Browse
Abstract
The retrieval of the mass-weighted mean diameter (D m ) is a fundamental component of spaceborne precipitation retrievals. The Dual-Frequency Precipitation Radar (DPR) on the Global Precipitation Measurement (GPM) satellite is the first satellite to use dual-wavelength ratio measurements—the quotient of radar reflectivity factors (Z) measured at Ku and Ka wavelengths—to retrieve D m . While it is established that DWR, being theoretically insensitive to changes in ice crystal mass and concentration, can provide a superior retrieval of D m compared to Z-based retrievals, the benefits of this retrieval have yet to be directly observed or quantified. In this study, DWR–D m and Z–D m relationships are empirically generated from collocated airborne radar and in situ cloud particle probe measurements. Data are collected during nine intensive observation periods (IOPs) from three experiments representing different locations and times of year. Across IOPs with varying ice crystal concentrations, cloud temperatures, and storm types, Z–D m relationships vary considerably while the DWR–D m relationship remains consistent. This study confirms that a DWR–D m relationship can provide a more accurate and consistent D m retrieval than a Z–D m relationship, quantified by a reduced overall RMSE (0.19 and 0.25 mm, respectively) and a reduced range of biases between experiments (0.11 and 0.32 mm, respectively).
Abstract
The retrieval of the mass-weighted mean diameter (D m ) is a fundamental component of spaceborne precipitation retrievals. The Dual-Frequency Precipitation Radar (DPR) on the Global Precipitation Measurement (GPM) satellite is the first satellite to use dual-wavelength ratio measurements—the quotient of radar reflectivity factors (Z) measured at Ku and Ka wavelengths—to retrieve D m . While it is established that DWR, being theoretically insensitive to changes in ice crystal mass and concentration, can provide a superior retrieval of D m compared to Z-based retrievals, the benefits of this retrieval have yet to be directly observed or quantified. In this study, DWR–D m and Z–D m relationships are empirically generated from collocated airborne radar and in situ cloud particle probe measurements. Data are collected during nine intensive observation periods (IOPs) from three experiments representing different locations and times of year. Across IOPs with varying ice crystal concentrations, cloud temperatures, and storm types, Z–D m relationships vary considerably while the DWR–D m relationship remains consistent. This study confirms that a DWR–D m relationship can provide a more accurate and consistent D m retrieval than a Z–D m relationship, quantified by a reduced overall RMSE (0.19 and 0.25 mm, respectively) and a reduced range of biases between experiments (0.11 and 0.32 mm, respectively).
Abstract
A key strategy in obtaining complete global coverage of high-resolution precipitation is to combine observations from multiple fields, such as the intermittent passive microwave observations, precipitation propagated in time using motion vectors, and geosynchronous infrared observations. These separate precipitation fields can be combined through weighted averaging, which produces estimates that are generally superior to the individual parent fields. However, the process of averaging changes the distribution of the precipitation values, leading to an increase in precipitating area and a decrease in the values of high precipitation rates, a phenomenon observed in IMERG. To mitigate this issue, we introduce a new scheme called SHARPEN (Scheme for Histogram Adjustment with Ranked Precipitation Estimates in the Neighborhood), which recovers the distribution of the averaged precipitation field based on the idea of quantile mapping applied to the local environment. When implemented in IMERG, precipitation estimates from SHARPEN exhibit a distribution that resembles that of the original instantaneous observations, with matching precipitating area and peak precipitation rates. Case studies demonstrate its improved ability in bridging between the parent precipitation fields. Evaluation against ground observations reveals a distinct improvement in precipitation detection skill, but also a slightly reduced correlation likely because of a sharper precipitation field. The increased computational demand of SHARPEN can be mitigated by striding over multiple grid boxes, which has only marginal impacts on the accuracy of the estimates. SHARPEN can be applied to any precipitation algorithm that produces an average from multiple input precipitation fields and is being considered for implementation in IMERG V07.
