Joint International Surface Working Group and Satellite Applications Facility on Land Surface Analysis Workshop, IPMA, Lisboa, 26-28 June 2018
Introduction Soil moisture Evapotranspiration Future plan
1. Introduction COMS status KMA hydrological products
Located in 128.2 E 4
- Combining Geo and Leo satellites 5
2. Hydrological Variables Soil Moisture Evapotranspiration
LST (COMS), NDVI (MODIS), Rainfall (AWS) LST spatial interpolation using Kriging method NDVI Low-peak correction Rainfall weighted average using Gaussian function Statistics analysis of In-situ Soil moisture NDVI Precipitation 7
CDF matching between AMSR-2 and GLDAS Comparisons AMSR2 and GLDAS to In-situ GCOM-W1/AMSR2 (local time is 13:30) GLDAS every 3h (0600 UTC) Collocated data using nearest neighborhood method AMSR2 : 7 days moving average 8
Linear Regression+CDF R: 0.692 RMSE: 0.052 Bias: -0.01165 In-situ GLDAS AMSR2 CDF Linear Regression R: 0.707 RMSE: 0.050 Bias: -0.00881 Time series Upper(Reg+CDF): similar variation with in-situ SM Lower(Reg): do not considered low SM domain(red arrow) Scatter plot Upper(Reg+CDF): Reasonable scatter, well matched one-to-one line Lower(Reg): overestimation < 0.2 m 3 /m 3, underestimation > 0.3 m 3 /m 3 9
Using R studio, H2O library Feed-forward network Supervised (using Truth) feedback propagation Hidden layer(3), Hidden node(100), Epochs(3000) Gaussian distribution (response variable distribution) Rectifier activation function 10
Grid search with early stopping Tolerance: 0.005 (improvement < 0.5% stop training, Mean Squared Error) Rounds: 3(average of recent 3 rounds compared with average of prior 3 rounds) Hidden RMSE R 2 R Epochs Hidden RMSE R 2 R Epochs [600, 600, 600, 600] 3.8246 0.7953 0.8918 480 [400, 400, 400, 400, 400] 3.9760 0.7788 0.8825 379 [700, 700, 700, 700] 3.8633 0.7911 0.8895 582 [400, 400, 400] 4.0020 0.7759 0.8808 546 [700, 700, 700] 3.8698 0.7904 0.8891 477 [300, 300, 300, 300] 4.0023 0.7758 0.8808 370 [600, 600, 600] 3.8750 0.7899 0.8887 429 [500, 500, 500, 500] 3.8809 0.7892 0.8884 435 [700, 700, 700, 700, 700, 700, 700] [400, 400, 400, 400, 400, 400, 400] 4.0581 0.7695 0.8772 333 4.0602 0.7693 0.8771 249 [600, 600, 600, 600, 600] 3.8810 0.7892 0.8884 353 [400, 400, 400, 400, 400, 400] 4.0716 0.7680 0.8764 339 [300, 300, 300, 300, 300] 3.9284 0.7840 0.8855 491 [500, 500, 500, 500, 500, 500] 4.0757 0.7675 0.8761 343 [600, 600, 600, 600, 600, 600] 3.9294 0.7839 0.8854 564 [600, 600, 600, 600, 600, 600, 600] 4.0773 0.7674 0.8760 414 [500, 500, 500] 3.9456 0.7822 0.8844 659 [300, 300, 300, 300, 300, 300] 4.1134 0.7632 0.8736 329 [400, 400, 400, 400] 3.9590 0.7807 0.8836 413 [500, 500, 500, 500, 500] 3.9639 0.7801 0.8832 372 [500, 500, 500, 500, 500, 500, 500] [300, 300, 300, 300, 300, 300, 300] 4.1235 0.7621 0.8730 304 4.1418 0.7599 0.8717 295 [700, 700, 700, 700, 700, 700] 3.9669 0.7798 0.8831 317 [300, 300, 300] 4.1574 0.7581 0.8707 493 [700, 700, 700, 700, 700] 3.9736 0.7790 0.8826 396 11
Youngcheon (R: 0.87/RMSE: 3.10/Bias: -0.157) Deep learning SM Jincheon (R: 0.68/RMSE: 4.03/Bias: 0.376) DL GLDAS AMSR2 R 0.89 0.36 0.22 RMSE 3.79 9.07 20.64 Bias -0.07-4.20-17.77 DL: relatively well matched with in-situ GLDAS: relatively high correlation with some bias AMSR2: different pattern with large bias 12
