Travel & Places

A linear merging methodology for high-resolution precipitation products using spatiotemporal regression

A linear merging methodology for high-resolution precipitation products using spatiotemporal regression
of 26
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
  This article was downloaded by: [Anish Turlapaty]On: 12 July 2012, At: 06:06Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK International Journal of RemoteSensing Publication details, including instructions for authors andsubscription information: A linear merging methodology for high-resolution precipitation products usingspatiotemporal regression Anish C. Turlapaty a   b  , Nicolas H. Younan a   b  & Valentine G.Anantharaj ca  Department of Electrical and Computer Engineering, MississippiState University, Mississippi State, MS, 39762, USA b  Geosystems Research Institute, Mississippi State University,Mississippi State, MS, 39762, USA c  The Scientific Computing Group, National Center forComputational Sciences, Oak Ridge National Laboratory,Starkville, MS, USAVersion of record first published: 12 Jul 2012 To cite this article:  Anish C. Turlapaty, Nicolas H. Younan & Valentine G. Anantharaj (2012):A linear merging methodology for high-resolution precipitation products using spatiotemporalregression, International Journal of Remote Sensing, 33:24, 7844-7867 To link to this article: PLEASE SCROLL DOWN FOR ARTICLEFull terms and conditions of use: conditionsThis article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden.The publisher does not give any warranty express or implied or make any representationthat the contents will be complete or accurate or up to date. The accuracy of anyinstructions, formulae, and drug doses should be independently verified with primarysources. The publisher shall not be liable for any loss, actions, claims, proceedings,  demand, or costs or damages whatsoever or howsoever caused arising directly orindirectly in connection with or arising out of the use of this material.    D  o  w  n   l  o  a   d  e   d   b  y   [   A  n   i  s   h   T  u  r   l  a  p  a   t  y   ]  a   t   0   6  :   0   6   1   2   J  u   l  y   2   0   1   2  International Journal of Remote Sensing  Vol. 33, No. 24, 20 December 2012, 7844–7867 A linear merging methodology for high-resolution precipitation productsusing spatiotemporal regression ANISH C. TURLAPATY*†‡, NICOLAS H. YOUNAN†‡ andVALENTINE G. ANANTHARAJ§ †Department of Electrical and Computer Engineering, Mississippi State University,Mississippi State, MS 39762, USA‡Geosystems Research Institute, Mississippi State University, Mississippi State, MS39762, USA§The Scientific Computing Group, National Center for Computational Sciences, OakRidge National Laboratory, Starkville, MS, USA ( Received 25 June 2011; in final form 26 January 2012 )Currently, the only viable option for a global precipitation product is the merger of several precipitation products from different modalities. In this article, we developa linear merging methodology based on spatiotemporal regression. Four highres-olution precipitation products (HRPPs), obtained through methods including theClimate Prediction Center’s Morphing (CMORPH), Geostationary OperationalEnvironmental Satellite-Based Auto-Estimator (GOES-AE), GOES-BasedHydro-Estimator (GOES-HE) and Self-Calibrating Multivariate PrecipitationRetrieval (SCAMPR) algorithms, are used in this study. The merged data areevaluated against the Arkansas Red Basin River Forecast Center’s (ABRFC’s)ground-based rainfall product. The evaluation is performed using the Heidkeskill score (HSS) for four seasons, from summer 2007 to spring 2008, and for twodifferent rainfall detection thresholds. It is shown that the merged data outperformall the other products in seven out of eight cases. A key innovation of this machinelearning method is that only 6% of the validation data are used for the initialtraining. The sensitivity of the algorithm to location, distribution of training data,selection of input data sets and seasons is also analysed and presented. 