Abstract
A key strategy in obtaining complete global coverage of high-resolution precipitation is to combine observations from multiple fields, such as the intermittent passive microwave observations, precipitation propagated in time using motion vectors, and geosynchronous infrared observations. These separate precipitation fields can be combined through weighted averaging, which produces estimates that are generally superior to the individual parent fields. However, the process of averaging changes the distribution of the precipitation values, leading to an increase in precipitating area and a decrease in the values of high precipitation rates, a phenomenon observed in IMERG. To mitigate this issue, we introduce a new scheme called SHARPEN (Scheme for Histogram Adjustment with Ranked Precipitation Estimates in the Neighborhood), which recovers the distribution of the averaged precipitation field based on the idea of quantile mapping applied to the local environment. When implemented in IMERG, precipitation estimates from SHARPEN exhibit a distribution that resembles that of the original instantaneous observations, with matching precipitating area and peak precipitation rates. Case studies demonstrate its improved ability in bridging between the parent precipitation fields. Evaluation against ground observations reveals a distinct improvement in precipitation detection skill, but also a slightly reduced correlation likely because of a sharper precipitation field. The increased computational demand of SHARPEN can be mitigated by striding over multiple grid boxes, which has only marginal impacts on the accuracy of the estimates. SHARPEN can be applied to any precipitation algorithm that produces an average from multiple input precipitation fields and is being considered for implementation in IMERG V07.
Abstract
Applications like drought monitoring and forecasting can profit from the global and near-real-time availability of satellite-based precipitation estimates once their related uncertainties and challenges are identified and treated. To this end, this study evaluates the IMERG V06B Late Run precipitation product from the Global Precipitation Measurement mission (GPM), a multisatellite product that combines space-based radar, passive microwave (PMW), and infrared (IR) data into gridded precipitation estimates. The evaluation is performed on the spatiotemporal resolution of IMERG (0.1° × 0.1°, 30 min) over the Netherlands over a 5-yr period. A gauge-adjusted radar precipitation product from the Royal Netherlands Meteorological Institute (KNMI) is used as reference, against which IMERG shows a large positive bias. To find the origin of this systematic overestimation, the data are divided into seasons, rainfall intensity ranges, echo top height (ETH) ranges, and categories based on the relative contributions of IR, morphing, and PMW data to the IMERG estimates. Furthermore, the specific radiometer is identified for each PMW-based estimate. IMERG’s detection performance improves with higher ETH and rainfall intensity, but the associated error and relative bias increase as well. Severe overestimation occurs during low-intensity rainfall events and is especially linked to PMW observations. All individual PMW instruments show the same pattern: overestimation of low-intensity events and underestimation of high-intensity events. IMERG misses a large fraction of shallow rainfall events, which is amplified when IR data are included. Space-based retrieval of shallow and low-intensity precipitation events should improve before IMERG can become accurate over the middle and high latitudes.
Abstract
Applications like drought monitoring and forecasting can profit from the global and near-real-time availability of satellite-based precipitation estimates once their related uncertainties and challenges are identified and treated. To this end, this study evaluates the IMERG V06B Late Run precipitation product from the Global Precipitation Measurement mission (GPM), a multisatellite product that combines space-based radar, passive microwave (PMW), and infrared (IR) data into gridded precipitation estimates. The evaluation is performed on the spatiotemporal resolution of IMERG (0.1° × 0.1°, 30 min) over the Netherlands over a 5-yr period. A gauge-adjusted radar precipitation product from the Royal Netherlands Meteorological Institute (KNMI) is used as reference, against which IMERG shows a large positive bias. To find the origin of this systematic overestimation, the data are divided into seasons, rainfall intensity ranges, echo top height (ETH) ranges, and categories based on the relative contributions of IR, morphing, and PMW data to the IMERG estimates. Furthermore, the specific radiometer is identified for each PMW-based estimate. IMERG’s detection performance improves with higher ETH and rainfall intensity, but the associated error and relative bias increase as well. Severe overestimation occurs during low-intensity rainfall events and is especially linked to PMW observations. All individual PMW instruments show the same pattern: overestimation of low-intensity events and underestimation of high-intensity events. IMERG misses a large fraction of shallow rainfall events, which is amplified when IR data are included. Space-based retrieval of shallow and low-intensity precipitation events should improve before IMERG can become accurate over the middle and high latitudes.