Based on the Surface Energy Equation COMS ET is characterized by simply simulated aerodynamic resistance, considering a variety of surface roughness length in the sensible heat flux (the so-called B-method of Jackson et al (1977)). Mapping Area : 20~50 N, 100~145 E (only clear skies, except on urban, desert area) Temporal/Spatial resolution : 1-day/1 km Estimation Period Z 0 Canopy height high: : April 1 ST 2011 ~ present Ecoclimap - Surface roughness (monthly) Z 0 Canopy height low: Z 0 =exp(-5.5+5.8 NDVI) (Gupta et al., 2002) 13
2012 COMS ET MODIS ET GLDAS ET N.obs. 125 300 361 R 0.598 0.696 RMSE 0.941 1.741 0.890 Bias 1.05 2.47 1.116 COMS ET MODIS GLDAS N. obs. 132 301 363 R 0.737 0.620 RMSE 1.122 0.853 0.710 Bias 1.205 1.182 0.923 (2012.08.07) COMS daily ET MODIS AET MODIS PET GLDAS ET 14
R studio, H2O library/ Feed-forward network Supervised (using Truth) Error back propagation Hidden layer(3), Hidden node(72), Epochs(3000) Gaussian distribution (response variable distribution) Rectifier activation function 15
COMS ET Himawari-8 (Ch12) 2016.04.05 06 UTC 16
Drought index (%) = AET/PET (Nicolas Ghilain, Royal Meteorological Institute, 2014) Actual Evapotranspiration : COMS daily Evapotranspiration Potential Evapotranspiration: COMS priestly-taylor Evapotranspiration (using COMS Ta, ins) - Simplified input data for estimating of evapotranspiration - Reference evapotranspiration (in grass conditions) - Proposed for application in humid areas 2012.06.09. COMS AET 16days α: Priestley-Taylor coefficient (dimensionless) : slope of the saturation vapour pressure curve (kpa C -1 ) γ: Psychrometric constant (kpa C -1 ) Rn : Net radiation (MJ m -2 day -1 ) G : Soil heat flux (MJ m -2 day -1 ) λ : Latent heat of vaporization 2012.06.09. COMS PET 16days 17
Reference data : MODIS VHI (Vegetation Health index, 16days) Satellite based meteorological-agriculture drought index Reflect on the Vegetation Condition Index (VCI) and Temperature Condition Index (TCI) Fig.1. KMA Drought Index. COMS AET/PET ratio : 16 days MODIS VHI : 16 days Fig.2. KMA average temperature of temperature anomaly during 30years Fig.3. KMA average rainfall of Rainfall anomaly during 30years The Possibility of COMS evapotranspiration Plan to comparison of land state index (LST, NDVI, NDWI, etc) 18
3. Future plan
COMS-2A is scheduled to launch on November 2018 COMS-2A will have 52 baseline products including hydrological variables GK-2A Products - Soil moisture, Vegetation Health Index - Land surface temperature, - Vegetation (VI, FVC), Snow (snow, snow depth) - Precipitation (potential precipitation, rainfall) - Radiation (upward, downward solar radiation, longwave radiation) - Flood detection Soil moisture - MW based on soil moisture : Low orbital satellite (AMSR2, SMOS, ASCAT) - IR based soil moisture : Geostationary satellite (COMS) - Deep learning based soil moisture using Himawari -8 Evapotranspiration - IR based evapotranspiration : COMS - Deep learning based evapotranspiration using Himawari-8(plan) Vegetation Index Fraction Vegetation Cover VHI (drougt index) IR based soil moisture Flood detection 20