1. Introduction Even though rainfall is one of the key processes in global water cycle studies andweather monitoring, there is no single precipitation product with sufficiently high spa-tial resolution, global coverage, temporal sampling and reliable accuracy. Achievingsuch a performance from a single sensor modality has proved to be a difficult chal-lenge. To improve current standards in precipitation observation from space, theNational Aeronautics and Space Administration (NASA) in partnership with theJapanese Space Agency (JAXA) is considering the use of a constellation of globalprecipitation measurement (GPM) satellites. Some of the principal objectives of thismission are to improve our understanding of rainfall dynamics, provide better input toweather and hydrological prediction models and archive data for long-term research *Corresponding author. Email: International Journal of Remote Sensing  ISSN 0143-1161 print/ISSN 1366-5901 online © 2012 Taylor & Francishttp://www.tandfonline.com    D  o  w  n   l  o  a   d  e   d   b  y   [   A  n   i  s   h   T  u  r   l  a  p  a   t  y   ]  a   t   0   6  :   0   6   1   2   J  u   l  y   2   0   1   2  Optimal merging of rainfall data from Arkansas Basin  7845(Bundas 2006, Hou 2006, 2008, Hou  et al  . 2010). A fused data set, which possessesthe useful characteristics of several individual measurement modalities, can also servethe purpose of a global precipitation product. For instance, a merger of satellite-basedrainfall data sets derived from different modalities, such as passive microwave (PMW)and infrared (IR) sensors, can be suitable for the development of global long-term pre-cipitation products for academic research and short-term meteorological studies. Themerging methods developed in the last five years address the combination of rainfalldata from a variety of sources, such as field observations and climate models, ground-based gauges and satellite-based sensors and PMW sensors and ground-based radarmeasurements. These merged products usually combine the advantages from individ-ual sources and eliminate some of the disadvantages. A review of the most recentmerging methods follows.Gupta  et al  . (2006) developed a Kalman filter-based scale recursive estimation(SRE) method for merging precipitation estimates from multiple sources. This methoddoes not require the multiplicative cascade models traditionally used in SRE methods;instead, it uses a data-driven expectation maximization approach for model estima-tion. A possible application of the method to the GPM mission was demonstrated bymerging of the synthetic examples with spatial and temporal resolutions similar to theGPM mission specifications. Luo  et al  . (2007) developed a Bayesian framework formerging multiple climate models based on hydrological forecasts on a seasonal basis.In this framework, a historical data set is used to construct the prior distribution. Themodel hindcasts are regressed against historical data to model the likelihood function.The conditional posterior probability of the predicted precipitation was computedusing the Bayes rule. The forecasts from multiple models were used as an ensembleto construct the posterior probability distribution. Over the Ohio River Basin, themerged predictions from these ensemble models proved to be much closer to the actualobservations than the model predictions or the climatology-based predictions alone(Luo  et al  . 2007).Chiang  et al  . (2007) developed a linear merging methodology to improve flash floodforecasting in Taiwan using both satellite-based rainfall estimates and ground-basedrain gauge data. The goal of this methodology is to determine the merging parame-ters to be multiplied to the gauge data in the linear model while the satellite estimateis weighted by its complementary value. The process consists of two cases: one caseis for the biased data and the other one is for the unbiased data. These parametersare determined using a grid search with a recurrent neural network that models thehydrological processes. The final data set produced was validated against heavy rain-fall events during Taiwan’s typhoon season. A linear merging method was developedby Sapiano  et al  . (2008), with the goal of producing a global precipitation prod-uct. The weights and errors were estimated by comparing the input data sets withthe available gauge data. The product was validated against other available data setswith very good performance in high latitudes and some biases in tropical regions(Sapiano  et al  . 2008). A recursive algorithm for merging rainfall data was developedby van de Vyver and Roulin (2009). This method is a variation of the Gupta  et al  .