Abstract
This paper describes a new Passive Microwave Empirical Cold Surface Classification Algorithm (PESCA) developed for snow-cover detection and characterization by using passive microwave satellite measurements. The main goal of PESCA is to support the retrieval of falling snow, since several studies have highlighted the influence of snow-cover radiative properties on the falling-snow passive microwave signature. The developed method is based on the exploitation of the lower-frequency channels (<90 GHz), common to most microwave radiometers. The method applied to the conically scanning Global Precipitation Measurement (GPM) Microwave Imager (GMI) and the cross-track-scanning Advanced Technology Microwave Sounder (ATMS) is described in this paper. PESCA is based on a decision tree developed using an empirical method and verified using the AutoSnow product built from satellite measurements. The algorithm performance appears to be robust both for sensors in dry conditions (total precipitable water < 10 mm) and for mean surface elevation < 2500 m, independent of the cloud cover. The algorithm shows very good performance for cold temperatures (2-m temperature below 270 K) with a rapid decrease of the detection capabilities between 270 and 280 K, where 280 K is assumed as the maximum temperature limit for PESCA (overall detection statistics: probability of detection is 0.98 for ATMS and 0.92 for GMI, false alarm ratio is 0.01 for ATMS and 0.08 for GMI, and Heidke skill score is 0.72 for ATMS and 0.69 for GMI). Some inconsistencies found between the snow categories identified with the two radiometers are related to their different viewing geometries, spatial resolution, and temporal sampling. The spectral signatures of the different snow classes also appear to be different at high frequency (>90 GHz), indicating potential impact for snowfall retrieval. This method can be applied to other conically scanning and cross-track-scanning radiometers, including the future operational EUMETSAT Polar System Second Generation (EPS-SG) mission microwave radiometers.
Abstract
This paper describes a new Passive Microwave Empirical Cold Surface Classification Algorithm (PESCA) developed for snow-cover detection and characterization by using passive microwave satellite measurements. The main goal of PESCA is to support the retrieval of falling snow, since several studies have highlighted the influence of snow-cover radiative properties on the falling-snow passive microwave signature. The developed method is based on the exploitation of the lower-frequency channels (<90 GHz), common to most microwave radiometers. The method applied to the conically scanning Global Precipitation Measurement (GPM) Microwave Imager (GMI) and the cross-track-scanning Advanced Technology Microwave Sounder (ATMS) is described in this paper. PESCA is based on a decision tree developed using an empirical method and verified using the AutoSnow product built from satellite measurements. The algorithm performance appears to be robust both for sensors in dry conditions (total precipitable water < 10 mm) and for mean surface elevation < 2500 m, independent of the cloud cover. The algorithm shows very good performance for cold temperatures (2-m temperature below 270 K) with a rapid decrease of the detection capabilities between 270 and 280 K, where 280 K is assumed as the maximum temperature limit for PESCA (overall detection statistics: probability of detection is 0.98 for ATMS and 0.92 for GMI, false alarm ratio is 0.01 for ATMS and 0.08 for GMI, and Heidke skill score is 0.72 for ATMS and 0.69 for GMI). Some inconsistencies found between the snow categories identified with the two radiometers are related to their different viewing geometries, spatial resolution, and temporal sampling. The spectral signatures of the different snow classes also appear to be different at high frequency (>90 GHz), indicating potential impact for snowfall retrieval. This method can be applied to other conically scanning and cross-track-scanning radiometers, including the future operational EUMETSAT Polar System Second Generation (EPS-SG) mission microwave radiometers.
