Search

Over the past decade the use of machine learning in meteorology has grown rapidly. Specifically neural networks and deep learning have been used at an unprecedented rate. To fill the dearth of resources covering neural networks with a meteorological lens, this paper discusses machine learning methods in a plain language format that is targeted to the operational meteorological community. This is the second paper in a pair that aim to serve as a machine learning resource for meteorologists. While the first paper focused on traditional machine learning methods (e.g., random forest), here a broad spectrum of neural networks and deep learning methods is discussed. Specifically, this paper covers perceptrons, artificial neural networks, convolutional neural networks, and U-networks. Like the Part I paper, this manuscript discusses the terms associated with neural networks and their training. Then the manuscript provides some intuition behind every method and concludes by showing each method used in a meteorological example of diagnosing thunderstorms from satellite images (e.g., lightning flashes). This paper is accompanied with an open-source code repository to allow readers to explore neural networks using either the dataset provided (which is used in the paper) or as a template for alternate datasets. [ABSTRACT FROM AUTHOR]

Conventional statistical postprocessing techniques offer limited ability to improve the skills of probabilistic guidance for heavy precipitation. This paper introduces two artificial neural network (ANN)-based, geographically aware, and computationally efficient postprocessing schemes, namely, the ANN-multiclass (ANN-Mclass) and the ANN–censored, shifted gamma distribution (ANN-CSGD). Both schemes are implemented to postprocess Global Ensemble Forecast System (GEFS) forecasts to produce probabilistic quantitative precipitation forecasts (PQPFs) over the contiguous United States (CONUS) using a short (60 days), rolling training window. The performances of these schemes are assessed through a set of hindcast experiments, wherein postprocessed 24-h PQPFs from the two ANN schemes were compared against those produced using the benchmark quantile mapping algorithm for lead times ranging from 1 to 8 days. Outcomes of the hindcast experiments show that ANN schemes overall outperform the benchmark as well as the raw forecast over the CONUS in predicting probability of precipitation over a range of thresholds. The relative performance varies among geographic regions, with the two ANN schemes broadly improving upon quantile mapping over the central, south, and southeast, and slightly underperforming along the Pacific coast where skills of raw forecasts are the highest. Between the two schemes, the hybrid ANN-CSGD outperforms at higher rainfall thresholds (i.e., >50 mm day−1), though the outperformance comes at a slight expense of sharpness and spatial specificity. Collectively, these results confirm the ability of the ANN algorithms to produce skillful PQPFs with a limited training window and point to the prowess of the hybrid scheme for calibrating PQPFs for rare-to-extreme rainfall events. [ABSTRACT FROM AUTHOR]

Tropical Rainfall Measuring Mission (TRMM) satellite products constitute valuable precipitation datasets over regions with sparse rain gauge networks. Downscaling is an effective approach to estimating the precipitation over ungauged areas with high spatial resolution. However, a large bias and low resolution of original TRMM satellite images constitute constraints for practical hydrologic applications of TRMM precipitation products. This study contributes two precipitation downscaling algorithms by exploring the nonstationarity relations between precipitation and various environment factors [daytime surface temperature (LTD), terrain slope, normalized difference vegetation index (NDVI), altitude, longitude, and latitude] to overcome bias and low-resolution constraints of TRMM precipitation. Downscaling of precipitation is achieved with the geographically weighted regression model (GWR) and the backwardpropagation artificial neural networks (BP_ANN). The probability density function (PDF) algorithm corrects the bias of satellite precipitation data with respect to spatial and temporal scales prior to downscaling. The principal component analysis algorithm (PCA) provides an alternative method of obtaining accurate monthly rainfall estimates during the wet rainfall season that minimizes the temporal uncertainties and upscaling effects introduced by direct accumulation (DA) of precipitation. The performances of the proposed downscaling algorithms are assessed by downscaling the latest version of TRMM3B42 V7 datasets within Hubei Province from 0.258 (about 25 km) to 1-km spatial resolution at the monthly scale. The downscaled datasets are systematically evaluated with in situ observations at 27 rain gauges from the years 2005 through 2010. This paper's results demonstrate the bias correction is necessary before downscaling. The highresolution precipitation datasets obtained with the proposed downscaling model with GWR relying on the NDVI and slope are shown to improve the accuracy of precipitation estimates. GWR exhibits more accurate downscaling results than BP_ANN coupled with the genetic algorithm (GA) in most dry and wet seasons. [ABSTRACT FROM AUTHOR]

Initial abstraction (Ia) is a sensitive parameter in hydrological models, and its value directly determines the amount of runoff. Ia, which is influenced by many factors related to antecedent watershed condition (AWC), is difficult to estimate due to lack of observed data. In the Soil Conservation Service curve number (SCS-CN) method, it is often assumed that Ia is 0.2 times the potential maximum retention S. Yet this assumption has frequently been questioned. In this paper, Ia/S and factors potentially influencing Ia were collected from rainfall–runoff events. Soil moisture and evaporation data were extracted from GLDAS-Noah datasets to represent AWC. Based on the driving factors of Ia, identified using the Pearson correlation coefficient and maximal information coefficient, artificial neural network (ANN)-estimated Ia was applied to simulate the selected flood events in the Hydrologic Engineering Center Hydrologic Modeling System (HEC-HMS) model. The results indicated that Ia/S varies over different events and different watersheds. Over 75% of the Ia/S values are less than 0.2 in the two study areas. The driving factors affecting Ia vary over different watersheds, and the antecedent precipitation index appears to be the most influential factor. Flood simulation by the HEC-HMS model using statistical Ia gives the best fitness, whereas applying ANN-estimated Ia outperforms the simulation with median Ia/S. For over 60% of the flood events, ANN-estimated Ia provided better fitness in flood peak and depth, with an average Nash–Sutcliffe efficiency coefficient of 0.76 compared to 0.71 for median Ia/S. The proposed ANN-estimated Ia is physically based and can be applied without calibration, saving time in constructing hydrological models. [ABSTRACT FROM AUTHOR]

Atmospheric visibility is an important element of meteorological observation. With existing methods, defining image features that reflect visibility accurately and comprehensively is difficult. This paper proposes a visibility detection method based on transfer learning using deep convolutional neural networks (DCNN) that addresses issues caused by a lack of sufficient visibility labeled datasets. In the proposed method, each image was first divided into several subregions, which were encoded to extract visual features using a pretrained no-reference image quality assessment neural network. Then a support vector regression model was trained to map the extracted features to the visibility. The fusion weight of each subregion was evaluated according to the error analysis of the regression model. Finally, the neural network was fine-tuned to better fit the problem of visibility detection using the current detection results conversely. Experimental results demonstrated that the detection accuracy of the proposed method exceeds 90% and satisfies the requirements of daily observation applications. [ABSTRACT FROM AUTHOR]

Machine learning algorithms are trained on a 10-yr archive of composite weather radar images in the Swiss Alps to nowcast precipitation growth and decay in the next few hours in moving coordinates (Lagrangian frame). The hypothesis of this study is that growth and decay is more predictable in mountainous regions, which represent a potential source of practical predictability by machine learning methods. In this paper, artificial neural networks (ANN) are employed to learn the complex nonlinear dependence relating the growth and decay to the input predictors, which are geographical location, mesoscale motion vectors, freezing level height, and time of the day. The average long-term growth and decay patterns are effectively reproduced by the ANN, which allows exploring their climatology for any combination of predictors. Due to the low intrinsic predictability of growth and decay, its prediction in real time is more challenging, but is substantially improved when adding persistence information to the predictors, more precisely the growth and decay and precipitation intensity in the immediate past. The improvement is considerable in mountainous regions, where, depending on flow direction, the root-mean-square error of ANN predictions can be 20%–30% lower compared with persistence. Because large uncertainty is associated with precipitation forecasting, deterministic machine learning predictions should be coupled with a model for the predictive uncertainty. Therefore, we consider a probabilistic perspective by estimating prediction intervals based on a combination of quantile decision trees and ANNs. The probabilistic framework is an attempt to address the problem of conditional bias, which often characterizes deterministic machine learning predictions obtained by error minimization. [ABSTRACT FROM AUTHOR]

