Search Results
for using model-derived fields are proposed. These NNs will be used to explore and evaluate their capability to improve the week-3–4 precipitation and 2-m air temperature forecasts. The rest of this paper is organized as follows: the dataset used for the NN training/testing and detailed NN methodology used in this study is highlighted in section 2 . The NN check, optimal hidden neurons, data representation, and analysis of the week-3–4 model forecast errors are described in section 3 . The NN
for using model-derived fields are proposed. These NNs will be used to explore and evaluate their capability to improve the week-3–4 precipitation and 2-m air temperature forecasts. The rest of this paper is organized as follows: the dataset used for the NN training/testing and detailed NN methodology used in this study is highlighted in section 2 . The NN check, optimal hidden neurons, data representation, and analysis of the week-3–4 model forecast errors are described in section 3 . The NN
global NWP models ( Bauer et al. 2015 ). The representation of these processes is especially challenging over continental areas from the Southern Hemisphere where the abundant vegetation and the sparse observations for evaluation and data assimilation have limited the models’ accuracy. Recent progress in forecasting tropical convection ( Bechtold et al. 2014 ; Subramanian et al. 2017 ) and the increasing quantity and quality of global information encourage the use of NWP for tropical precipitation
global NWP models ( Bauer et al. 2015 ). The representation of these processes is especially challenging over continental areas from the Southern Hemisphere where the abundant vegetation and the sparse observations for evaluation and data assimilation have limited the models’ accuracy. Recent progress in forecasting tropical convection ( Bechtold et al. 2014 ; Subramanian et al. 2017 ) and the increasing quantity and quality of global information encourage the use of NWP for tropical precipitation
are 1D with lower spatial resolution), which would present a major difficulty for non-ML-based postprocessing methods such as SSPF. The rest of this paper is organized as follows. Section 2 briefly describes the inner workings of CNNs [a more thorough description is provided in Lagerquist et al. (2019) , hereafter L19 ], section 3 describes the input data and preprocessing, section 4 describes experiments used to find the best CNNs, section 5 evaluates performance of the best CNNs, and
are 1D with lower spatial resolution), which would present a major difficulty for non-ML-based postprocessing methods such as SSPF. The rest of this paper is organized as follows. Section 2 briefly describes the inner workings of CNNs [a more thorough description is provided in Lagerquist et al. (2019) , hereafter L19 ], section 3 describes the input data and preprocessing, section 4 describes experiments used to find the best CNNs, section 5 evaluates performance of the best CNNs, and
southeastern United States where the snow has little influence on runoff. The LSTM performance is evaluated in comparison with a physics-based hydrologic model calibrated using the same training data as in the LSTM. Results from this study should have important implications for streamflow simulation in rural watersheds where data quality and availability are a critical issue. The paper is organized as follows. Section 2 describes the LSTM network, various regularization techniques, Bayesian LSTM and the
southeastern United States where the snow has little influence on runoff. The LSTM performance is evaluated in comparison with a physics-based hydrologic model calibrated using the same training data as in the LSTM. Results from this study should have important implications for streamflow simulation in rural watersheds where data quality and availability are a critical issue. The paper is organized as follows. Section 2 describes the LSTM network, various regularization techniques, Bayesian LSTM and the
the model learns to fit the training data. The paper is organized as follows. In section 2 , the U-Net architecture is described, as are the numerical metrics used to evaluate its success. Section 3 describes both qualitatively and quantitatively the design and performance of the best U-Net model obtained for identifying tropical cyclones using the Global Forecast System (GFS) total precipitable water field as inputs. In section 4 , three additional U-Net models are introduced that identify
the model learns to fit the training data. The paper is organized as follows. In section 2 , the U-Net architecture is described, as are the numerical metrics used to evaluate its success. Section 3 describes both qualitatively and quantitatively the design and performance of the best U-Net model obtained for identifying tropical cyclones using the Global Forecast System (GFS) total precipitable water field as inputs. In section 4 , three additional U-Net models are introduced that identify
used through an unbiased method that trains a representative model learning universal hydrologic behaviors across scales. For this purpose, a training strategy is designed that applies multifold cross validation across multiple independent subsets of training, testing, and validation datasets. Reproducibility across different basins is monitored through various regression evaluation metrics that quantify the model performances. Note that the predictors are not transformed because tree
used through an unbiased method that trains a representative model learning universal hydrologic behaviors across scales. For this purpose, a training strategy is designed that applies multifold cross validation across multiple independent subsets of training, testing, and validation datasets. Reproducibility across different basins is monitored through various regression evaluation metrics that quantify the model performances. Note that the predictors are not transformed because tree
, by trying different sets of hyperparameters, training a complete model for each set, evaluating the resulting model, and then deciding which hyperparameter set results in best performance. Algorithms range from simple exhaustive grid search (as illustrated in the “Using performance measures for NN tuning” section) to sophisticated algorithms ( Kasim et al. 2020 ; Hertel et al. 2020 ). Sample application: Image-to-image translation from GOES to MRMS. We demonstrate many of the concepts in this
, by trying different sets of hyperparameters, training a complete model for each set, evaluating the resulting model, and then deciding which hyperparameter set results in best performance. Algorithms range from simple exhaustive grid search (as illustrated in the “Using performance measures for NN tuning” section) to sophisticated algorithms ( Kasim et al. 2020 ; Hertel et al. 2020 ). Sample application: Image-to-image translation from GOES to MRMS. We demonstrate many of the concepts in this
. However, it is difficult to generalize this difference because of the small sample size for category 5. Overall, the improvement is enough to justify limiting the remaining model evaluation to only the two-channel version of DeepMicroNet going forward. Fig . 5. (a) Intensity error (RMSE) according to best track MSW for the three model versions labeled in the legend, and (b) average standard deviation of the PDFs according to best track MSW. b. Model performance The following describes a two
. However, it is difficult to generalize this difference because of the small sample size for category 5. Overall, the improvement is enough to justify limiting the remaining model evaluation to only the two-channel version of DeepMicroNet going forward. Fig . 5. (a) Intensity error (RMSE) according to best track MSW for the three model versions labeled in the legend, and (b) average standard deviation of the PDFs according to best track MSW. b. Model performance The following describes a two
trained model, which is useful in selecting hyperparameters (see section 2d ). However, by choosing hyperparameter values that optimize performance on the validation set, the hyperparameters can be overfit to the validation set, just like model weights (those adjusted by training) can be overfit to the training set. Thus, the selected model is also evaluated on the testing set, which is independent of the data used to fit both the model weights and hyperparameters. c. Model architecture CNNs use a
trained model, which is useful in selecting hyperparameters (see section 2d ). However, by choosing hyperparameter values that optimize performance on the validation set, the hyperparameters can be overfit to the validation set, just like model weights (those adjusted by training) can be overfit to the training set. Thus, the selected model is also evaluated on the testing set, which is independent of the data used to fit both the model weights and hyperparameters. c. Model architecture CNNs use a
training set and an evaluation set (called the validation set) that ensures the absence of overfitting to the training set. When ANNs are specifically used, multiple network sizes and/or architectures are typically experimented with, so that two evaluation sets (validation and test) are required. The test set is used only with the final “best” trained network to confirm similar performance as the validation set. This avoids overfitting to the validation set itself from any possible hidden preferential
training set and an evaluation set (called the validation set) that ensures the absence of overfitting to the training set. When ANNs are specifically used, multiple network sizes and/or architectures are typically experimented with, so that two evaluation sets (validation and test) are required. The test set is used only with the final “best” trained network to confirm similar performance as the validation set. This avoids overfitting to the validation set itself from any possible hidden preferential