Abstract
Surface snowfall rate estimates from the Global Precipitation Measurement (GPM) mission’s Core Observatory sensors and the CloudSat radar are compared to those from the Multi-Radar Multi-Sensor (MRMS) radar composite product over the continental United States during the period from November 2014 to September 2020. The analysis includes the Dual-Frequency Precipitation Radar (DPR) retrieval and its single-frequency counterparts, the GPM Combined Radar Radiometer Algorithm (CORRA), the CloudSat Snow Profile product (2C-SNOW-PROFILE), and two passive microwave retrievals, i.e., the Goddard Profiling algorithm (GPROF) and the Snow Retrieval Algorithm for GMI (SLALOM). The 2C-SNOW retrieval has the highest Heidke skill score (HSS) for detecting snowfall among the products analyzed. SLALOM ranks second; it outperforms GPROF and the other GPM algorithms, all detecting only 30% of the snow events. Since SLALOM is trained with 2C-SNOW, it suggests that the optimal use of the information content in the GMI observations critically depends on the precipitation training dataset. All the retrievals underestimate snowfall rates by a factor of 2 compared to MRMS. Large discrepancies (RMSE of 0.7–1.5 mm h−1) between spaceborne and ground-based snowfall rate estimates are attributed to the complexity of the ice scattering properties and to the limitations of the remote sensing systems: the DPR instrument has low sensitivity, while the radiometric measurements are affected by the confounding effects of the background surface emissivity and of the emission of supercooled liquid droplet layers.
Abstract
Surface snowfall rate estimates from the Global Precipitation Measurement (GPM) mission’s Core Observatory sensors and the CloudSat radar are compared to those from the Multi-Radar Multi-Sensor (MRMS) radar composite product over the continental United States during the period from November 2014 to September 2020. The analysis includes the Dual-Frequency Precipitation Radar (DPR) retrieval and its single-frequency counterparts, the GPM Combined Radar Radiometer Algorithm (CORRA), the CloudSat Snow Profile product (2C-SNOW-PROFILE), and two passive microwave retrievals, i.e., the Goddard Profiling algorithm (GPROF) and the Snow Retrieval Algorithm for GMI (SLALOM). The 2C-SNOW retrieval has the highest Heidke skill score (HSS) for detecting snowfall among the products analyzed. SLALOM ranks second; it outperforms GPROF and the other GPM algorithms, all detecting only 30% of the snow events. Since SLALOM is trained with 2C-SNOW, it suggests that the optimal use of the information content in the GMI observations critically depends on the precipitation training dataset. All the retrievals underestimate snowfall rates by a factor of 2 compared to MRMS. Large discrepancies (RMSE of 0.7–1.5 mm h−1) between spaceborne and ground-based snowfall rate estimates are attributed to the complexity of the ice scattering properties and to the limitations of the remote sensing systems: the DPR instrument has low sensitivity, while the radiometric measurements are affected by the confounding effects of the background surface emissivity and of the emission of supercooled liquid droplet layers.
Abstract
Rainfall retrieval algorithms for passive microwave radiometers often exploit the brightness temperature depression due to ice scattering at high-frequency channels (≥85 GHz) over land. This study presents an alternate method to estimate the daily rainfall amount using the emissivity temporal variation (i.e., Δe) under rain-free conditions at low-frequency channels (19, 24, and 37 GHz). Emissivity is derived from 10 passive microwave radiometers, including the Global Precipitation Measurement (GPM) Microwave Imager (GMI), the Advanced Microwave Scanning Radiometer 2 (AMSR2), three Special Sensor Microwave Imager/Sounders (SSMIS), the Advanced Technology Microwave Sounder (ATMS), and four Advanced Microwave Sounding Units-A (AMSU-A). Four different satellite combination schemes are used to derive the Δe for daily rainfall estimates. They are all 10 satellites, 5 imagers, 6 satellites with very different equator crossing times, and GMI only. Results show that Δe from all 10 satellites has the best performance with a correlation of 0.60 and RMSE of 6.52 mm, compared with the Integrated Multisatellite Retrievals for GPM (IMERG) Final run product. The 6-satellites scheme has comparable performance with the all-10-satellites scheme. The 5-imagers scheme performs noticeably worse with a correlation of 0.49 and RMSE of 7.28 mm, while the GMI-only scheme performs the worst with a correlation of 0.25 and RMSE of 11.36 mm. The inferior performance from the 5-imagers and GMI-only schemes can be explained by the much longer revisit time, which cannot accurately capture the emissivity temporal variation.
