高分解能 GSMaP アルゴリズムの 構造と考え方 牛尾知雄 ( 大阪大 )
Curret GSMaP products GSMaP_MWR Microwave radiometer product GSMaP_MVK Global precipitatio mappig from microwave ad ifrared radiometric data GSMaP_Gauge Gauge adjusted GSMaP_MVK GSMaP_NRT ad GSMaP_Now Near real time versio of the GSMaP_MVK product GSMaP_Gauge_NRT Near real time product of GSMaP_Gauge 2
GSMaP sesors Microwave Imagers & Souders GPM-Core GMI GCOM-W AMSR2 DMSP SSM/I, SSMIS NOAA/MetOp AMSU GSMaP Microwave Radiometer Retrieval Algorithm IR Imagers Raifall Data from each Microwave Radiometer Merged Microwave Raifall Data Geostatioary Satellites Microwave-IR Merged Algorithm (CMV, K/F) Global Raifall Map + Gauge-calibrated Raifall Map (0.1 degree grid, Hourly)
Coverage of the 6 MWRs i 1 hour
6 hourly MWR combied map TMI AMSR & AMSR-E Combied 6 hourly SSM/I (F13, F14, F15)
GSMaP sesors Microwave Imagers & Souders GPM-Core GMI GCOM-W AMSR2 DMSP SSM/I, SSMIS NOAA/MetOp AMSU GSMaP Microwave Radiometer Retrieval Algorithm IR Imagers Raifall Data from each Microwave Radiometer Merged Microwave Raifall Data Geostatioary Satellites Microwave-IR Merged Algorithm (CMV, K/F) Global Raifall Map + Gauge-calibrated Raifall Map (0.1 degree grid, Hourly)
Forward process Rai area obtaied from the radiometer Swath of the sesor
Forward process forward 1hr 2hr 3hr 0hr Propagatig rai pixels alog with cloud motio vector every hour
Forward process forward 1hr 2hr 3hr 0hr Apply Kalma filter
Forward process forward 1hr 2hr 3hr 0hr Newly idetified rai area Revisted Path of the microwave sesor
Forward process forward 1 時 2 時 0 時 3 時
GSMaP_NRT
GSMaP_NRT This is the basic methodology for the GSMaP_NRT product. However, sometimes the error ca be large for this ear real time product. To reduce the bias error, ot oly the forward process but the backward process are itroduced i the GSMaP_MVK product.
Backward process forward 1 時 2 時 0 時 3 時 1 時 2 時 backward
Combie 1 時 2 時 0 時 3 時 Combie
t t+1 t+2 + + + = = =
GSMaP algorithm flow CPC Global Daily Gauge MWR data GSMaP Geo-IR image Atmospheric Motio Vector Kalma Filter GSMaP_NRT ad GSMaP_Now GSMaP_Gauge_NRT GSMaP_MVK Backward propagatio Optimizatio module GSMaP_Gauge
Compariso with radar-rai gauge etwork i Japa Rai rate [mm/h] 50 40 30 20 10 0 50 45 2005. 7. 8 ~10 th (UTC) lat. 33.1 lo. 130.9 RA GSMaP mwr 0 6 12 18 0 6 12 18 0 6 12 18 Time (hour) GSMaP_MVK [mm/h] 40 35 30 25 20 15 10 5 0 0 5 10 15 20 25 30 35 40 45 50 RadarAMeDAS [mm/h] GSMaP_MVK teds to uderestimate the rai fall rate with 1 hour resolutio But, basically the GSMaP_MVK s estimates strogly depeds o the estimates from microwave radiometer
What is our approach? Problem Potetial data sources for the gauge adjusted GSMaP_MVK product GPCC data set Mothly base/ 0.5 degree grid Coversio of the time resolutio of Mothly -> Hourly is difficult CPC global gauge data aalysis Solutio CPC global gauge data aalysis by Che et al. (2008) Daily base/ 0.5 degree grid
GSMaP_Gauge algorithm Data used i the GSMaP_Gauge algorithm NOAA CPC Uified Gauge-Based Aalysis of Global Daily Precipitatio (CPC). The data is a daily ad 0.5 grid precipitatio fields o a global basis. Based o the assumptio that the GSMaP_Gauge CPC Gauge data (Gauge term) has the gaussia distributio, the probability desity fuctio of the GSMaP_Gauge estimatio multiplied by the Gauge term is maximized