Echoes in radar reflectivity data do not always correspond to precipitating particles. Echoes on radar may result from biological targets such as insects, birds, or wind-borne particles; from anomalous propagation or ground clutter; or from test and interference patterns that inadvertently seep into the final products. Although weather forecasters can usually identify and account for the presence of such contamination, automated weather-radar algorithms are drastically affected. Several horizontal and vertical features have been proposed to discriminate between precipitation echoes and echoes that do not correspond to precipitation. None of these features by themselves can discriminate between precipitating and nonprecipitating areas. In this paper, a neural network is used to combine the individual features, some of which have already been proposed in the literature and some of which are introduced in this paper, into a single discriminator that can distinguish between “good” and “bad” echoes (i.e., precipitation and nonprecipitation, respectively). The method of computing the horizontal features leads to statistical anomalies in their distributions near the edges of echoes. Also described is how to avoid presenting such range gates to the neural network. The gate-by-gate discrimination provided by the neural network is followed by more holistic postprocessing based on spatial contiguity constraints and object identification to yield quality-controlled radar reflectivity scans that have most of the bad echo removed while leaving most of the good echo untouched. A possible multisensor extension, utilizing satellite data and surface observations, to the radar-only technique is also demonstrated. It is demonstrated that the resulting technique is highly skilled and that its skill exceeds that of the currently operational algorithm. [ABSTRACT FROM AUTHOR]

Despite the advantage of global coverage at high spatiotemporal resolutions, satellite remotely sensed precipitation estimates still suffer from insufficient accuracy that needs to be improved for weather, climate, and hydrologic applications. This paper presents a framework of a deep neural network (DNN) that improves the accuracy of satellite precipitation products, focusing on reducing the bias and false alarms. The state-of-the-art deep learning techniques developed in the area of machine learning specialize in extracting structural information from a massive amount of image data, which fits nicely into the task of retrieving precipitation data from satellite cloud images. Stacked denoising autoencoder (SDAE), a widely used DNN, is applied to perform bias correction of satellite precipitation products. A case study is conducted on the Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks Cloud Classification System (PERSIANN-CCS) with spatial resolution of 0.08° × 0.08° over the central United States, where SDAE is used to process satellite cloud imagery to extract information over a window of 15 × 15 pixels. In the study, the summer of 2012 (June-August) and the winter of 2012/13 (December-February) serve as the training periods, while the same seasons of the following year (summer of 2013 and winter of 2013/14) are used for validation purposes. To demonstrate the effectiveness of the methodology outside the study area, three more regions are selected for additional validation. Significant improvements are achieved in both rain/no-rain ( R/NR) detection and precipitation rate quantification: the results make 33% and 43% corrections on false alarm pixels and 98% and 78% bias reductions in precipitation rates over the validation periods of the summer and winter seasons, respectively. [ABSTRACT FROM AUTHOR]

Accurate temperature forecasts are critical for various industries and sectors. We propose a probabilistic neural network (PNN), an extension of the distributional regression network (DRN), for 2-m temperature forecasts, consisting of three variants with different inputs and target variables. The first variant, standardized anomaly probabilistic neural network (SAPNN), employs a two-step approach involving standardized anomalies and global PNN modeling to effectively capture underlying features and anomalies. The second variant, PNN with geographical predictors (PNNGE), incorporates raw and static geographical predictors to enhance predictive performance. The third variant, PNN with station one-hot encoding (PNNEN), utilizes raw with station one-hot encoding predictors to represent geographical information effectively. We compare three PNN variants with two benchmarks: 1) standardized anomaly model output statistics (SAMOS) and 2) three DRN variants identical to those applied to PNN. These evaluations utilize ECMWF data from 2019 to 2020 at 6-h intervals up to 72 h over Hebei, China. Results show that SAPNN and PNNGE are better than SAMOS, while PNNEN notably exhibits a significant 14% improvement in the continuous ranked probability skill score (CRPSS). Moreover, the PNN variants exhibit comparable or superior performance to DRN regarding forecast accuracy, CRPSS, and reliability, showcasing a better-calibrated spread–error relationship. This study highlights the value of the proposed PNN variants with a distribution output in capturing nonlinear relationships within different sources of predictors and improving temperature forecast skills. Significance Statement: This study aims to improve the accuracy, skills, and reliability of 2-m temperature forecasts, which are crucial in agriculture and energy management. To achieve this, we extend a popular artificial intelligence framework and explore four data sources with two schemes to systematically compare the predictive performances in making temperature forecasts. The findings of this research are vital as they offer novel ways to improve forecast skills. Imagine having a weather app that is significantly more accurate, enabling you to plan your day better. This study is about discovering innovative approaches to enhance forecast skills and reliability, which could benefit various aspects of our daily lives. One of the new methods even exhibits a 14% improvement in forecast skills. [ABSTRACT FROM AUTHOR]

Accurately estimating tropical cyclone (TC) intensity is one of the most critical steps in TC forecasting and disaster warning/management. For over 40 years, the Dvorak technique (and several improved versions) has been applied for estimating TC intensity by forecasters worldwide. However, the operational Dvorak techniques primarily used in various agencies have several deficiencies, such as inherent subjectivity leading to inconsistent intensity estimates within various basins. This collaborative study between meteorologists and data scientists has developed a deep-learning model using satellite imagery to estimate TC intensity. The conventional convolutional neural network (CNN), which is a mature technology for object classification, requires several modifications when being used for directly estimating TC intensity (a regression task). Compared to the Dvorak technique, the CNN model proposed here is objective and consistent among various basins; it has been trained with satellite infrared brightness temperature and microwave rain-rate data from 1097 global TCs during 2003–14 and optimized with data from 188 TCs during 2015–16. This paper also introduces an upgraded version that further improves the accuracy by using additional TC information (i.e., basin, day of year, local time, longitude, and latitude) and applying a postsmoothing procedure. An independent testing dataset of 94 global TCs during 2017 has been used to evaluate the model performance. A root-mean-square intensity difference of 8.39 kt (1 kt ≈ 0.51 m s−1) is achieved relative to the best track intensities. For a subset of 482 samples analyzed with reconnaissance observations, a root-mean-square intensity difference of 8.79 kt is achieved. [ABSTRACT FROM AUTHOR]

Compared to ground precipitation measurements, satellite-based precipitation estimation products have the advantage of global coverage and high spatiotemporal resolutions. However, the accuracy of satellitebased precipitation products is still insufficient to serve many weather, climate, and hydrologic applications at high resolutions. In this paper, the authors develop a state-of-the-art deep learning framework for precipitation estimation using bispectral satellite information, infrared (IR), and water vapor (WV) channels. Specifically, a two-stage framework for precipitation estimation from bispectral information is designed, consisting of an initial rain/no-rain (R/NR) binary classification, followed by a second stage estimating the nonzero precipitation amount. In the first stage, the model aims to eliminate the large fraction of NR pixels and to delineate precipitation regions precisely. In the second stage, the model aims to estimate the pointwise precipitation amount accurately while preserving its heavily skewed distribution. Stacked denoising autoencoders (SDAEs), a commonly used deep learning method, are applied in both stages. Performance is evaluated along a number of common performance measures, including both R/NR and real-valued precipitation accuracy, and compared with an operational product, Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks-Cloud Classification System (PERSIANN-CCS). For R/NR binary classification, the proposed two-stage model outperforms PERSIANN-CCS by 32.56% in the critical success index (CSI). For real-valued precipitation estimation, the two-stage model is 23.40% lower in average bias, is 44.52% lower in average mean squared error, and has a 27.21% higher correlation coefficient. Hence, the two-stage deep learning framework has the potential to serve as a more accurate and more reliable satellite-based precipitation estimation product. The authors also provide some future directions for development of satellite-based precipitation estimation products in both incorporating auxiliary information and improving retrieval algorithms. [ABSTRACT FROM AUTHOR]

This paper evaluates the quality of neural network classifiers for wind speed and wind gust prediction with prediction lead times between +1 and +24 h. The predictions were realized based on local time series and model data. The selection of appropriate input features was initiated by time series analysis and completed by empirical comparison of neural network classifiers trained on several choices of input features. The selected input features involved day time, yearday, features from a single wind observation device at the site of interest, and features derived from model data. The quality of the resulting classifiers was benchmarked against persistence for two different sites in Switzerland. The neural network classifiers exhibited superior quality when compared with persistence judged on a specific performance measure, hit and false-alarm rates. [ABSTRACT FROM AUTHOR]