Abstract
Rainfall retrieval algorithms for passive microwave radiometers often exploit the brightness temperature depression due to ice scattering at high-frequency channels (≥85 GHz) over land. This study presents an alternate method to estimate the daily rainfall amount using the emissivity temporal variation (i.e., Δe) under rain-free conditions at low-frequency channels (19, 24, and 37 GHz). Emissivity is derived from 10 passive microwave radiometers, including the Global Precipitation Measurement (GPM) Microwave Imager (GMI), the Advanced Microwave Scanning Radiometer 2 (AMSR2), three Special Sensor Microwave Imager/Sounders (SSMIS), the Advanced Technology Microwave Sounder (ATMS), and four Advanced Microwave Sounding Units-A (AMSU-A). Four different satellite combination schemes are used to derive the Δe for daily rainfall estimates. They are all 10 satellites, 5 imagers, 6 satellites with very different equator crossing times, and GMI only. Results show that Δe from all 10 satellites has the best performance with a correlation of 0.60 and RMSE of 6.52 mm, compared with the Integrated Multisatellite Retrievals for GPM (IMERG) Final run product. The 6-satellites scheme has comparable performance with the all-10-satellites scheme. The 5-imagers scheme performs noticeably worse with a correlation of 0.49 and RMSE of 7.28 mm, while the GMI-only scheme performs the worst with a correlation of 0.25 and RMSE of 11.36 mm. The inferior performance from the 5-imagers and GMI-only schemes can be explained by the much longer revisit time, which cannot accurately capture the emissivity temporal variation.
Abstract
With the launch of the Global Precipitation Measurement Dual-Frequency Precipitation Radar (GPM-DPR) in 2014, renewed interest in retrievals of snowfall in the atmospheric column has occurred. The current operational GPM-DPR retrieval largely underestimates surface snowfall accumulation. Here, a neural network (NN) trained on data that are synthetically derived from state-of-the-art ice particle scattering models and measured in situ particle size distributions (PSDs) is used to retrieve two parameters of the PSD: liquid equivalent mass-weighted mean diameter
Abstract
With the launch of the Global Precipitation Measurement Dual-Frequency Precipitation Radar (GPM-DPR) in 2014, renewed interest in retrievals of snowfall in the atmospheric column has occurred. The current operational GPM-DPR retrieval largely underestimates surface snowfall accumulation. Here, a neural network (NN) trained on data that are synthetically derived from state-of-the-art ice particle scattering models and measured in situ particle size distributions (PSDs) is used to retrieve two parameters of the PSD: liquid equivalent mass-weighted mean diameter
Abstract
Two power-law relations linking equivalent radar reflectivity factor Z e and snowfall rate S are derived for a K-band Micro Rain Radar (MRR) and for a W-band cloud radar. For the development of these Z e –S relationships, a dataset of calculated and measured variables is used. Surface-based video-disdrometer measurements were collected during snowfall events over five winters at the high-latitude site in Hyytiälä, Finland. The data from 2014 to 2018 include particle size distributions (PSD) and their fall velocities, from which snowflake masses were derived. The K- and W-band Z e values are computed using these surface-based observations and snowflake scattering properties as provided by T-matrix and single-particle scattering tables, respectively. The uncertainty analysis shows that the K-band snowfall-rate estimation is significantly improved by including the intercept parameter N 0 of the PSD calculated from concurrent disdrometer measurements. If N 0 is used to adjust the prefactor of the Z e –S relationship, the RMSE of the snowfall-rate estimate can be reduced from 0.37 to around 0.11 mm h−1. For W-band radar, a Z e –S relationship with constant parameters for all available snow events shows a similar uncertainty when compared with the method that includes the PSD intercept parameter. To demonstrate the performance of the proposed Z e –S relationships, they are applied to measurements of the MRR and the W-band microwave radar for Arctic clouds at the Arctic research base operated by the German Alfred Wegener Institute Helmholtz Centre for Polar and Marine Research (AWI) and the French Polar Institute Paul Emile Victor (IPEV) (AWIPEV) in Ny-Ålesund, Svalbard, Norway. The resulting snowfall-rate estimates show good agreement with in situ snowfall observations while other Z e –S relationships from literature reveal larger differences.
Abstract
Two power-law relations linking equivalent radar reflectivity factor Z e and snowfall rate S are derived for a K-band Micro Rain Radar (MRR) and for a W-band cloud radar. For the development of these Z e –S relationships, a dataset of calculated and measured variables is used. Surface-based video-disdrometer measurements were collected during snowfall events over five winters at the high-latitude site in Hyytiälä, Finland. The data from 2014 to 2018 include particle size distributions (PSD) and their fall velocities, from which snowflake masses were derived. The K- and W-band Z e values are computed using these surface-based observations and snowflake scattering properties as provided by T-matrix and single-particle scattering tables, respectively. The uncertainty analysis shows that the K-band snowfall-rate estimation is significantly improved by including the intercept parameter N 0 of the PSD calculated from concurrent disdrometer measurements. If N 0 is used to adjust the prefactor of the Z e –S relationship, the RMSE of the snowfall-rate estimate can be reduced from 0.37 to around 0.11 mm h−1. For W-band radar, a Z e –S relationship with constant parameters for all available snow events shows a similar uncertainty when compared with the method that includes the PSD intercept parameter. To demonstrate the performance of the proposed Z e –S relationships, they are applied to measurements of the MRR and the W-band microwave radar for Arctic clouds at the Arctic research base operated by the German Alfred Wegener Institute Helmholtz Centre for Polar and Marine Research (AWI) and the French Polar Institute Paul Emile Victor (IPEV) (AWIPEV) in Ny-Ålesund, Svalbard, Norway. The resulting snowfall-rate estimates show good agreement with in situ snowfall observations while other Z e –S relationships from literature reveal larger differences.
Abstract
High-resolution datasets offer the potential to improve our understanding of spatial and temporal precipitation patterns and storm structures. The goal of this study is to evaluate the similarities and differences of object-based storm characteristics as observed using space- or land-based sensors. The Method of Object-based Diagnostic Evaluation (MODE) Time Domain (MTD) is used to identify and track storm objects in two high-resolution merged datasets: the Integrated Multisatellite Retrievals for Global Precipitation Measurement (IMERG) final product V06B and gauge-corrected ground-radar-based Multi-Radar Multi-Sensor (MRMS) quantitative precipitation estimations. Characteristics associated with landfalling hurricanes were also examined as a separate category of storm. The results reveal that IMERG and MRMS agree reasonably well across many object-based storm characteristics. However, there are some discrepancies that are statistically significant. MRMS storms are more concentrated, with smaller areas and higher peak intensities, which implies higher flash flood risks associated with the storms. On the other hand, IMERG storms can travel longer distances with a higher volume of precipitation, which implies higher risk of riverine flooding. Agreement between the datasets is higher for faster-moving hurricanes in terms of the averaged intensity. Finally, MRMS indicates a higher average precipitation intensity during the hurricane’s lifetime. However, in non-hurricanes, the opposite result was observed. This is likely related to MRMS having higher resolution; monitoring the hurricanes from many viewing angles, leading to different signal saturation properties compared to IMERG; and/or the dominance of droplet aggregation effects over evaporation effects at lower altitudes.
