Machine Learning(ML) models have emerged as a powerful tool in many fields, including climate modeling. In climate modeling, ML is often used to analyze climate data, satellite imagery and other environmental data and are compared to more traditional analysis methods and proved its ability to analyze vast amounts of data and identify complex patterns. Convolutional Neural Network(CNN), Recurrent Neural Network(RNN). Extreme have become more frequent and severe in recent years. In this study, we concentrate on the extreme precipitation events. These events can cause severe flooding and other hazards, leading to significant social, economic, and environmental impacts. As these events become more frequent and intense due to climate change, there is an urgent need to develop more accurate and timely methods for predicting and managing their impacts. In this study, we completed two main experiments to understand if ML algorithm can detect the extreme events. In both experiment the predictors that have been used are eastern and northern wind, geopotential height, specific humidity, and temperature at four pressure levels, which are 1000hpa, 850hpa, 700hpa, and 500hpa. The frequency for the predictors is 3hours, while the predictand being the precipitation accumulated over 3hours. In the first experiment two main models, CNN and LSTM, have been used. The data used in this part of this study are the Re-Analysis -5th generation-(ERA5) with a resolution of 25 km at different pressure levels and for the surface (precipitation in our case). The spatial domain used was the window between 2 to 20 longitude degree North and 36 to 52 latitude degree East. During the validation both models showed the tendency to overestimate the frequency and underestimate the intensity of the precipitation events. Although they are able to detect the most active regions of the high precipitation convection events, they underestimate the high quantile (99th percentile) of the distribution. The main difference between the two models is that CNN model is more efficient for the prediction of extreme events than the LSTM model. However, LSTM model produces more accurate results for the events less than 1mm/day that are classified as dry days. The second experiment in this study is the precipitation downscaling by mean of ML algorithms. The target domain is the Alpine region within the window of 0 to 22 longitude degree East and 35 to 55 latitude degree North. The predictors are the same ERA5 variables from the first experiment at 25km resolution, and the predictand being the observed rainfall at 3km from the Gridded Italian Precipitation Hourly Observations dataset(GRIPHO). The ML model used in this experiment is a combination of convolutional layers, LSTM layers and fully connected layers. The idea of the model is to take advantages of using LSTM layers for memorizing the long and short events, while convolutional layers for feature extraction and classification. The model will handle every variable along the pressure levels individually. This experiment referred as ERA5-GRIPHO trial, shows more intense precipitation than the GRIPHO, and more frequent wet events, and underestimates the tails of the precipitation distribution over the whole domain, but the model capture the seasonal distribution of precipitation and correctly identify the convective regions. underestimate the tails of the precipitation distribution over the whole domains. The last experiment was to use the above ERA5-GRIPHO trial methodology by replacing the ERA5 reanalysis predictors with the Regional Climate Model(RegCM) predictors at 3km resolution, upscaled to 25km resolution and the observed precipitation dataset with the modelled one at 3km resolution. The reason being that we wanted to test if the ML algorithm can perform better in a simpler problem that is to "learn" the downscaling function typical of each Regional Climate Model.
Machine Learning(ML) models have emerged as a powerful tool in many fields, including climate modeling. In climate modeling, ML is often used to analyze climate data, satellite imagery and other environmental data and are compared to more traditional analysis methods and proved its ability to analyze vast amounts of data and identify complex patterns. Convolutional Neural Network(CNN), Recurrent Neural Network(RNN). Extreme have become more frequent and severe in recent years. In this study, we concentrate on the extreme precipitation events. These events can cause severe flooding and other hazards, leading to significant social, economic, and environmental impacts. As these events become more frequent and intense due to climate change, there is an urgent need to develop more accurate and timely methods for predicting and managing their impacts. In this study, we completed two main experiments to understand if ML algorithm can detect the extreme events. In both experiment the predictors that have been used are eastern and northern wind, geopotential height, specific humidity, and temperature at four pressure levels, which are 1000hpa, 850hpa, 700hpa, and 500hpa. The frequency for the predictors is 3hours, while the predictand being the precipitation accumulated over 3hours. In the first experiment two main models, CNN and LSTM, have been used. The data used in this part of this study are the Re-Analysis -5th generation-(ERA5) with a resolution of 25 km at different pressure levels and for the surface (precipitation in our case). The spatial domain used was the window between 2 to 20 longitude degree North and 36 to 52 latitude degree East. During the validation both models showed the tendency to overestimate the frequency and underestimate the intensity of the precipitation events. Although they are able to detect the most active regions of the high precipitation convection events, they underestimate the high quantile (99th percentile) of the distribution. The main difference between the two models is that CNN model is more efficient for the prediction of extreme events than the LSTM model. However, LSTM model produces more accurate results for the events less than 1mm/day that are classified as dry days. The second experiment in this study is the precipitation downscaling by mean of ML algorithms. The target domain is the Alpine region within the window of 0 to 22 longitude degree East and 35 to 55 latitude degree North. The predictors are the same ERA5 variables from the first experiment at 25km resolution, and the predictand being the observed rainfall at 3km from the Gridded Italian Precipitation Hourly Observations dataset(GRIPHO). The ML model used in this experiment is a combination of convolutional layers, LSTM layers and fully connected layers. The idea of the model is to take advantages of using LSTM layers for memorizing the long and short events, while convolutional layers for feature extraction and classification. The model will handle every variable along the pressure levels individually. This experiment referred as ERA5-GRIPHO trial, shows more intense precipitation than the GRIPHO, and more frequent wet events, and underestimates the tails of the precipitation distribution over the whole domain, but the model capture the seasonal distribution of precipitation and correctly identify the convective regions. underestimate the tails of the precipitation distribution over the whole domains. The last experiment was to use the above ERA5-GRIPHO trial methodology by replacing the ERA5 reanalysis predictors with the Regional Climate Model(RegCM) predictors at 3km resolution, upscaled to 25km resolution and the observed precipitation dataset with the modelled one at 3km resolution. The reason being that we wanted to test if the ML algorithm can perform better in a simpler problem that is to "learn" the downscaling function typical of each Regional Climate Model.