Principal component analysis (PCA) is one of the most widely used methods in the examination of climate data. However, PCA of a dataset is handicapped if the data size is large. PCA of a large dataset would require huge computer resources in terms of memory and CPU time. Neural network principal component analysis (NNPCA), which has been used mainly in the signal-processing field, can be a useful tool in the analysis of large climate datasets. NNPCA requirements of computer memory and CPU time are far less than what are needed by conventional methods of PCA, such as singular value decomposition of the data matrix. In this paper, an NNPCA application to climate data is introduced. NNPCA is applied to reanalysis data of monthly and daily global maps of the 850-hPa geopotential height from the National Centers for Environmental Prediction–National Center for Atmospheric Research reanalysis data. These data, covering the period from 1948 to 2003, are composed of 648 monthly maps and 20 117 daily ones. The first 10 principal components (PCs) of the daily data contribute up to 58.27% of the data total variance, whereas the first 10 PCs of monthly data represent 80.72% of the total variance. The first, second, and third PCs describe the annual variation, southern annular mode, and northern annular mode, respectively. [ABSTRACT FROM AUTHOR]

This study details a two-method, machine learning approach to predict current and short-term intensity change in global tropical cyclones (TCs), "D-MINT" and "D-PRINT." D-MINT and D-PRINT use infrared imagery and environmental scalar predictors, while D-MINT also employs microwave imagery. Results show that current TC intensity estimates from D-MINT and D-PRINT are more skillful than three established intensity estimation methods routinely used by operational forecasters for North Atlantic and eastern and western North Pacific TCs. Short-term intensity predictions are validated against five operational deterministic guidances at 6-, 12-, 18-, and 24-h lead times. D-MINT and D-PRINT are less skillful than NHC and consensus TC intensity predictions in North Atlantic and eastern North Pacific TCs, but are more skillful than the other guidances for at least half of the lead times. In western North Pacific, north Indian Ocean, and Southern Hemisphere TCs, D-MINT is more skillful than the JTWC and other individual TC intensity forecasts for over half of the lead times. When probabilistically predicting TC rapid intensification (RI), D-MINT is more skillful in North Atlantic and western North Pacific TCs than three operationally used RI guidances, but less skillful for yes–no RI forecasts. In addition, this work demonstrates the importance of microwave imagery, as D-MINT is more skillful than D-PRINT. Since D-MINT and D-PRINT are convolutional neural network models interrogating two-dimensional structures within TC satellite imagery, this study also demonstrates that those features can yield better short-term predictions than existing scalar statistics of satellite imagery in operational models. Finally, a diagnostics tool is revealed to aid the attribution of the D-MINT/D-PRINT intensity predictions. Significance Statement: This study develops a method to predict current and short-term forecasts of tropical cyclone (TC) intensity using artificial intelligence. The resultant models use a convolutional neural network (CNN) that can identify two-dimensional features in satellite imagery that are indicative of TC intensity and future intensity change. The performance results indicate that in several TC basins, the CNN approach is generally more skillful than alternative satellite-based estimates of TC intensity as well as operational short-term forecasts of deterministic intensity change and of similar skill to probabilistic rapid intensification forecasts. [ABSTRACT FROM AUTHOR]

The prediction of sea surface temperature (SST) in real-time or online mode has applications in planning marine operations and forecasting climate. This paper demonstrates how SST measurements can be combined with numerical estimations with the help of neural networks and how reliable site-specific forecasts can be made accordingly. Additionally, this work demonstrates the skill of a special wavelet neural network in this task. The study was conducted at six different locations in the Indian Ocean and over three time scales (daily, weekly, and monthly). At every time step, the difference between the numerical estimation and the SST measurement was evaluated, an error time series was formed, and errors over future time steps were forecasted. The time series forecasting was affected through neural networks. The predicted errors were added to the numerical estimation, and SST predictions were made over five time steps in the future. The performance of this procedure was assessed through various error statistics, which showed a highly satisfactory functioning of this scheme. The wavelet neural network based on the particular basic or mother wavelet called the 'Meyer wavelet with discrete approximation' worked more satisfactorily than other wavelets. [ABSTRACT FROM AUTHOR]

This paper describes the development and testing of a cloud-regime-dependent short-range solar irradiance forecasting system for predictions of 15-min-average clearness index (global horizontal irradiance). This regime-dependent artificial neural network (RD-ANN) system classifies cloud regimes with a k-means algorithm on the basis of a combination of surface weather observations, irradiance observations, and GOES-East satellite data. The ANNs are then trained on each cloud regime to predict the clearness index. This RD-ANN system improves over the mean absolute error of the baseline clearness-index persistence predictions by 1.0%, 21.0%, 26.4%, and 27.4% at the 15-, 60-, 120-, and 180-min forecast lead times, respectively. In addition, a version of this method configured to predict the irradiance variability predicts irradiance variability more accurately than does a smart persistence technique. [ABSTRACT FROM AUTHOR]

This paper presents an approach for regional flood frequency analysis (RFFA) in the presence of nonlinearity and problematic stations, which require adapted methodologies. To this end, the projection pursuit regression (PPR) is proposed. The PPR is a family of regression models that applies smooth functions on intermediate predictors to fit complex patterns. The PPR approach can be seen as a hybrid method between the generalized additive model (GAM) and the artificial neural network (ANN), which combines the advantages of both methods. Indeed, the PPR approach has the structure of a GAM to describe nonlinear relations between hydrological variables and other basin characteristics. On the other hand, PPR can consider interactions between basin characteristics to improve the predictive capabilities in a similar way to ANN, but simpler. The methodology developed in the present study is applied to a case study represented by hydrometric stations from southern Québec, Canada. It is shown that flood quantiles are mostly associated with a dominant intermediate predictor, which provides a parsimonious representation of the nonlinearity in the flood-generating processes. The model performance is compared to eight other methods available in the literature for the same dataset, including GAM and ANN. When using the same basin characteristics, the results indicate that the simpler structure of PPR does not affect the global performance and that PPR is competitive with the best existing methods in RFFA. Particular attention is also given to the performance resulting from the choice of the basin characteristics and the presence of problematic stations. [ABSTRACT FROM AUTHOR]

This paper evaluates the applicability of neural networks for estimating wind speeds at various target locations using neighboring reference locations along the south coast of Newfoundland, Canada. The stations were chosen to cover a variety of topographic features and span distances in excess of 100 km. The goal of the study is to provide a general description of the summer wind conditions along the south coast of Newfoundland and to assess the potential application of neural networks for wind speed predictions. Analysis of wind data from July to October showed the wind going dominantly toward the northeast with speeds ranging from 0 to 45 m s−1. The efficacy of neural networks to predict wind speeds varied among stations and was largely influenced by the presence/absence of wind barriers. Sensitivity analysis on neural network performance concluded that an absolute minimum of 3000 h of continuous monitoring is needed to effectively train neural networks to predict wind speeds. The conclusions of this study have implications for future work utilizing wind speed data where a generalization of uniform wind speeds is assumed. [ABSTRACT FROM AUTHOR]

A constructive back-propagation code that was designed to run as a single-hidden-layer, feed-forward neural network (SLFFNN) has been adapted and used to estimate subsurface temperature from a small volume of magnetotelluric (MT)-derived electrical resistivity data and borehole thermograms. The code was adapted to use a looping procedure in searching for better initialization conditions that can optimally solve nonlinear problems using the random weight initialization approach. Available one-dimensional (1D) MT-derived resistivity data and borehole temperature records from the Tattapani geothermal field, central India, were collated and digitized at 10-m intervals. The two datasets were paired to form a set of input–output pairs. The paired data were randomized, standardized, and partitioned into three mutually exclusive subsets. The various subsets had 52% (later increased to 61%), 30%, and 18% (later reduced to 9%) for training, validation, and testing, respectively, in the first and second training phases. The second training phase was meant to assess the influence of the training data volume on network performance. Standard statistical techniques including adjusted coefficient of determination (R2a), relative error (ϵ), absolute average deviation (AAD), root-mean-square error (RMSE), and regression analysis were used to quantitatively rate network performance. A manually designed two-hidden-layer, feed-forward network with 20 and 15 neurons in the first and second layers was also adopted in solving the same problem. Performance ratings were observed to be 0.97, 3.75, 4.09, 1.41, 1.18, and 1.08 for R2a, AAD, ϵ, RMSE, slope, and intercept, respectively, compared to an ϵ of 20.33 observed with the manually designed network. The SLFFNN is thus a structurally flexible network that performs better in spite of the small volume of data used in testing the network. The network needs to be tested further. [ABSTRACT FROM AUTHOR]