Abstract
High-resolution datasets offer the potential to improve our understanding of spatial and temporal precipitation patterns and storm structures. The goal of this study is to evaluate the similarities and differences of object-based storm characteristics as observed using space- or land-based sensors. The Method of Object-based Diagnostic Evaluation (MODE) Time Domain (MTD) is used to identify and track storm objects in two high-resolution merged datasets: the Integrated Multisatellite Retrievals for Global Precipitation Measurement (IMERG) final product V06B and gauge-corrected ground-radar-based Multi-Radar Multi-Sensor (MRMS) quantitative precipitation estimations. Characteristics associated with landfalling hurricanes were also examined as a separate category of storm. The results reveal that IMERG and MRMS agree reasonably well across many object-based storm characteristics. However, there are some discrepancies that are statistically significant. MRMS storms are more concentrated, with smaller areas and higher peak intensities, which implies higher flash flood risks associated with the storms. On the other hand, IMERG storms can travel longer distances with a higher volume of precipitation, which implies higher risk of riverine flooding. Agreement between the datasets is higher for faster-moving hurricanes in terms of the averaged intensity. Finally, MRMS indicates a higher average precipitation intensity during the hurricane’s lifetime. However, in non-hurricanes, the opposite result was observed. This is likely related to MRMS having higher resolution; monitoring the hurricanes from many viewing angles, leading to different signal saturation properties compared to IMERG; and/or the dominance of droplet aggregation effects over evaporation effects at lower altitudes.
Abstract
This study evaluates a machine learning–based precipitation ensemble technique (MLPET) over three mountainous tropical regions. The technique, based on quantile regression forests, integrates global satellite precipitation datasets from CMORPH, PERSIANN, GSMaP (V6), and 3B42 (V7) and an atmospheric reanalysis precipitation product (EI_GPCC) with daily soil moisture, specific humidity, and terrain elevation datasets. The complex terrain study areas include the Peruvian and Colombian Andes in South America and the Blue Nile in East Africa. Evaluation is performed at a daily time scale and 0.25° spatial resolution based on 13 years (2000–12) of reference rainfall data derived from dense in situ rain gauge networks. The technique is evaluated using K-fold, separately in each region, and leave-one-region-out validation experiments. Comparison of MLPET with the individual satellite and reanalysis precipitation datasets used for the blending and the recent Multi-Source Weighted-Ensemble Precipitation (MSWEP) global precipitation product exhibited improved systematic and random error statistics for all regions. In addition, it is shown that observations are encapsulated well within the ensemble envelope generated by the blending technique.
Abstract
This study evaluates a machine learning–based precipitation ensemble technique (MLPET) over three mountainous tropical regions. The technique, based on quantile regression forests, integrates global satellite precipitation datasets from CMORPH, PERSIANN, GSMaP (V6), and 3B42 (V7) and an atmospheric reanalysis precipitation product (EI_GPCC) with daily soil moisture, specific humidity, and terrain elevation datasets. The complex terrain study areas include the Peruvian and Colombian Andes in South America and the Blue Nile in East Africa. Evaluation is performed at a daily time scale and 0.25° spatial resolution based on 13 years (2000–12) of reference rainfall data derived from dense in situ rain gauge networks. The technique is evaluated using K-fold, separately in each region, and leave-one-region-out validation experiments. Comparison of MLPET with the individual satellite and reanalysis precipitation datasets used for the blending and the recent Multi-Source Weighted-Ensemble Precipitation (MSWEP) global precipitation product exhibited improved systematic and random error statistics for all regions. In addition, it is shown that observations are encapsulated well within the ensemble envelope generated by the blending technique.