Example of GPM-GSMaP mothly Mothly Mea i April 2014(V03A) GSMaP_MVK Referece TRMM/PR 3A25 GSMaP_Gauge GPCC Gauge Mothly Moitorig GPCC: Global Precipitatio Climatology Ceter GPM-GSMaP_MVK hourly (00Z o April 1) NOAA/CPC Daily Gauge Aalysis CPC: Climate Predictio Ceter
Evaluatio of GPM-GSMaP GPM-GSMaP compared with GPCC Mothly accumulated Gauge data. GSMaP_Gauge (calibrated by NOAA CPC Daily Gauge Aalysis) shows 20-30% better accuracy i RMSE tha GSMaP_MVK. GSMaP_MVK Mothly Raifall [mm/moth] GSMaP_MVK April 2014 April 2014 GSMaP_Gauge Mothly Raifall [mm/moth] GSMaP_Gauge GPCC Mothly Raifall [mm/moth] GPCC Mothly raifall [mm/moth] GSMaP_MVK [mm/moth] GSMaP_Gauge [mm/moth] Bias R RMSE Bias R RMSE April 2014-0.54 0.73 62.45-15.58 0.88 44.26 May 2014 1.64 0.68 72.22-17.93 0.87 47.99 Jue 2014-4.40 0.69 83.10-12.70 0.82 64.16
Evaluatio of GPM-GSMaP Daily averaged raifall aroud Japa i 0.25 degree grid was compared with JMA s Radar AMeDAS (gauge-calibrated radar aalysis raifall). A example o Apr. 12, 2014 GSMaP_MVK GSMaP_Gauge Gauge-Radar Aalysis JMA radar coverage RMSE=0.35mm/h Corr.=0.825 RMSE=0.16mm/h Corr.=0.871 GSMaP_Gauge shows better correlatio with less Root Mea Square Error (RMSE) o Apr. 12, 2014.
Evaluatios by RMSE Better Root Mea Square Error (RMSE) (mm/hr) 0.6 0.5 0.4 0.3 0.2 0.1 0 Daily series of Root Mea Square Error (RMSE) for GSMaP product with referece to JMA radar-amedas data durig 1st April to 30th Jue 2014 TRMM-GSMaP_NRT GPM-GSMaP_MVK GPM-GSMaP_NRT GPM-GSMaP_Gauge GPM-GSMaP_Gauge GPM-GSMaP_MVK GPM-GSMaP_NRT TRMM-GSMaP_NRT Mea RMSE (Apr.-Ju. 2014): GPM-GSMaP_Gauge: 0.29 mm/h GPM-GSMaP_MVK: 0.33 mm/h GPM-GSMaP_NRT:0.34mm/hr TRMM-GSMaP_NRT: 0.35 mm/h 30-Ju-14 20-Ju-14 10-Ju-14 31-May-14 21-May-14 11-May-14 1-May-14 21-Apr-14 11-Apr-14 1-Apr-14 GSMaP_Gauge shows the least RMSE. GSMaP_Gauge algorithm is workig well.
However.. GSMaP_Gauge has latecies of a few days. The latecies of the GSMaP_Gauge sometimes cause problem for real time applicatios (ex. Flash flood warig system ). I received so may requests for real time versio of the GSMaP_Gauge product. 29
GSMaP_Gauge_NRT algorithm I the real time product, ufortuately, we do ot have real time gauge data. Istead, we prepare the data base which characterize the precipitatio i this model. Those are sigmaw, sigmav ad c. GSMaP Gauge x:true precipitatio rate y: Observatio I order to have the model parameters (N, C) i GSMaP Gauge NRT, x:gsmap Gauge precipitatio rate y: GSMaP MVK precipitatio rate
How to get model parameters GSMaP MVK GSMaP Gauge 1 2 3.... 29 30 31 32 day 30 days c, σ v, (σ w, μ w ) GSMaP Gauge NRT model GSMaP Gauge NRT
Distributio of Parameters of GSMaP Gauge NRT c σ v μ w GSMaP Gauge c : 0.7 σ v :1.0 μ w :0 σ w :1.5 σ w
Zoal Mea much improved. 60-60
GSMaP Gauge NRT Mothly raifall distributio, Jue 2015
Daily precipitatio amout, Jauary 2015 Vs GSMaP Gauge GSMaP MVK GSMaP Gauge NRT GSMaP Gauge NRT Correlatio RMSE Coefficiet of regressio lie GSMaP MVK 0.844 40.3 0.894 Gauge NRT (Costat) 0.853 62.3 1.26 Gauge NRT (c, σ v ) 0.873 34.0 0.873
Curret GSMaP products GSMaP_MWR Microwave radiometer product GSMaP_MVK Global precipitatio mappig from microwave ad ifrared radiometric data GSMaP_Gauge Gauge adjusted GSMaP_MVK GSMaP_NRT ad GSMaP_Now Near real time versio of the GSMaP_MVK product GSMaP_Gauge_NRT Near real time product of GSMaP_Gauge 36
Summary ad Coclusios Methodology ad algorithm structure of the high resolutio GSMaP was described.
