Upload
mara-radley
View
215
Download
0
Tags:
Embed Size (px)
Citation preview
What I have been doing working on:
1) Visualization of RMSE fit for Geographically Weighted Regression •Writing a code in R to visualize the RMSE using- Stations location- Kriged error surface from stations
2) Producing LST daily mean Python script (with IDRISI API but with GDAL in mind) to calculate:- Daily mean- Number of valid observation per day.
3) GAM prediction• Some GAM predictions with interaction terms• Including daily mean LST and LC in the GAM regression
1) Visualization of RMSE fit for Geographically Weighted Regression •Writing a code in R to visualize the RMSE using- Stations location- Kriged error surface from stations
1)VISUALIZATION OF RMSE Moving beyond aggregate statistic…
0
5
10
15
20
25
30
35
40
45
RM
SE fi
t (d
eg
C *
10
)
Interpolation Date
RMSE FIT USING GWR WITH 30% RETAINED FOR VALIDATION
Fit residuals from gwr using 20100701Run 6-Fit residuals from gwr using 20100502
run dates ns RMSE_gwr16 20100502 114 21.33324
Potentially useful to have the 2 sd thresholds…
Run 3-Fit residuals from gwr using 20100301
NO KRIGED FIT
run dates ns RMSE_gwr13 20100301 120 18.19032
Run 8-Fit residuals from gwr using 20100301Run 2-Fit residuals from gwr using 20100102
run dates ns RMSE_gwr12 20100102 115 23.73444
Run 8-Fit residuals from gwr using 20100301Run 9-Fit residuals from gwr using 20100102
Run 1-Fit residuals from gwr using 20100102
run dates ns RMSE_gwr11 20100101 113 32.1132
•Python script (with IDRISI API but with GDAL in mind) to calculate:- Daily mean- Number of valid observation per day.
LST DAILY MEAM PRODUCTION
MOD11A1hdf
OR83M.rst
MosaicReprojection
QC flagsLevel 1 and 2
Masking Low quality
Daily Mean Daily Valid Obs.
WORKFLOW DAILY MEAN CALCULATION
Part of the process is automated in python with IDRISI API.
DownloadingMissing Data Assessment
SUMMARY INFORMATION OF THE DAILY MEAN CALCULATION
A full assessment of the temporal and spatial distribution of mean would be necessary:- Most dates have 10 images (on average 9.88 images).- The number of valid values seems to be lower in Winter (more check needed).- Average per month may be quite helpful.
Missing data:
The average was done over the 2001-2010 time period and there were 45 missing images (out of a total of 3652).
Missing DOY 78 to 88: 2002-03-19 to 2002-03-28Missing DOY 166 to 181: 2001-06-15 to 2001-07-02 (with July 01 missing 2)Missing DOY 301 to 305Missing DOY 351 to 357: 2003-12-17 to 2003-12-23 (355 to 357 missing 2)
3)GAM MODELING USING LST AND LC
GAM regressions:• Some GAM predictions with interaction terms• Including daily mean LST and LC in the GAM regression
AggregatedClassification class
Class No.
GLC20001 UMD MODIS GlobCover2
Forest 1 1,2,3,4,5,6,7,8
1,2,3,4,5,6
1,2,3,4,5,8
40,50,60,70,90,100,160,170
Shrub 2 9,10,11,12,14 7,8,9 6,7,9 110,120,130,150Grass 3 13 10 10 140Crop 4 16 11 12 11,14Mosaic3 5 17,18 14 20,30Urban 6 22 13 13 190Barren 7 19 12 16 200Snow 8 21 15 220Wetland 9 15 11 180Water body 10 20 0 17 210
Table 5. Legend for the 10 aggregated land cover classes and the corresponding classes from the six individual global land cover legends. Modified from (Nakaegawa 2011).1I added class 3 to ‘forest’ since it was missing in original table. The class 2 entry under ‘shrub’ is probably an error and so is removed.2GlobCover class assignment needs to be finalized.3Mosaic is composed of cropland and natural vegetation.
LAND COVER CONSENSUS CATEGORIES
GAM MODELS USED FOR THIS ANALYSIS
mod1<- tmax~ s(lat) + s (lon) + s (ELEV_SRTM) mod2<- tmax~ s(lat,lon,ELEV_SRTM) mod3<- tmax~ s(lat) + s (lon) + s (ELEV_SRTM) + s (Northness)+ s (Eastness) + s(DISTOC) mod4<- tmax~ s(lat) + s (lon) + s(ELEV_SRTM) + s(Northness) + s (Eastness) + s(DISTOC) + s(LST) mod5<- tmax~ s(lat,lon) +s(ELEV_SRTM) + s(Northness,Eastness) + s(DISTOC) + s(LST)
mod6<- tmax~ s(lat,lon) +s(ELEV_SRTM) + s(Northness,Eastness) + s(DISTOC) + s(LST,LC1) mod7<- tmax~ s(lat,lon) +s(ELEV_SRTM) + s(Northness,Eastness) + s(DISTOC) + s(LST,LC3)
PROBLEM WITH MISSING DATA
If screening is used such as LST> 258 & LST<313)… the number of observations can drop to 48 and 20 for training and testing compared to 120 and 50 stations.