Detection of High Convective Precipitation Events Using Advanced Machine Learning Methods / Abed, WAED H. A.. - (2024 Mar 21).
Detection of High Convective Precipitation Events Using Advanced Machine Learning Methods
ABED, WAED H.A.
2024-03-21
Abstract
Machine Learning(ML) models have emerged as a powerful tool in many fields, including climate modeling. In climate modeling, ML is often used to analyze climate data, satellite imagery and other environmental data and are compared to more traditional analysis methods and proved its ability to analyze vast amounts of data and identify complex patterns. Convolutional Neural Network(CNN), Recurrent Neural Network(RNN). Extreme have become more frequent and severe in recent years. In this study, we concentrate on the extreme precipitation events. These events can cause severe flooding and other hazards, leading to significant social, economic, and environmental impacts. As these events become more frequent and intense due to climate change, there is an urgent need to develop more accurate and timely methods for predicting and managing their impacts. In this study, we completed two main experiments to understand if ML algorithm can detect the extreme events. In both experiment the predictors that have been used are eastern and northern wind, geopotential height, specific humidity, and temperature at four pressure levels, which are 1000hpa, 850hpa, 700hpa, and 500hpa. The frequency for the predictors is 3hours, while the predictand being the precipitation accumulated over 3hours. In the first experiment two main models, CNN and LSTM, have been used. The data used in this part of this study are the Re-Analysis -5th generation-(ERA5) with a resolution of 25 km at different pressure levels and for the surface (precipitation in our case). The spatial domain used was the window between 2 to 20 longitude degree North and 36 to 52 latitude degree East. During the validation both models showed the tendency to overestimate the frequency and underestimate the intensity of the precipitation events. Although they are able to detect the most active regions of the high precipitation convection events, they underestimate the high quantile (99th percentile) of the distribution. The main difference between the two models is that CNN model is more efficient for the prediction of extreme events than the LSTM model. However, LSTM model produces more accurate results for the events less than 1mm/day that are classified as dry days. The second experiment in this study is the precipitation downscaling by mean of ML algorithms. The target domain is the Alpine region within the window of 0 to 22 longitude degree East and 35 to 55 latitude degree North. The predictors are the same ERA5 variables from the first experiment at 25km resolution, and the predictand being the observed rainfall at 3km from the Gridded Italian Precipitation Hourly Observations dataset(GRIPHO). The ML model used in this experiment is a combination of convolutional layers, LSTM layers and fully connected layers. The idea of the model is to take advantages of using LSTM layers for memorizing the long and short events, while convolutional layers for feature extraction and classification. The model will handle every variable along the pressure levels individually. This experiment referred as ERA5-GRIPHO trial, shows more intense precipitation than the GRIPHO, and more frequent wet events, and underestimates the tails of the precipitation distribution over the whole domain, but the model capture the seasonal distribution of precipitation and correctly identify the convective regions. underestimate the tails of the precipitation distribution over the whole domains. The last experiment was to use the above ERA5-GRIPHO trial methodology by replacing the ERA5 reanalysis predictors with the Regional Climate Model(RegCM) predictors at 3km resolution, upscaled to 25km resolution and the observed precipitation dataset with the modelled one at 3km resolution. The reason being that we wanted to test if the ML algorithm can perform better in a simpler problem that is to "learn" the downscaling function typical of each Regional Climate Model.File | Dimensione | Formato | |
---|---|---|---|
thesis_wbed_final.pdf
embargo fino al 21/03/2025
Descrizione: Detection of High Convective Precipitation Events Using Advanced Machine Learning Methods
Tipologia:
Tesi di dottorato
Dimensione
33.73 MB
Formato
Adobe PDF
|
33.73 MB | Adobe PDF | Visualizza/Apri Richiedi una copia |
Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.