(2006) method, with experiments being done on real multi-sensor data sets. It com-bines rainfall estimates from ground-based radar with those from a satellite-basedmicrowave sensor on board the National Oceanic and Atmospheric Administration(NOAA) satellite over Belgium. An artificial neural network (ANN)-based precipita-tionestimationfromdata obtained through visibleand infrared sensors was developedby Behrangi  et al  . (2009). Self-organizing feature maps (SOFM) were used for the    D  o  w  n   l  o  a   d  e   d   b  y   [   A  n   i  s   h   T  u  r   l  a  p  a   t  y   ]  a   t   0   6  :   0   6   1   2   J  u   l  y   2   0   1   2  7846  A. C. Turlapaty  et al . classification process. A few more recent methods for combining the rainfall data fromsuch complementary sources are a shuffled complex evolutionary search approach(Hsu  et al.  2007), a history-based merging method for simulation modelling (Frezghiand Smithers 2008) and a kernel smoothing algorithm (Li and Shao 2010).One of the objectives of the feasibility study conducted at Mississippi StateUniversity was to develop and evaluate a merging method for high-resolution pre-cipitation products (HRPPs). Some of these HRPPs are usually merged data setsthemselves; for example, one data set is a combination of PMW data with IR-basedestimates. In the context of our study, an important characteristic of these HRPPs isthat they have better skill with heavy rainfall (Ebert  et al  . 2007, Turk  et al  . 2008). Themain goal of our study is to develop a merging methodology that fuses HRPPs intoa statistically superior product. Turlapaty  et al  . (2010) developed a merging methodbased on a non-linear vector transformation and ANNs. The goal of the methodol-ogy was to improve the rain detection rate and reduce the false alarm rates (FARs).However,animportantaspectofthispatternrecognitionmethodwasthatitrequiredalargeamountoftrainingdata.Inthatstudy,theHRPPsforfourseasons,fromsummer2007 to spring 2008, were considered, and in order to achieve a satisfactory perfor-mance, the method needed to be trained with at least one season of the validation datato fuse the remaining three seasons.In this article, we propose a new spatiotemporal regression-based approach thatfuses HRPPs into a statistically better product. Taking into consideration variationswith respect to seasons, geographical region, amount of rainfall and type of on-boardsensors, this merging method outperforms each of the individual HRPPs and requiresonly a fraction of the validation data, currently as low as 6%, to train the mergingmethod. Another important difference in this approach is that it is a linear mergingmethod and the merging parameters are adaptable to spatial locations. 2. Methodology2.1  Motivation As described in Chiang  et al  . (2007), a linear merging method was developed for datasets from different types of sources such as satellites and gauges. In this article, wepropose a linear merging method for the data sets derived from very similar sensorsor algorithms. For instance, two of the HRPPs used in this study are derived from thesame sensor on board the Geostationary Operational Environmental Satellite (GOES)using two related algorithms, the auto-estimator (Vicente  et al  . 1998) and the hydro-estimator (Scofield and Kuligowski 2003, Ramirez-Beltran  et al  . 2008). These methodsare discussed in §3. The main idea behind the proposed methodology is that the indi-vidual HRPPs perform differently under different conditions, such as seasons – forinstance, the Climate Prediction Center’s Morphing algorithm (CMORPH) performsbetter in warm seasons (Turk  et al  . 2008); geographical region – satellite-based algo-rithms perform well in the central USA (Ebert  et al  . 2007); the rainfall amount – asmentioned in §1, the HRPP performance depends on the amount of rainfall; and thetype of on-board sensor – IR or microwave. For example, an ANN-based classifier wasdeveloped by Capacci and Porcu (2008) and used to compare the skills of the spinningenhanced visible and infrared imager (SEVIRI) sensor on board the MeteorologicalSatellite, METEOSAT-8, with the Advanced Microwave Scanning Radiometer EarthObserving System (AMSR-E) on board the Aqua satellite. The key findings are thatSEVIRI had better skill in rain area detection, AMSR-E had better skill in rain-rate    D  o  w  n   l  o  a   d  e   d   b  y   [   A  n   i  s   h   T  u  r   l  a  p  a   t  y   ]  a   t   0   6  :   0   6   1   2   J  u   l  y   2   0   1   2
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks