Retrieval of snow parameters by iterative inversion of a neuralnetwork

The inversion of snow parameters from passive microwave remote sensing measurements is performed, using an iterative inversion of a neural network (NN) trained with a dense-media multiple-scattering model. Inversion of four parameters is performed based on five brightness temperatures. The four parameters are mean grain size of ice particles in snow, snow density, snow temperature, and snow depth. Iterative inversion of a data-driven forward NN model is justified on a theoretical and methodological basis. An error analysis is performed, comparing iterative inversion of a forward model with the use of an explicit inverse for the retrieval of independent snow parameters from their corresponding measurements. The NN iterative inversion algorithm is further illustrated by reconstructing a synthetic terrain of snow parameters from their corresponding measurements, inverting all four parameters simultaneously. The reconstructed parameter contours are in good agreement with the original synthetic parameter contours     

Brightness Temperatures of Snow Melting/Refreezing Cycles: Observations and Modeling Using a Multilayer Dense Medium Theory-Based Model

The ability of electromagnetic models to accurately predict microwave emission of a snowpack is complicated by the need to account for, among other things, nonindependent scattering by closely packed snow grains, stratigraphic variations, and the occurrence of wet snow. A multilayer dense medium model can account for the first two effects. While microwave remote sensing is well known to be capable of binary wet/dry discrimination, the ability to model brightness as a function of wetness opens up the possibility of ultimately retrieving a percentage wetness value during such hydrologically significant melting conditions. In this paper, the first application of a multilayer dense medium radiative transfer theory (DMRT) model is proposed to simulate emission from both wet and dry snow during melting and refreezing cycles. Wet snow is modeled as a mixture of ice particles surrounded by a thin film of water embedded in an air background. Melting/refreezing cycles are studied by means of brightness temperatures at 6.7, 19, and 37 GHz recorded by the University of Michigan Truck-Mounted Radiometer System at the Local Scale Observation Site during the Cold Land Processes Experiment-1 in March 2003. Input parameters to the DMRT model are obtained from snow pit measurements carried out in conjunction with the microwave observations. The comparisons between simulated and measured brightness temperatures show that the electromagnetic model is able to reproduce the brightness temperatures with an average percentage error of 3% (~8 K) and a maximum relative percentage error of around 8% (~20 K)     

Radiometric profiling of temperature using algorithm based on neural networks

The retrieval of atmospheric temperature profiles from microwave radiometer brightness temperatures requires the solution of a nonlinear inversion problem. An inversion technique based on neural networks (NN) is developed. The NN technique, compared with the classical inversion methods, exhibits better results in terms of retrieval accuracy, vertical resolution and elaboration time.     

Inversion of snow parameters from passive microwave remote sensingmeasurements by a neural network trained with a multiple scatteringmodel

The inversion of snow parameters from passive microwave remote sensing measurements is performed with a neural network trained with a dense-media multiple-scattering model. The input-output pairs generated by the scattering model are used to train the neural network. Simultaneous inversion of three parameters, mean-grain size of ice particles in snow, snow density, and snow temperature from five brightness temperatures, is reported. It is shown that the neural network gives good results for simulated data. The absolute percentage errors for mean-grain size of ice particles and snow density are less than 10%, and the absolute error for snow temperature is less than 3 K. The neural network with the trained weighting coefficients of the three-parameter model is also used to invert SSMI data taken over the Antarctic region     

Retrieval of dry-snow parameters from microwave radiometric data using a dense-medium model and genetic algorithms

A numerical technique based on genetic algorithms (GAs) is used to invert the equations of an electromagnetic model based on dense-medium radiative transfer theory (DMRT) to retrieve snow depth, mean grain size, and fractional volume from microwave brightness temperatures. In order to study the sensitivity of the GA to its parameters, the technique is initially tested on simulated microwave data with and without adding a random noise. A configuration of GA parameters is selected and used for the retrieval of snow parameters from both ground-based observations and brightness temperatures recorded by the Advanced Microwave Scanning Radiometer-EOS (AMSR-E). Retrieved snow parameters are then compared with those measured on ground. Although more investigation is required, results suggest that the proposed technique is able to retrieve snow parameters with good accuracy.     

Bayesian algorithm for microwave-based precipitation retrieval: description and application to TMI measurements over ocean

A physically oriented inversion algorithm to retrieve precipitation from satellite-based passive microwave measurements named the Bayesian algorithm for microwave-based precipitation retrieval (BAMPR) is proposed. First, we illustrate the procedure that BAMPR follows to produce precipitation estimates from observed multichannel brightness temperatures. Retrieval products are the surface rain rates, columnar equivalent water contents, and hydrometeor content profiles, together with the associated estimation uncertainties. Numerical tests performed on simulated measurements show that retrieval errors are reduced when a rain type and pattern classification procedure is employed, and that estimates are quite sensitive to the adopted error model. Finally, for different tropical storms that were observed by the Tropical Rainfall Measuring Mission (TRMM) Microwave Imager (TMI), we compare the rain retrieved from BAMPR relative to those retrieved from the Goddard Profiling (Gprof) algorithm and the Precipitation Radar-adjusted TMI estimation of rainfall (PATER) algorithm. Despite a similar inversion approach, the algorithms exhibit different performances that can be mainly related to different training databases and retrieval constraints such as cloud classification.     

神经网络用于反演微波亮度温度分布

将神经网络用于反演目标的微波亮度温度分布,提出了视在温度的神经网络迭代反演法。该方法利用神经网络正向模型提供反演误差对输入的梯度信息,利用该梯度信息进行迭代得到最优反演结果。利用该方法对水面在35GHz频段的视在温度进行反演所得的结果显示,该方法优于Holms反演法及HFR反演法。从视在温度中剔除散射辐射温度得到目标的亮度温度。     

Global Millimeter-Wave Precipitation Retrievals Trained With a Cloud-Resolving Numerical Weather Prediction Model, Part I: Retrieval Design

This paper develops a global precipitation rate retrieval algorithm for the advanced microwave sounding unit (AMSU), which observes 23-191 GHz. The algorithm was trained using a numerical weather prediction (NWP) model (MM5) for 106 globally distributed storms that predicted brightness temperatures consistent with those observed simultaneously by AMSU. Neural networks were trained to retrieve hydrometeor water-paths, peak vertical wind, and 15-min average surface precipitation rates for rain and snow at 15-km resolution at all viewing angles. Different estimators were trained for land and sea, where surfaces classed as snow or ice were generally excluded from this paper. Surface-sensitive channels were incorporated by using linear combinations [principal components (PCs)] of their brightness temperatures that were observed to be relatively insensitive to the surface, as determined by visual examination of global images of each brightness temperature spectrum PC. This paper also demonstrates that multiple scattering in high microwave albedo clouds may help explain the observed consistency for a global set of 122 storms between AMSU-observed 50-191-GHz brightness temperature distributions and corresponding distributions predicted using a cloud-resolving mesoscale NWP model (MM5) and a two-stream radiative transfer model that models icy hydrometeors as spheres with frequency-dependent densities. The AMSU/MM5 retrieval algorithm developed in Part I of this paper is evaluated in Part II on a separate paper.     

Satellite Retrievals of Arctic and Equatorial Rain and Snowfall Rates Using Millimeter Wavelengths

A new global precipitation retrieval algorithm for the millimeter-wave Advanced Microwave Sounding Unit is presented that also retrieves Arctic precipitation rates over surface snow and ice. This algorithm improves upon its predecessor by excluding some surface-sensitive channels and by reducing the number of principal components (PCs) used to represent those that remain. The training sets were also modified to better represent cold regions. The algorithm still incorporates conversion of brightness temperatures to nadir, spatial filtering to better detect pixels scattering near 54 GHz, PC filtering of surface effects, and use of separate neural networks trained with the fifth-generation National Center for Atmospheric Research/Penn State Mesoscale Model (MM5) for land and sea, where warm and cold ocean are now treated differently. The validity of the snowfall detections is supported by nearly coincident CloudSat radar observations, and the physics of the model is largely validated by the reasonable agreement in annual precipitation obtained for 231 globally distributed rain gauges, including many at latitudes where snowfall dominates. Observed annual global precipitation statistics are also presented to permit comparisons with other algorithms and sensors.      

Snow water equivalent retrieval in a Canadian boreal environment from microwave measurements using the HUT snow emission model

Snow water equivalent (SWE) is a critical parameter for climatological and hydrological studies over northern high-latitude areas. In this paper, we study the usability of the Helsinki University of Technology (HUT) snow emission model for the estimation of SWE in a Canadian boreal forest environment. The experimental data (airborne passive microwave and ground-based data) were acquired during the Boreal Ecosystem-Atmosphere Study winter field campaign held in February 1994 in Central Canada. Using the experimental dataset, surface brightness temperatures at 18 and 37 GHz (vertical polarization) were simulated with the HUT snow emission model and compared to those acquired by the airborne sensors. The results showed an important underestimation at 37 GHz (-27 K) and an overestimation at 18 GHz (10 K). In this paper, we demonstrate that the errors in the model simulations are due mainly to the extinction coefficient modeling, which is a function of snow grain size. Therefore, we propose a new semiempirical function for the extinction coefficient, based on an empirical correction to the Rayleigh scattering expression. Results presented in this paper show that the proposed function improves the HUT model accuracy to predict brightness temperature in the experimental context considered, with a mean error of /spl plusmn/5 K and /spl plusmn/9 K, respectively, at 18 and 37 GHz, and a negligible bias (less than 4 K) in both cases. These errors are comparable in magnitude to the accuracy of the radiometers used during the airborne flights. SWE was retrieved using the modified HUT snow emission model based on an iterative inversion technique. SWE was estimated with a mean error of /spl plusmn/10 mm and a negligible bias. Only a rough knowledge of mean snow grain size /spl phi/~ was required in the inversion procedure. The effects of possible errors on mean snow grain size /spl phi/~ are presented and discussed.     

基于遗传BP神经网络算法的主被动遥感协同反演土壤水分

提出了一种基于遗传神经网络算法的主被动遥感协同反演地表土壤水分的方法.首先,建立一个BP神经网络,并采用遗传算法对BP网络的节点权值进行了优化.然后分别将TM数据(TM3,TM4,TM6)、不同极化和极化比的(VV,VH,VH/VV)ASAR数据作为神经网络的输入,土壤水分含量作为网络的输出,用部分实测数据对网络进行训练并反演得到研究区土壤水分布图.最后,利用地面实测数据分别对遗传神经网络优化算法的有效性和主被动遥感协同反演的效果进行了验证,结果表明,新优化算法是有效可行的,且TM和ASAR协同反演的结果比两者单独反演的结果明显要好,体现了主被动遥感协同反演土壤水分的优势与潜力.     

Microwave radiometric technique to retrieve vapor, liquid and ice.I. Development of a neural network-based inversion method

With the advent of the microwave radiometer, passive remote sensing of clouds and precipitation has become an indispensable tool in a variety of meteorological and oceanographical applications. There is wide interest in the quantitative retrieval of water vapor, cloud liquid, and ice using brightness temperature observations in scientific studies such as Earth's radiation budget and microphysical processes of winter and summer clouds. Emission and scattering characteristics of hydrometeors depend on the frequency of observation. Thus, a multifrequency radiometer has the capability of profiling cloud microphysics. Sensitivities of vapor, liquid, and ice with respect to 20.6, 31.65 and 90 GHz brightness temperatures are studied. For the model studies, the atmosphere is characterized by vapor density and temperature profiles and layers of liquid and ice components. A parameterized radiative transfer model is used to quantify radiation emanating from the atmosphere. It is shown that downwelling scattering of radiation by an ice layer results in enhancement at 90 GHz brightness temperature. Once absorptive components such as vapor and liquid are estimated accurately, then it is shown that the ice water path can be retrieved using ground-based three-channel radiometer observations. In this paper the authors developed two- and three-channel neural network-based inversion models. Success of a neural network-based approach is demonstrated using a simulated time series of vapor, liquid, and ice. Performance of the standard explicit inversion model is compared with an iterative inversion model. In part II of this paper, actual radiometer, and radar field measurements are utilized to show practical applicability of the inverse models     

Metamorphic signature of snow revealed in SSM/I measurements

Brightness temperatures (19, 22, 37, 85 GHz) measured by the special sensor microwave/imager (SSM/I) are analyzed using data from the snow monitoring network within the former Soviet Union during the 1987-1988 winter period. It is shown that in the beginning of winter, the SSM/I measurements display the classical snow scattering signature, i.e., the brightness temperatures decrease with increasing depth, and the largest decrease occurs at the highest frequency. Dramatic deviations from this pattern are observed in the middle of winter, where the brightness temperature approaches a minimum and then begins to increase despite the fact that the snow depth remains constant or continues to grow. The two-stream radiative transfer model is combined with results from dense media theory to help explain the phenomenon. Model results suggest that the increase in brightness temperature is due to a decrease of the single scattering albedo as the snowpack ages. This decrease of the albedo is related to changes in the snow crystalline structure due to metamorphism. Consequences for the interpretation of satellite measurements and development of algorithms for deriving snow water equivalent are discussed     

Applications of dense media radiative transfer theory for passivemicrowave remote sensing of foam covered ocean

The effect of the foam covered ocean surface on the passive microwave remote sensing measurements is studied based on the electromagnetic scattering theory. In formulating an electromagnetic scattering model, the authors treat the foam as densely packed sticky air bubbles coated with thin seawater coating. The layer of foam covers the ocean surface that has air bubbles. They then use dense media radiative transfer (DMRT) theory with quasi-crystalline approximation (QCA) for densely distributed sticky moderate size particles to calculate the brightness temperatures of the foam-covered ocean surface. Results are illustrated for 19 GHz and 37 GHz and for both vertical and horizontal polarizations as a function of foam microstructure properties and foam layer thickness. Comparisons are also made with experimental measurements     

A physical model to determine snowfall over land by microwave radiometry

Falling snow is an important component of global precipitation in extratropical regions. This paper describes the methodology and results of physically based retrievals of snow falling over land surfaces. Because microwave brightness temperatures emitted by snow-covered surfaces are highly variable, precipitating snow above such surfaces is difficult to observe using window channels that occur at low frequencies (/spl nu/<100 GHz). Furthermore, at frequencies /spl nu//spl les/37 GHz, sensitivity to liquid hydrometeors is dominant. These problems are mitigated at high frequencies (/spl nu/>100 GHz) where water vapor screens the surface emission, and sensitivity to frozen hydrometeors is significant. However, the scattering effect of snowfall in the atmosphere at those higher frequencies is also impacted by water vapor in the upper atmosphere. The theory of scattering by randomly oriented dry snow particles at high microwave frequencies appears to be better described by regarding snow as a concatenation of "equivalent" ice spheres rather than as a sphere with the effective dielectric constant of an air-ice mixture. An equivalent sphere snow scattering model was validated against high-frequency attenuation measurements. Satellite-based high-frequency observations from an Advanced Microwave Sounding Unit (AMSU-B) instrument during the March 5-6, 2001 New England blizzard were used to retrieve snowfall over land. Vertical distributions of snow, temperature, and relative humidity profiles were derived from the Mesoscale Model (MM5) cloud model. Those data were applied and modified in a radiative transfer model that derived brightness temperatures consistent with the AMSU-B observations. The retrieved snowfall distribution was validated with radar reflectivity measurements obtained from a ground-based radar network.     

Nonlinear statistical retrievals of ice content and rain rate frompassive microwave observations of a simulated convective storm

A numerical simulator for analysis of multispectral passive microwave mapping and retrieval is described. This simulator allows evaluation and optimization of satellite-based cloud and precipitation parameter retrieval algorithms. It contains three major components: the forward radiative transfer model, the sensor observation model, and the parameter retrieval algorithm. Simulated spaceborne observations of an oceanic tropical squall sampled at five stages in time are demonstrated for a simplified version of the proposed Earth Observation System (EOS) Multifrequency Imaging Microwave Radiometer (MIMR). The simulator uses a nonlinear statistical retrieval algorithm consisting of a Karhunen-Loeve (KL) transform, a projection operator, a nonlinear inverse mapping and a linear minimum mean-square error estimator. Retrievals of rain rate and integrated ice content are performed for each evolutionary frame at both full spatial resolution (1.5 km) and the degraded spatial resolution of a MIMR-class system. Results are presented for both KL-based and brightness temperature-based retrieval algorithms. It is found that the KL-based algorithm has a reduced complexity and performs better than the brightness temperature-based algorithm for degraded resolution imagery, especially for rain rate retrievals. In addition, rain rate retrievals are more affected by low image resolution than are integrated ice content retrievals. Retrieval accuracy of both rain and integrated ice is also found to depend on the evolutionary stage of the storm     

Mapping the spatial distribution and time evolution of snow water equivalent with passive microwave measurements

This paper presents an algorithm that estimates the spatial distribution and temporal evolution of snow water equivalent and snow depth based on passive remote sensing measurements. It combines the inversion of passive microwave remote sensing measurements via dense media radiative transfer modeling results with snow accumulation and melt model predictions to yield improved estimates of snow depth and snow water equivalent, at a pixel resolution of 5 arc-min. In the inversion, snow grain size evolution is constrained based on pattern matching by using the local snow temperature history. This algorithm is applied to produce spatial snow maps of Upper Rio Grande River basin in Colorado. The simulation results are compared with that of the snow accumulation and melt model and a linear regression method. The quantitative comparison with the ground truth measurements from four Snowpack Telemetry (SNOTEL) sites in the basin shows that this algorithm is able to improve the estimation of snow parameters.     

Sensitivity of passive microwave snow depth retrievals to weather effects and snow evolution

Snow fall and snow accumulation are key climate parameters due to the snow's high albedo, its thermal insulation, and its importance to the global water cycle. Satellite passive microwave radiometers currently provide the only means for the retrieval of snow depth and/or snow water equivalent (SWE) over land as well as over sea ice from space. All algorithms make use of the frequency-dependent amount of scattering of snow over a high-emissivity surface. Specifically, the difference between 37- and 19-GHz brightness temperatures is used to determine the depth of the snow or the SWE. With the availability of the Advanced Microwave Scanning Radiometer (AMSR-E) on the National Aeronautics and Space Administration's Earth Observing System Aqua satellite (launched in May 2002), a wider range of frequencies can be utilized. In this study we investigate, using model simulations, how snow depth retrievals are affected by the evolution of the physical properties of the snow (mainly grain size growth and densification), how they are affected by variations in atmospheric conditions and, finally, how the additional channels may help to reduce errors in passive microwave snow retrievals. The sensitivity of snow depth retrievals to atmospheric water vapor is confirmed through the comparison with precipitable water retrievals from the National Oceanic and Atmospheric Administration's Advanced Microwave Sounding Unit (AMSU-B). The results suggest that a combination of the 10-, 19-, 37-, and 89-GHz channels may significantly improve retrieval accuracy. Additionally, the development of a multisensor algorithm utilizing AMSR-E and AMSU-B data may help to obtain weather-corrected snow retrievals.     

Land-surface-type classification using microwave brightnesstemperatures from the Special Sensor Microwave/Imager

The use of empirical parameter retrieval algorithms over land requires the prior classification of surface types according to their microwave emission properties. A land-surface-type classification scheme was developed to be used with the Special Sensor Microwave/Imager (SSM/I) algorithm package. The classification rules were based on statistical analysis of SSM/I brightness temperature combinations from several surfaces, including dense vegetation, rangeland and agricultural soils, deserts, snow, precipitation, surface moisture, etc. A set of independent classification rules was derived which should result in increased confidence of parameter retrievals     

A recurrent neural network classifier for improved retrievals ofareal extent of snow cover

Accurate detection of areal extent of snow in mountainous regions is important. Areal extent of snow is a useful climatic indicator. Moreover, snow melt is a major source of water supply for many arid regions (e.g., western United States, Morocco) and affects regional ecosystems. Unfortunately, accurate satellite retrievals of areal extent of snow have been difficult to achieve. Two approaches to effectively and accurately detect clear land, cloud, and areal extent of snow in satellite data are developed. A feed-forward neural network (FFNN) is used to classify individual images, and a recurrent NN is used to classify sequences of images. The continuous outputs of the NN, combined with a linear mixing model, provide support for mixed-pixel classification. Validation with independent in situ data confirms the classification accuracy (94% for feed-forward NN, 97% for recurrent NN). The combination of rapid temporal sampling (e.g., GOES) and a recurrent NN classifier is recommended (relative to an isolated scene (e.g., AVHRR) and a feed-forward NN classifier)