
|
113
|
vbg%i% = aggregate(w%j% * sqr(h%j% - h%i%), 3, sum) / w%i%
|
114
|
&if %i% eq 1 &then vwg%i% = n%i% - n%i% /* zero, but with window and cell size set for us
|
115
|
&else vwg%i% = aggregate(w%j% * vg%j%, 3, sum) / w%i%
|
116
|
vg%i% = vbg%i% + vwg%i%
|
117
|
vm%i% = 1.0 / w%i%
|
118
|
mv%i% = n%i% / w%i%
|
119
|
|
120
|
chisq%i% = 1 + %chisqa% / sqrt(neff%i% - 1) + %chisqb% / (neff%i% - 1)
|
121
|
v%i% = con(vg%i% / chisq%i% < mv%i%, vm%i%, vg%i%)
|
122
|
|
123
|
/* remove everything except h%i% and v%i%
|
124
|
kill w%j%
|
125
|
kill wsq%j%
|
126
|
kill n%j%
|
127
|
kill neff%i%
|
128
|
kill vbg%i%
|
129
|
kill vwg%i%
|
130
|
kill vg%j%
|
131
|
kill vm%i%
|
132
|
kill mv%i%
|
133
|
kill chisq%i%
|
134
|
|
135
|
&sv done %i% eq 4
|
136
|
|
137
|
&sv i [calc %i% + 1]
|
138
|
&end
|
139
|
|
140
|
|
141
|
&sv maxstep [calc %i% - 1]
|
142
|
&sv bigvar [calc %bigvar% * 10]
|
143
|
|
144
|
kill w%maxstep%
|
145
|
kill wsq%maxstep%
|
146
|
kill n%maxstep%
|
147
|
kill vg%maxstep%
|
148
|
|
149
|
/* smooth, refine and combine each layer in turn
|
150
|
|
151
|
|
152
|
copy h%maxstep% hs%maxstep%
|
153
|
copy v%maxstep% vs%maxstep%
|
154
|
kill h%maxstep%
|
155
|
kill v%maxstep%
|
156
|
setcell hs%maxstep%
|
157
|
setwindow hs%maxstep%
|
158
|
|
159
|
&do j := %maxstep% &to 1 &by -1
|
160
|
&sv i [calc %j% - 1]
|
161
|
|
162
|
&type Refine from %j% to %i%
|
163
|
|
164
|
/* for the first stage where the coarser grid is refined and smoothed, set window to the coarse grid
|
165
|
setcell h%i%
|
166
|
setwindow maxof
|
167
|
|
168
|
/* create smoothed higher resolution versions of h and v_bg, hopefully with no nulls!
|
169
|
hs%j%_%i% = focalmean(hs%j%, circle, 2)
|
170
|
vs%j%_%i% = focalmean(vs%j%, circle, 2)
|
171
|
|
172
|
setcell h%i%
|
173
|
&describe h%i%
|
174
|
&sv cellsize %grd$dx%
|
175
|
&describe bboxgrid
|
176
|
setwindow [calc %grd$xmin% - 4 * %cellsize%] [calc %grd$ymin% - 4 * %cellsize%] [calc %grd$xmax% + 4 * %cellsize%] [calc %grd$ymax% + 4 * %cellsize%] h%i%
|
177
|
|
178
|
/* create no-null version of finer h and v
|
179
|
h%i%_c = con(isnull(h%i%), 0, h%i%)
|
180
|
v%i%_c = con(isnull(v%i%), %bigvar%, v%i%)
|
181
|
|
182
|
/* combine two values using least variance
|
183
|
hs%i% = (h%i%_c / v%i%_c + hs%j%_%i% / vs%j%_%i% ) / (1.0 / v%i%_c + 1.0 / vs%j%_%i%)
|
184
|
vs%i% = 1 / (1.0 / v%i%_c + 1.0 / vs%j%_%i%)
|
185
|
|
186
|
kill v%i%_c
|
187
|
kill h%i%_c
|
188
|
kill v%i%
|
189
|
kill h%i%
|
190
|
kill vs%j%_%i%
|
191
|
kill hs%j%_%i%
|
192
|
kill hs%j%
|
193
|
kill vs%j%
|
194
|
&end
|
195
|
|
196
|
/* result is hs0, with variance vs0
|
197
|
|
198
|
kill bboxgrid
|