280 likes | 403 Views
ENVIRONMENTAL LAYERS MEETING IPLANT TUCSON 2012-04-03 Roundup Benoit Parmentier. What I have been doing working on: 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
E N D
ENVIRONMENTAL LAYERS MEETING IPLANT TUCSON 2012-04-03 Roundup Benoit Parmentier
What I have been doing working on: • 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 • 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 Moving beyond aggregate statistic… • 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
Fit residuals from gwr using 20100701 Run 6-Fit residuals from gwr using 20100502 Potentially useful to have the 2 sd thresholds…
Run 3-Fit residuals from gwr using 20100301 NO KRIGED FIT
Run 8-Fit residuals from gwr using 20100301 Run 2-Fit residuals from gwr using 20100102
Run 8-Fit residuals from gwr using 20100301 Run 9-Fit residuals from gwr using 20100102 Run 1-Fit residuals from gwr using 20100102
LST DAILY MEAM PRODUCTION • Python script (with IDRISI API but with GDAL in mind) to calculate: • Daily mean • Number of valid observation per day.
WORKFLOWDAILY MEAN CALCULATION Downloading Missing Data Assessment MOD11A1 hdf Mosaic Reprojection OR83M .rst QC flags Level 1 and 2 Masking Low quality Part of the process is automated in python with IDRISI API. Daily Mean Daily Valid Obs.
OREGON- DAILY MEAN FOR DOY 001 mean_day001_rescaled.rst
OREGON-NUMBER OF VALID OBSERVATION FOR DOY 001 mean_day_valid_obs_001_Sum.rst
OREGON- DAILY MEAN FOR DOY 182 mean_day182_rescaled.rst
OREGON-NUMBER OF VALID OBSERVATION FOR DOY 182 mean_day_valid_obs_182_Sum.rst
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 2002-03-28 Missing DOY 166 to 181: 2001-06-15 to 2001-07-02 (with July 01 missing 2) Missing DOY 301 to 305 Missing 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
LAND COVER CONSENSUS CATEGORIES 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.
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.
What's next..? • Continue the Visualization of RMSE for GAM and GWR • 2) Influence of sampling on results • GWR • Prediction • Producing LST monthly • 4) GAM using LST and Land Cover • 5) Use Kriging and co-kriging to predict tmax • 6) Documentation of the analysis