Wind speed and its evolution over time and space are important for many purposes, particularly for climate studies. However, direct measurement at the ocean surface is problematic, and remote prediction using satellite and radar systems rarely offers the desired resolution or accuracy, especially in polar waters. Being able to predict wind speed from underwater acoustic noise is, therefore, a key aspect to improving climate models and monitoring systems. To date, empirical equations using the noise spectrum at a given frequency, fitted to the available data, have been adopted. To leverage on the whole noise spectrum, only recently supervised machine-learning regression has been proposed. This article exploits the concurrent acquisition of underwater noise and anemometer measurements over a period of 16 months, at 10-min steps, to demonstrate the superiority of regression models based on supervised learning over empirical equations. It is also shown that depending on the type of compounding implemented, different tradeoffs are achieved between accuracy and temporal resolution. Considering all samples in the data set, including those in which rainfall and passing ships occurred, and separating by cross-validation the samples used for training from those used for testing, it is shown that a regressor based on the random forest technique, followed by the compounding of the predictions over an interval of 1 h, provides a mean absolute error of 0.62 m/s and a correlation coefficient of 0.96. In addition, the robustness with respect to the parameter setting and the performance obtained in performing prediction in chronological order are evaluated.

Compounding Approaches for Wind Prediction From Underwater Noise by Supervised Learning

Andrea Trucco;Annalisa Barla;Emanuele Fava;Alessandro Verri;David Solarna
2022

Abstract

Wind speed and its evolution over time and space are important for many purposes, particularly for climate studies. However, direct measurement at the ocean surface is problematic, and remote prediction using satellite and radar systems rarely offers the desired resolution or accuracy, especially in polar waters. Being able to predict wind speed from underwater acoustic noise is, therefore, a key aspect to improving climate models and monitoring systems. To date, empirical equations using the noise spectrum at a given frequency, fitted to the available data, have been adopted. To leverage on the whole noise spectrum, only recently supervised machine-learning regression has been proposed. This article exploits the concurrent acquisition of underwater noise and anemometer measurements over a period of 16 months, at 10-min steps, to demonstrate the superiority of regression models based on supervised learning over empirical equations. It is also shown that depending on the type of compounding implemented, different tradeoffs are achieved between accuracy and temporal resolution. Considering all samples in the data set, including those in which rainfall and passing ships occurred, and separating by cross-validation the samples used for training from those used for testing, it is shown that a regressor based on the random forest technique, followed by the compounding of the predictions over an interval of 1 h, provides a mean absolute error of 0.62 m/s and a correlation coefficient of 0.96. In addition, the robustness with respect to the parameter setting and the performance obtained in performing prediction in chronological order are evaluated.
File in questo prodotto:
File Dimensione Formato  
Compounding_Approaches_for_Wind_Prediction_From_Underwater_Noise_by_Supervised_Learning.pdf

accesso chiuso

Descrizione: Articolo su rivista
Tipologia: Documento in versione editoriale
Dimensione 4.89 MB
Formato Adobe PDF
4.89 MB Adobe PDF   Visualizza/Apri   Richiedi una copia

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11567/1100636
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? 0
social impact