System Model of the GSMaP_Gauge product a x 24 = 1 + 1 a = αa a = = W + + w v a: Rai rate at each pixel W: Rai rate from CPC global gauge data sets : Time, a: Rai rate from GSMaP_Gauge x: Rai rate from GSMaP_MVK, v, w: Noise 2 2 w N( µ w, σ ) v (, ) w N µ v σ v
System Model of the GSMaP_Gauge product a x + 1 = a = αa + + w v a: Rai rate at each pixel W: Rai rate from CPC global gauge data sets : Time, a: Rai rate from GSMaP_Gauge x: Rai rate from GSMaP_MVK, v, w: Noise 2 2 w ( w, ) N µ σ v (, ) w N µ v σ v L( a) 24 24 = l Pr( x, a) = l Pr( a ) Pr( a m am 1) Pr( x m= 1 = 1 1 ) a
System Model of the GSMaP_Gauge product a x 24 = 1 + 1 = a = αa a = W + w + v a: Rai rate at each pixel W: Rai rate from CPC global gauge data sets : Time, a: Rai rate from GSMaP_Gauge x: Rai rate from GSMaP_MVK, v, w: Noise 2 2 w ( w, ) N µ σ v (, ) w N µ v σ v
= l Pr( a Cost Fuctio L( a) = l Pr( x, a) e 1 ) 24 m= 1 Pr( a m λ N a 2 a = 1 ) e I a word, based o the assumptio that the GSMaP_Gauge CPC Gauge data (Gauge term) has the gaussia distributio, we maximize the probability desity fuctio of the GSMaP_Gauge estimatio multiplied by the Gauge term. The solutio ca be determied by calculatig the dl/da = 0 equatio m 1 ) W 24 = 1 2 Pr( x a λ N a 2 = 1 Gauge term 2 W
From JAXA Examples GSMaP_NRT (FW) Mothly Mea i April 2014(V03A) GPCC Gauge Mothly Moitorig GSMaP_MVK NOAA/CPC Daily Gauge Aalysis GSMaP_Gauge
Zoal mea aalysis Compariso betwee MVK(red) ad GSMaP_Gauge(black) 2004/01 2004/04 2004/07 2004/10 Estimatio from the GSMaP_Gauge product teds to be larger tha that from the GSMaP_MVK
GSMaP_Gauge Radar Rai Gauge Network GSMaP_NRT 44
Correlatio 45
RMS Error
モデルへの期待 大気移動ベクトルが, 過去から将来まで多層にわたって求まる GSMaP_MVK, GSMaP_Now, GSMaP_Future まで使用することが可能となる より効果的な高度のベクトルの使用が可能となる モデルによる地表面降水強度の使用 高緯度では観測より精度が良い? 一方, 熱帯では??? MVK でのカルマンフィルタの観測方程式の y にモデルからの地表面降水量を追加することが可能 47
モデルとの複合案 モデルからの移動ベクトルを使用 多層で使えるか GSMaP_MVK GSMaP_Now にも使用できる さらに,GSMaP_Future も作れそう MVK でのカルマンフィルタ 観測方程式の y にモデルからの地表面降水量を追加 48
State ad observatio equatio used i x k +1 = x k + σw Kalma filter ( State Equatio) 350 x k y k x k +1 v :Rai rate at time k :Ifrared Tb, R from model :Rai rate at time k+1 :System oise :Observatio oise w 150 IR brightess temperature [K 300 250 200 0 10 20 30 Precipitatio [mm/h]
CPC vs GSMaP_MVK ad GSMaP_Gauge
State ad observatio equatio used i x k +1 = x k + σw Kalma filter ( State Equatio) 350 x k y k x k +1 w v :Rai rate at time k :Ifrared Tb :Rai rate at time k+1 :System oise :Observatio oise IR brightess temperature [K 300 250 200 150 0 10 20 30 Precipitatio [mm/h]