Scarcity or unavailability of precipitation observation creates difficulties in hydrologic modeling of mountainous sections of the arid region of northwest China (34°-50°N, 72°-107°E). Tropical Rainfall Measuring Mission (TRMM) precipitation products may be a potential substitute, but they should be evaluated and corrected with ground observation data before application. In this paper, two TRMM Multisatellite Precipitation Analysis (TMPA) precipitation products were evaluated by gauge observations, using indices such as frequency bias index, probability of detection, false alarm ratio, relative mean bias, Nash-Sutcliffe efficiency, and correlation coefficient. Terrain variables were extracted from a digital elevation model, and their rotated principal components were determined to establish a stepwise regression model to adjust TMPA precipitation. Additionally, a back-propagation (BP) neural network was established to correct TMPA precipitation. The results showed that TMPA had an unsatisfactory detection ability in the study area for both precipitation occurrence and amount. TMPA precipitation corrected by a stepwise regression method showed some improvement, but only the results for TRMM 3B43 on a subregion scale were acceptable. The BP neural network method showed better results than the stepwise regression method, and both TRMM 3B42 and TRMM 3B43 corrected by the former method on a subregion scale could be acceptable. Both methods were spatial-scale dependent and showed better results on a subregion scale than on a larger scale. [ABSTRACT FROM AUTHOR]

The current Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER)-based broadband albedomodel requires shortwave infrared bands 5 (2.145-2.185 nm), 6 (2.185-2.225 nm), 8 (2.295-2.365 nm), and 9 (2.360-2.430 nm) and visible/near-infrared bands 1 (0.52-0.60 nm) and 3 (0.78-0.86 nm). However, because of sensor irregularities at high temperatures, shortwave infrared wavelengths are not recorded in the ASTER data acquired after April 2008. Therefore, this study seeks to evaluate the performance of artificial neural networks (ANN) in estimating surface albedo using visible/nearinfrared bands available in the data obtained after April 2008. It also compares the outcomes with the results of multiple linear regression (MLR) modeling. First, the most influential spectral bands used in the current model as well as band 2 (0.63-0.69 nm) (which is also available after April 2008 in the visible/nearinfrared part) were determined by a primary analysis of the data acquired before April 2008. Then, multiple linear regression and ANN models were developed by using bands with a relatively high level of contribution. The results showed that bands 1 and 3 were the most important spectral ones for estimating albedo where land cover consisted of soil and vegetation. These two bands were used as the study input, and the albedo (estimated through a model that utilized bands 1, 3, 5, 6, 8, and 9) served as a target to remodel albedo. Because of its high collinearity with band 1, band 2 was identified less effectively by MLR as well as ANN. The study confirmed that a combination of bands 1 and 3, which are available in the current ASTER data, could be modeled through ANN and MLR to estimate surface albedo. However, because of its higher accuracy, ANN method was superior to MLR in developing objective functions. [ABSTRACT FROM AUTHOR]

Satellite-derived high-resolution precipitation products (HRPP) have been developed to address the needs of the user community and are now available with 0.25° × 0.25° (or less) subdaily resolutions. This paper evaluates a number of commonly available satellite-derived HRPPs covering northwest Europe over a 6-yr period. Precipitation products include the Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis (TMPA), the Climate Prediction Center (CPC) morphing (CMORPH) technique, the CPC merged microwave technique, the Naval Research Laboratory (NRL) blended technique, and the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) technique. In addition, the Geosynchronous Operational Environmental Satellite (GOES) precipitation index (GPI) and the European Centre for Medium-Range Weather Forecasting (ECMWF) operational forecast model products are included for comparison. Surface reference data from the European radar network is used as ground truth, supported by the Global Precipitation Climatology Centre (GPCC) precipitation gauge analysis and gauge data over the United Kingdom. Measures of correlation, bias ratio, probability of detection, and false alarm ratio are used to evaluate the products. Results show that satellite products generally exhibit a seasonal cycle in correlation, bias ratio, probability of detection, and false alarm ratio, with poorer statistics during the winter. The ECMWF model also shows a seasonal cycle in the correlation, although the results are poorer during the summer, while the bias ratio, probability of detection, and false alarm ratio are consistent through all seasons. Importantly, all the satellite HRPPs underestimate precipitation over northwest Europe in all seasons. [ABSTRACT FROM AUTHOR]

The paper presents an approach for conditional airmass classification based on local precipitation rate distributions. The method seeks, within the potential region, three-dimensional atmospheric predictor domains with high impact on the local-scale phenomena. These predictor domains are derived by an algorithm consisting of a clustering method, namely, self-organizing maps, and a nonlinear optimization method, simulated annealing. The findings show that the resulting spatial structures can be attributed to well-known atmospheric processes. Since the optimized predictor domains probably contain relevant information for precipitation generation, these grid points may also be potential inputs for nonlinear downscaling methods. Based on this assumption, the potential of these optimized large-scale predictors for downscaling has been investigated by applying an artificial neural network as a nonparametric statistical downscaling model. Compared to preset local predictors, using the optimized predictors improves the accuracy of the downscaled time series, particularly in summer and autumn. However, optimizing predictors by a conditional classification does not guarantee that a predictor increases the explained variance of the downscaling model. To study the contribution of each predictor to the output variance, either individually or by interactions with other parameters, the sources of uncertainty have been estimated by global sensitivity analysis, which provides model-free sensitivity measures. It is shown that predictor interactions play an important part in the modeling process and should be taken into account in the predictor screening. [ABSTRACT FROM AUTHOR]

In a recent paper by Kit et al., a novel algorithm for the calibration of hot-film probes using a collocated sonic anemometer combined with the neural network approach is described. An important step in the algorithm is the generation of a calibration dataset by an appropriate low-pass filtering of the voltage and velocity time series obtained from hot-film probes and a sonic anemometer, correspondingly. Kit et al. report that a polynomial least squares fit that was used to approximate the relations of these voltage--velocity data from the dataset failed while a neural network approach worked satisfactorily. The same polynomial fit worked successfully with a calibration dataset obtained using a standard calibration unit that enables one to generate calibration data at evenly distributed yaw angles, varying in a wide range (−−30°°, 30°°). In the current study, an attempt is made to uncover the reason for the failure of the polynomial fit algorithm with a sonic anemometer--based calibration dataset (SBS-PF). The probability densities of the velocity angles for the calibration dataset, as well as for a full velocity dataset obtained using the neural network approach, are computed. Also developed are theoretical expressions for the same angular density probability distributions based on the following assumptions: (i) an axisymmetric turbulent velocity field, (ii) Gaussian density probability distribution for velocity components, and (iii) weak correlations between the velocity components (i.e., the probability density distribution of the entire velocity vector is a product of probabilities of its components). The agreement between measured and theoretical angular probability distributions is good. The results herein indicate that the angular density probability of the low-pass-filtered calibration dataset is twice as narrow as that of the full velocity time series. This result can explain the failure of the polynomial fit to reconstruct the full velocity time series satisfactorily as resulting from the intrinsic property of this algorithm to ascribe a large weight to the highly concentrated points and a light weight to the thinly concentrated points while performing fitting. [ABSTRACT FROM AUTHOR]

Retrieval schemes often use two important components: 1) a radiative transfer model (RTM) inside the retrieval procedure or to construct the learning dataset for the training of the statistical retrieval algorithms and 2) a numerical weather prediction (NWP) model to provide a first guess or, again, to construct a learning dataset. This is particularly true in operational centers. As a consequence, any physical retrieval or similar method is limited by inaccuracies in the RTM and NWP models on which it is based. In this paper, a method for partially compensating for these errors as part of the sensor calibration is presented and evaluated. In general, RTM/NWP errors are minimized as best as possible prior to the training of the retrieval method, and then tolerated. The proposed method reduces these unknown and generally nonlinear residual errors by training a separate preprocessing neural network (NN) to produce calibrated radiances from real satellite data that approximate those radiances produced by the ''flawed'' NWP and RTM models. The final ''compensated/flawed'' retrieval assures better internal consistency of the retrieval procedure and then produces more accurate results. To the authors'' knowledge, this type of NN model has not been used yet for this purpose. The calibration approach is illustrated here on one particular application: the retrieval of atmospheric water vapor from the Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) and the Humidity Sounder for Brazil (HSB) measurements for nonprecipitating scenes, over land and ocean. Before being inverted, the real observations are ''projected'' into the space of the RTM simulation space from which the retrieval is designed. Validation of results is performed with radiosonde measurements and NWP analysis departures. This study shows that the NN calibration of the AMSR-E/HSB observations improves water vapor inversion, over ocean and land, for both clear and cloudy situations. The NN calibration is efficient and very general, being applicable to a large variety of problems. The nonlinearity of the NN allows for the calibration procedure to be state dependent and adaptable to specific cases (e.g., the same correction will not be applied to medium-range measurement and to extreme conditions). Its multivariate nature allows for a full exploitation of the complex correlation structure among the instrument channels, making the calibration of each single channel more robust. The procedure would make it possible to project the satellite observations in a reference observational space defined by radiosonde measurements, RTM simulations, or other instrument observational space. [ABSTRACT FROM AUTHOR]

Visible and infrared data obtained from instruments onboard geostationary satellites have been extensively used for monitoring clouds and their evolution. The Advanced Baseline Imager (ABI) that will be launched onboard the Geostationary Operational Environmental Satellite-R (GOES-R) series in the near future will offer a larger range of spectral bands; hence, it will provide observations of cloud and rain systems at even finer spatial, temporal, and spectral resolutions than are possible with the current GOES. In this paper, a new method called Precipitation Estimation from Remotely Sensed information using Artificial Neural Networks–Multispectral Analysis (PERSIANN-MSA) is proposed to evaluate the effect of using multispectral imagery on precipitation estimation. The proposed approach uses a self-organizing feature map (SOFM) to classify multidimensional input information, extracted from each grid box and corresponding textural features of multispectral bands. In addition, principal component analysis (PCA) is used to reduce the dimensionality to a few independent input features while preserving most of the variations of all input information. The above method is applied to estimate rainfall using multiple channels of the Spinning Enhanced Visible and Infrared Imager (SEVIRI) onboard the Meteosat Second Generation (MSG) satellite. In comparison to the use of a single thermal infrared channel, the analysis shows that using multispectral data has the potential to improve rain detection and estimation skills with an average of more than 50% gain in equitable threat score for rain/no-rain detection, and more than 20% gain in correlation coefficient associated with rain-rate estimation. [ABSTRACT FROM AUTHOR]

Reliable precipitation measurement is a crucial component in hydrologic studies. Although satellite-based observation is able to provide spatial and temporal distribution of precipitation, the measurements tend to show systematic bias. This paper introduces a grid-based precipitation merging procedure in which satellite estimates from the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks–Cloud Classification System (PERSIANN–CCS) are adjusted based on the Climate Prediction Center (CPC) daily rain gauge analysis. To remove the bias, the hourly CCS estimates were spatially and temporally accumulated to the daily 1° × 1° scale, the resolution of CPC rain gauge analysis. The daily CCS bias was then downscaled to the hourly temporal scale to correct hourly CCS estimates. The bias corrected CCS estimates are called the adjusted CCS (CCSA) product. With the adjustment from the gauge measurement, CCSA data have been generated to provide more reliable high temporal/spatial-resolution precipitation estimates. In the case study, the CCSA precipitation estimates from the proposed approach are compared against ground-based measurements in high-density gauge networks located in the southwestern United States. [ABSTRACT FROM AUTHOR]

Data from geosynchronous Earth-orbiting (GEO) satellites equipped with visible (VIS) and infrared (IR) scanners are commonly used in rain retrieval algorithms. These algorithms benefit from the high spatial and temporal resolution of GEO observations, either in stand-alone mode or in combination with higher-quality but less frequent microwave observations from low Earth-orbiting (LEO) satellites. In this paper, a neural network–based framework is presented to evaluate the utility of multispectral information in improving rain/no-rain (R/NR) detection. The algorithm uses the powerful classification features of the self-organizing feature map (SOFM), along with probability matching techniques to map single- or multispectral input space into R/NR maps. The framework was tested and validated using the 31 possible combinations of the five Geostationary Operational Environmental Satellite 12 (GOES-12) channels. An algorithm training and validation study was conducted over the conterminous United States during June–August 2006. The results indicate that during daytime, the visible channel (0.65 μm) can yield significant improvements in R/NR detection capabilities, especially when combined with any of the other four GOES-12 channels. Similarly, for nighttime detection the combination of two IR channels—particularly channels 3 (6.5 μm) and 4 (10.7 μm)—resulted in significant performance gain over any single IR channel. In both cases, however, using more than two channels resulted only in marginal improvements over two-channel combinations. Detailed examination of event-based images indicate that the proposed algorithm is capable of extracting information useful to screen no-rain pixels associated with cold, thin clouds and identifying rain areas under warm but rainy clouds. Both cases have been problematic areas for IR-only algorithms. [ABSTRACT FROM AUTHOR]

Sea ice thickness (SIT) is a key parameter of scientific interest because understanding the natural spatiotemporal variability of ice thickness is critical for improving global climate models. In this paper, changes in Arctic SIT during 1982–2003 are examined using a neural network (NN) algorithm trained with in situ submarine ice draft and surface drilling data. For each month of the study period, the NN individually estimated SIT of each ice-covered pixel (25-km resolution) based on seven geophysical parameters (four shortwave and longwave radiative fluxes, surface air temperature, ice drift velocity, and ice divergence/convergence) that were cumulatively summed at each monthly position along the pixel’s previous 3-yr drift track (or less if the ice was <3 yr old). Average January SIT increased during 1982–88 in most regions of the Arctic (+7.6 ± 0.9 cm yr-1), decreased through 1996 Arctic-wide (-6.1 ± 1.2 cm yr-1), then modestly increased through 2003 mostly in the central Arctic (+2.1 ± 0.6 cm yr-1). Net ice volume change in the Arctic Ocean from 1982 to 2003 was negligible, indicating that cumulative ice growth had largely replaced the estimated 45 000 km3 of ice lost by cumulative export. Above 65°N, total annual ice volume and interannual volume changes were correlated with the Arctic Oscillation (AO) at decadal and annual time scales, respectively. Late-summer ice thickness and total volume varied proportionally until the mid-1990s, but volume did not increase commensurate with the thickening during 1996–2002. The authors speculate that decoupling of the ice thickness–volume relationship resulted from two opposing mechanisms with different latitudinal expressions: a recent quasi-decadal shift in atmospheric circulation patterns associated with the AO’s neutral state facilitated ice thickening at high latitudes while anomalously warm thermal forcing thinned and melted the ice cap at its periphery. [ABSTRACT FROM AUTHOR]

This paper describes a neural network–based approach to estimate the conditional distribution function (cdf) of rainfall with respect to multidimensional satellite-derived input data. The methodology [Conditional Histogram of Precipitation (CHIP)] employs a histogram-based approximation of the cdf. In addition to the conditional expected rainfall rate, it provides conditional probabilities for that rate falling within each of a fixed set of intervals or bins. A test algorithm based on the CHIP approach was calibrated against Goddard profiling algorithm (GPROF) rainfall data for June–August 2002 and then used to produce a 30-min, 0.5° rainfall product from global (60°N–60°S) composite geostationary thermal infrared imagery for June–August 2003. Estimated rainfall rates and conditional probabilities were validated against 2003 GPROF data. The CHIP methodology provides the means to extend existing probabilistic and ensemble satellite rainfall error models, conditioned on a single, scalar, satellite rainfall predictor or upon scalar rainfall-rate outputs, to make full use of multidimensional input data. [ABSTRACT FROM AUTHOR]

The occurrence of fog can significantly impact air transport operations, and plays an important role in aviation safety. The economic value of aviation forecasts for Sydney Airport alone in 1993 was estimated at $6.8 million (Australian dollars) for Quantas Airways. The prediction of fog remains difficult despite improvements in numerical weather prediction guidance and models of the fog phenomenon. This paper assesses the ability of artificial neural networks (ANNs) to provide accurate forecasts of such events at Canberra International Airport (YSCB). Unlike conventional statistical techniques, ANNs are well suited to problems involving complex nonlinear interactions and therefore have potential in application to fog prediction. A 44-yr database of standard meteorological observations obtained from the Australian Bureau of Meteorology was used to develop, train, test, and validate ANNs designed to predict fog occurrence. Fog forecasting aids were developed for 3-, 6-, 12-, and 18-h lead times from 0600 local standard time. The forecasting skill of various ANN architectures was assessed through analysis of relative operating characteristic curves. Results indicate that ANNs are able to offer good discrimination ability at all four lead times. The results were robust to error perturbation for various input parameters. It is recommended that such models be included when preparing forecasts for YSCB, and that the technique should be extended in its application to cover other similarly fog-prone aviation locations. [ABSTRACT FROM AUTHOR]

Data assimilation in the field of predictive land surface modeling is generally limited to using observational data to estimate optimal model states or restrict model parameter ranges. To date, very little work has attempted to systematically define and quantify error resulting from a model's inherent inability to simulate the natural system. This paper introduces a data assimilation technique that moves toward this goal by accounting for those deficiencies in the model itself that lead to systematic errors in model output. This is done using a supervised artificial neural network to “learn” and simulate systematic trends in the model output error. These simulations in turn are used to correct the model's output each time step. The technique is applied in two case studies, using fluxes of latent heat flux at one site and net ecosystem exchange (NEE) of carbon dioxide at another. Root-mean-square error (rmse) in latent heat flux per time step was reduced from 27.5 to 18.6 W m-2 (32%) and monthly from 9.91 to 3.08 W m-2 (68%). For NEE, rmse per time step was reduced from 3.71 to 2.70 μmol m-2 s-1 (27%) and annually from 2.24 to 0.11 μmol m-2 s-1 (95%). In both cases the correction provided significantly greater gains than single criteria parameter estimation on the same flux. [ABSTRACT FROM AUTHOR]

In this paper several indices derived from atmospheric sounding data are used to develop a short-term thunderstorm forecast tool. The Udine (Italy) sounding (WMO code 16044, taken every 6 h, hereafter 6 h) is used to describe the initial conditions in which a thunderstorm may develop in the Friuli Venezia Giulia region. The tool forecasts the convective activity in the 6 h after the sounding is launched. Sounding, lightning, and mesonet station data, from April to September of the years 1995–2002, were used to train and validate artificial neural networks (ANNs) and the data of 2003 were used as an independent test sample. Special emphasis was given to avoid one of the major ANN problems, that of data overfitting, which requires limiting the possible complexity of the ANN, that is, the total number of inputs (predictors) and of hidden neurons. To select the best input set, the training–validation mechanism was repeated in eight different ways. Two types of ANNs were developed: the first is a classification model and is built for forecasting the thunderstorm occurrence, defined as the report of at least one cloud-to-ground lightning strike. If this first ANN forecasts convective activity, then a second ANN, built as a regression model, is used for forecasting the thunderstorm intensity. The relative operating characteristic (ROC) diagram and some parameters derived from a contingency table were used to compare the different classification performances. The linear cross-correlation coefficient, R2, was used to assess the performance of the regression ANNs. The ANN inputs were not the only “raw” sounding-derived index values; in addition, two derived sets of variables, corresponding to the z scores and the “posterior probability” transformed data, were used as preprocessing methods. The last method yields the best results. The tool reported on here has been used operationally since 2001 in the regional meteorological office of the Friuli Venezia Giulia region [Osservatorio Meteorologico Regionale/Agenzia Regionale per la Protezione dell’Ambiente (OSMER/ARPA)]. [ABSTRACT FROM AUTHOR]

Ensemble prediction has become an indispensable tool in weather forecasting. One of the issues in ensemble prediction is that, regardless of the method, the prediction error does not map well to the underlying physics (i.e., error estimates do not project strongly onto physical structures). This paper is driven by the hypothesis that prediction error includes a deterministic component, which can be isolated and then removed, and that removing the error would enable researchers and forecasters to better map the error to the physics and improve prediction of atmospheric transitions. Here, preliminary experimental evidence is provided that supports this hypothesis. This evidence is provided from results obtained from two low-order but highly chaotic systems, one of which incorporates atmospheric flow transitions. Using neural networks to probe the deterministic component of forecast error, it is shown that the error recovery relates to the underlying type of flow and that it can be used to better forecast transitions in the atmospheric flow using ensemble data. A discussion of methods to extend these ideas to more realistic forecast settings is provided. [ABSTRACT FROM AUTHOR]

Precipitation is arguably the most important meteorological forcing variable in land surface modeling. Many types of precipitation datasets exist (with various pros and cons) and include those from atmospheric data assimilation systems, satellites, rain gauges, ground radar, and merged products. These datasets are being evaluated in order to choose the most suitable precipitation forcing for real-time and retrospective simulations of the Global Land Data Assimilation System (GLDAS). This paper first presents results of a comparison for the period from March 2002 to February 2003. Later, GLDAS simulations 14 months in duration are analyzed to diagnose impacts on GLDAS land surface states when using the Mosaic land surface model (LSM). A comparison of seasonal total precipitation for the continental United States (CONUS) illustrates that the Climate Prediction Center (CPC) Merged Analysis of Precipitation (CMAP) has the closest agreement with a CPC rain gauge dataset for all seasons except winter. The European Centre for Medium-Range Weather Forecasts (ECMWF) model performs the best of the modeling systems. The satellite-only products [the Tropical Rainfall Measuring Mission (TRMM) Real-time Multi-satellite Precipitation Analysis and the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN)] suffer from a few deficiencies—most notably an overestimation of summertime precipitation in the central United States (200–400 mm). CMAP is the most closely correlated with daily rain gauge data for the spring, fall, and winter seasons, while the satellite-only estimates perform best in summer. GLDAS land surface states are sensitive to different precipitation forcing where percent differences in volumetric soil water content (SWC) between simulations ranged from -75% to +100%. The percent differences in SWC are generally 25%–75% less than the percent precipitation differences, indicating that GLDAS and specifically the Mosaic LSM act to generally “damp” precipitation differences. Areas where the percent changes are equivalent to the percent precipitation changes, however, are evident. Soil temperature spread between GLDAS runs was considerable and ranged up to ±3.0 K with the largest impact in the western United States. [ABSTRACT FROM AUTHOR]

Gated experts (GE) neural networks have been developed in order to retrieve atmospheric liquid water content over ocean from radiometer data. Gated experts neural networks are statistical models, which can model any general class of function. This paper focuses on the case where the complex transfer functions can be split on different simpler functions in order to improve the accuracy. Two atmospheric quantities are considered: the integrated cloud liquid water (iclw) and the surface rain rate (RR). In the case of iclw, the GE neural network finds two modes, splitting the problem into low and high iclw values. The physical meaning of those modes is discussed. A comparison with a standard regression algorithm and a multilayer perceptron neural network is done on simulated data and an “indirect comparison” is done using Special Sensor Microwave Imager (SSM/I) data. In the case of RR, the focus is on the ability of GE neural networks to perform a classification between rainy and nonrainy situations. Tropical Rainfall Measuring Mission (TRMM) data are used for rain-rate validation: rain-rate retrieval from the GE algorithm applied to actual TRMM Microwave Imager (TMI) measurements are compared with collocated precipitation radar (PR) rain rate. [ABSTRACT FROM AUTHOR]

Recent research has shown that neural network techniques can be used successfully for ground rainfall estimation from radar measurements. The neural network is a nonparametric method for representing the relationship between radar measurements and rainfall rate. The relationship is derived directly from a dataset consisting of radar measurements and rain gauge measurements. The effectiveness of the rainfall estimation by using neural networks can be influenced by many factors such as the representativeness and sufficiency of the training dataset, the generalization capability of the network to new data, season change, location change, and so on. In this paper, a novel scheme of adaptively updating the structure and parameters of the neural network for rainfall estimation is presented. This adaptive neural network scheme enables the network to account for any variability in the relationship between radar measurements and precipitation estimation and also to incorporate new information to the network without retraining the complete network from the beginning. This precipitation estimation scheme is a good compromise between the competing demands of accuracy and generalization. Data collected by a Weather Surveillance Radar—1988 Doppler (WSR-88D) and a rain gauge network were used to evaluate the performance of the adaptive network for rainfall estimation. It is shown that the adaptive network can estimate rainfall fairly accurately. The implementation of the adaptive network is very efficient and convenient for real-time rainfall estimation to be used with WSR-88D. [ABSTRACT FROM AUTHOR]

Radiance measurements from satellites offer the opportunity to retrieve atmospheric variables at much higher spatial resolution than is presently afforded by in situ measurements (e.g., radiosondes). However, the accuracy of these retrievals is crucial to their usefulness, and the ill-posed nature of the problem precludes a straightforward solution. A number of retrieval approaches have been investigated, including empirical techniques, coupling with numerical weather prediction models, and data analysis techniques such as regression. In this paper, artificial neural networks are used to retrieve vertical temperature and dewpoint profiles from infrared and microwave brightness temperatures from a polar-orbiting satellite. This approach allows retrievals to be performed even in cloudy conditions—a limitation of infrared-only retrievals. In a direct comparison of this technique with results from the operational Advanced Television and Infrared Observation Satellite Operational Vertical Sounder (ATOVS) retrievals, it was found that the neural-network temperature retrievals had larger errors than the ATOVS retrievals (though generally smaller than the first guess used in the ATOVS retrievals) but that the dewpoint retrievals showed consistent improvement over the comparable ATOVS retrievals. [ABSTRACT FROM AUTHOR]

This study investigated the uncertainty involved in statistical downscaling of hydroclimatic time series obtained by artificial neural networks (ANNs). Phase 6 of the Coupled Model Intercomparison Project (CMIP6) general circulation model (GCM) Canadian Earth System Model, version 5 (CanESM5), was used as large-scale predictor data for downscaling temperature and precipitation parameters. Two ANNs, feed forward and long short-term memory (LSTM), were utilized for statistical downscaling. To quantify the uncertainty of downscaling, prediction intervals (PIs) were estimated via the lower upper bound estimation (LUBE) method. To assess performance of proposed models in different climate regimes, data from the Tabriz and Rasht stations in Iran were employed. The calibrated models via historical GCM data were used for future projections via the high-forcing and fossil fuel–driven development scenario shared socioeconomic pathway (SSP) 5-8.5. Projections were compared with the Canadian Regional Climate Model 4 (Can-RCM4) projections via the same scenario. Results indicated that both LSTM-based point predictions and PIs are more accurate than the feedforward neural network (FFNN)-based predictions, with an average of 55% higher Nash–Sutcliffe efficiency (NSE) for point predictions and 25% lower coverage width criterion (CWC) for PIs. Projections suggested that Tabriz is going to experience a warmer climate with an increase in average temperature of 2° and 5°C for near and far futures, respectively, and a drier climate with a 20% decrease in precipitation until 2100. Future projections for the Rasht station, however, suggested a more uniform climate with less seasonal variability. Average precipitation will increase by up to 25% and 70% until near and far future periods, respectively. Ultimately, point predictions show that the average temperature in Rasht will increase by 1°C until the near future and then be a constant average temperature until the far future. Significance Statement: The downscaling of hydroclimatic parameters is subject to uncertainty. The best way is to provide an area with the highest contingency of them as a prediction interval. The reduction width of such an interval leads to increased confidence in explaining and predicting these processes. We proposed and applied a deep learning–based machine learning method for both point prediction and prediction interval estimation of temperature and precipitation parameters for the future over two different climatic regions. The results show the superiority of such a machine learning–based prediction interval estimation for quantification of the downscaling uncertainty. [ABSTRACT FROM AUTHOR]

Although the resolution of numerical weather prediction models continues to improve, many of the processes that influence precipitation are still not captured adequately by the scales of present operational models, and consequently precipitation forecasts have not yet reached the level of accuracy needed for hydrologic forecasting. Postprocessing of model output to account for local differences can enhance the accuracy and usefulness of these forecasts. Model Output Statistics have performed this important function for a number of years via regression techniques; this paper presents an alternate approach that uses artificial neural networks to produce 6-h precipitation forecasts for specific locations. Tests performed on four locations in the middle Atlantic region of the United States show that the accuracy of the forecasts produced using neural networks compares favorably with those generated using linear regression, especially for heavier precipitation amounts. [ABSTRACT FROM AUTHOR]

Accurate, timely, site-specific forecasts of precipitation are important for accurately predicting streamflow and flash floods in small drainage basins. However, presently available numerical weather prediction models do not generally provide forecasts with the accuracy and/or resolution appropriate for this task. A wide variety of approaches to small-scale, short-term precipitation forecasting have been investigated by numerous authors; this paper describes a simple precipitation forecasting model based on artificial neural networks. The model uses the radiosonde-based 700-hPa wind direction and antecedent precipitation data from a rain gauge network to generate short-term (0--6 h) precipitation forecasts for a target location. The performance of the model is illustrated for a gauge in eastern Pennsylvania. [ABSTRACT FROM AUTHOR]

Forecast skill from dynamical forecast models decreases quickly with projection time due to various errors. Therefore, postprocessing methods, from simple bias correction methods to more complicated multiple linear regression-based model output statistics, are used to improve raw model forecasts. Usually, these methods show clear forecast improvement over the raw model forecasts, especially for short-range weather forecasts. However, linear approaches have limitations because the relationship between predictands and predictors may be nonlinear. This is even truer for extended range forecasts, such as week-3-4 forecasts. In this study, neural network techniques are used to seek or model the relationships between a set of predictors and predictands, and eventually to improve week-3-4 precipitation and 2-m temperature forecasts made by the NOAA/NCEP Climate Forecast System. Benefitting from advances in machine learning techniques in recent years, more flexible and capable machine learning algorithms and availability of big datasets enable us not only to explore nonlinear features or relationships within a given large dataset, but also to extract more sophisticated pattern relationships and covariabilities hidden within the multidimensional predictors and predictands. Then these more sophisticated relationships and high-level statistical information are used to correct the model week-3-4 precipitation and 2-m temperature forecasts. The results show that to some extent neural network techniques can significantly improve the week-3-4 forecast accuracy and greatly increase the efficiency over the traditional multiple linear regression methods. [ABSTRACT FROM AUTHOR]

We evaluate the short-term weather forecast performance of three flavors of artificial neural networks (NNs): feed forward back propagation, radial basis function, and generalized regression. To prepare the application of the NNs to an operational setting, we tune NN hyperparameters using over two years of historical data. Five objective guidance products serve as predictors to the NNs: North American Mesoscale and Global Forecast System model output statistics (MOS) forecasts, the High-Resolution Rapid Refresh (HRRR) model, National Weather Service forecasts, and the National Blend of Models product. We independently test NN performance using 96 real-time forecasts of temperature, wind, and precipitation across 11 U.S. cities made during the WxChallenge, a weather forecasting competition. We demonstrate that all NNs significantly improve short-range weather forecasts relative to the traditional objective guidance aids used to train the networks. For example, 1-day maximum and minimum temperature forecast error is 20%–30% lower than MOS. However, NN improvement over multiple linear regression for short-term forecasts is not significant. We suggest this may be attributed to the small number of training samples, the operational nature of the experiment, and the short forecast lead times. Regardless, our results are consistent with previous work suggesting that applying NNs to model forecasts can have a positive impact on operational forecast skill and will become valuable tools when integrated into the forecast enterprise. Significance Statement: We used approximately two years of historical weather data and objective forecasts for a number of cities to tune a series of artificial neural networks (NNs) to forecast 1-day values of maximum and minimum temperature, maximum sustained wind speed, and quantitative precipitation. We compare forecast error against common objective guidance and multiple linear regression. We found that the NNs exhibit about 25% lower error than common objective guidance for temperature forecasting and 50% lower error for wind speed. Our results suggest that NNs will be a valuable contributor to improving weather forecast skill when adopted into the existing forecast enterprise. [ABSTRACT FROM AUTHOR]

The evaluation of the representation of the surface downward shortwave flux (DSF) from atmospheric reanalysis data products is required to obtain reliable information for the resource assessment of surface solar energy. The representation of the DSF from the Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2), reanalysis data product was evaluated using surface solar radiation from ground-based observations in Japan. The cloud fraction (CFR) and cloud optical thickness (COT) from Moderate Resolution Imaging Spectroradiometer (MODIS) were also used as references. The CFR from MERRA-2 tends to be smaller than that from MODIS, and the correlation between the difference in the CFR and that in the DSF is negative. The correlation between the difference in the COT and that in the DSF is weakly negative. To quantify the effects of the difference in the CFR and COT to that of the DSF, a regression model based on an artificial neural network architecture that emulates the process of the DSF in MERRA-2 was constructed. Numerical experiments using the emulator quantify contributions of each of the differences in the CFR and COT and joint contributions of the two variables. In addition, a cluster analysis was performed to clarify the differences in the seasonal changes in the monthly mean bias error (MBE) in the DSF among ground observation stations, and three clusters were identified. Contributions of the differences in the CFR and COT to the seasonal change in the monthly MBE were also clarified using the results of the numerical experiments. [ABSTRACT FROM AUTHOR]

MODIS (Moderate Resolution Imaging Spectroradiometer) land surface temperature measurements in combination with in situ air temperature records from 119 meteorological stations are used to reconstruct a monthly near-surface air temperature product over the Antarctic Ice Sheet (AIS) by means of a neural network model. The product is generated on a regular grid of 0.05° 3 0.05°, spanning from 2001 to 2018. Comparison with independent in situ air temperature measurements shows low uncertainty, with a mean bias of 0.098C, a mean absolute error of 2.23°C, and a correlation coefficient of 97%. Furthermore, the performance of the reconstruction is better than ERA5 (the fifth-generation ECMWF reanalysis model) against in situ measurements. For the 2001-18 period, the MODIS-based near-surface air temperature product yields annual warming in the East Antarctica, but cooling in the Antarctic Peninsula and West Antarctica. However, they are not statistically significant. This product can also be used to investigate the impact of the Southern Hemisphere annual mode (SAM) on year-to-year variability of air temperature. The enhanced positive phase of SAMin recent decades in austral summer has a cooling effect on East andWest Antarctica. In addition, the dataset has the potential application for climate model validation and data assimilation due to the independence of the input of a numerical weather prediction model. [ABSTRACT FROM AUTHOR]

Wind gusts, and in particular intense gusts, are societally relevant but extremely challenging to forecast. This study systematically assesses the skill enhancement that can be achieved using artificial neural networks (ANNs) for forecasting of wind gust occurrence and magnitude. Geophysical predictors from the ERA5 reanalysis are used in conjunction with an autoregressive term in regression and ANN models with different predictors, and varying model complexity. Models are derived and assessed for the warm (April–September) and cold (October–March) seasons for three high passenger volume airports in the United States. Model uncertainty is assessed by deriving models for 1000 different randomly selected training (70%) and testing (30%) subsets. Gust prediction fidelity in independent test samples is critically dependent on inclusion of an autoregressive term. Gust occurrence probabilities derived using five-layer ANNs exhibit consistently higher fidelity than those from regression models and shallower ANNs. Inclusion of the autoregressive term and increasing the number of hidden layers in ANNs from 1 to 5 also improve the model performance for gust magnitudes (lower RMSE, increased correlation, and model standard deviations that more closely approximate observed values). Deeper ANNs (e.g., 20 hidden layers) exhibit higher skill in forecasting strong (17–25.7 m s−1) and damaging (≥25.7 m s−1) wind gusts. However, such deep networks exhibit evidence of overfitting and still substantially underestimate (by 50%) the frequency of strong and damaging wind gusts at the three airports considered herein. Significance Statement: Improved short-term forecasting of wind gusts will enhance aviation safety and logistics and may offer other societal benefits. Here we present a rigorous investigation of the relative skill of models of wind gust occurrence and magnitude that employ different statistical methods. It is shown that artificial neural networks (ANNs) offer considerable skill enhancement over regression methods, particularly for strong and damaging wind gusts. For wind gust magnitudes in particular, application of deeper learning networks (e.g., five or more hidden layers) offers tangible improvements in forecast accuracy. However, deeper networks are vulnerable to overfitting and exhibit substantial variability with the specific training and testing data subset used. Also, even deep ANNs reproduce only half of strong and damaging wind gusts. These results indicate the need for future work to elucidate the dynamical mechanisms of intense wind gusts and advance solutions to their prediction. [ABSTRACT FROM AUTHOR]

A climate model emulator is developed using neural network techniques and trained with the data from the multithousand-member climateprediction.net perturbed physics GCM ensemble. The method recreates nonlinear interactions between model parameters, allowing a simulation of a much larger ensemble that explores model parameter space more fully. The emulated ensemble is used to search for models closest to observations over a wide range of equilibrium response to greenhouse gas forcing. The relative discrepancies of these models from observations could be used to provide a constraint on climate sensitivity. The use of annual mean or seasonal differences on top-of-atmosphere radiative fluxes as an observational error metric results in the most clearly defined minimum in error as a function of sensitivity, with consistent but less well-defined results when using the seasonal cycles of surface temperature or total precipitation. The model parameter changes necessary to achieve different values of climate sensitivity while minimizing discrepancy from observation are also considered and compared with previous studies. This information is used to propose more efficient parameter sampling strategies for future ensembles. [ABSTRACT FROM AUTHOR]

We introduce an adaptive form of postprocessor where algorithm structures are neural networks where the number of hidden nodes and the network training features evolve. Key potential advantages of this system are the flexible, nonlinear mapping capabilities of neural networks and, through backpropagation, the ability to rapidly establish capable predictors in an algorithm population. The system can be implemented after one initial training process and future changes to postprocessor inputs (new observations, new inputs, or model upgrades) are incorporated as they become available. As in prior work, the implementation in the form of a predator–prey ecosystem allows for the ready construction of ensembles. Computational requirements are minimal, and the use of a moving data window means that data storage requirements are constrained. The system adds predictive skill to a demonstration dynamical model representing the hemispheric circulation, with skill competitive with or exceeding that obtainable from multiple linear regression and standard artificial neural networks constructed under typical operational limitations. The system incorporates new information rapidly and the dependence of the approach on the training data size is similar to multiple linear regression. A loss of performance occurs relative to a fixed neural network architecture in which only the weights are adjusted after training, but this loss is compensated for by gains from the ensemble predictions. While the demonstration dynamical model is complex, current numerical weather prediction models are considerably more so, and thus a future step will be to apply this technique to operational weather forecast data. [ABSTRACT FROM AUTHOR]

The Global Precipitation Measurement (GPM) mission provides satellite precipitation products with an unprecedented spatiotemporal resolution and spatial coverage. However, its near-real-time (NRT) product still suffers from low accuracy. This study aims to improve the early run of the Integrated Multisatellite Retrievals for GPM (IMERG) by using four machine learning approaches, that is, support vector machine (SVM), random forest (RF), artificial neural network (ANN), and extreme gradient boosting (XGB). The cloud properties are selected as the predictors in addition to the original IMERG in these approaches. All of the four approaches show similar improvement, with 53%–60% reduction of root-mean-square error (RMSE) compared with the original IMERG in a humid area, that is, the Dongjiang River basin (DJR) in southeastern China. The improvements are even greater in a semiarid area, that is, the Fenhe River basin (FHR) in central China, where the RMSE reduction ranges from 63% to 66%. The products generated by the machine learning methods perform similarly to or even outperform the final run of IMERG. Feature importance analysis, a technique to evaluate input features based on how useful they are in predicting a target variable, indicates that the cloud height and the brightness temperature are the most useful information in improving satellite precipitation products, followed by the atmospheric reflectivity and the surface temperature. This study shows that a more accurate NRT precipitation product can be produced by combining machine learning approaches and cloud information, which is of importance for hydrological applications that require NRT precipitation information including flood monitoring. [ABSTRACT FROM AUTHOR]

ABSTRACT A procedure to enhance neural network (NN) predictions of tropical Pacific sea surface temperature anomalies and calculating their estimated errors is presented. A simple linear correction enables more accurate predictions of warm and cold events but can result in introduction of larger errors in other cases. The prediction error estimates aid recognizing erroneously magnified anomalies and are used to sort the predictions into El Nino, La Nina, and neutral states. The error estimation process is based on bootstrap resamplings of the data and construction of a large number of bootstrap prediction replicas. A statistic calculated on the set of bootstrap replicas that corresponds to each of the actual predictions is used to estimate the prediction's errors. The method is demonstrated on NN prediction of the Nino-3.4 index. [ABSTRACT FROM AUTHOR]