Structured neural network modelling of multi-valued functions for wind retrieval from scatterometer measurements

David J. Evans, Dan Cornford, Ian T. Nabney

Research output: Working paperTechnical report

Abstract

A conventional neural network approach to regression problems approximates the conditional mean of the output vector. For mappings which are multi-valued this approach breaks down, since the average of two solutions is not necessarily a valid solution. In this article mixture density networks, a principled method to model conditional probability density functions, are applied to retrieving Cartesian wind vector components from satellite scatterometer data. A hybrid mixture density network is implemented to incorporate prior knowledge of the predominantly bimodal function branches. An advantage of a fully probabilistic model is that more sophisticated and principled methods can be used to resolve ambiguities.
Original languageEnglish
Place of PublicationBirmingham
PublisherAston University
Number of pages8
ISBN (Print)NCRG/98/022
Publication statusPublished - 22 Oct 1998

Fingerprint

scatterometer
probability density function
modeling
method

Keywords

  • wind vector retrieval
  • ERS-1 satellite
  • probabilistic models
  • mixture density networks
  • neural networks

Cite this

Evans, D. J., Cornford, D., & Nabney, I. T. (1998). Structured neural network modelling of multi-valued functions for wind retrieval from scatterometer measurements. Birmingham: Aston University.
Evans, David J. ; Cornford, Dan ; Nabney, Ian T. / Structured neural network modelling of multi-valued functions for wind retrieval from scatterometer measurements. Birmingham : Aston University, 1998.
@techreport{838cb4bde32d40799ef2af2d58bd8f38,
title = "Structured neural network modelling of multi-valued functions for wind retrieval from scatterometer measurements",
abstract = "A conventional neural network approach to regression problems approximates the conditional mean of the output vector. For mappings which are multi-valued this approach breaks down, since the average of two solutions is not necessarily a valid solution. In this article mixture density networks, a principled method to model conditional probability density functions, are applied to retrieving Cartesian wind vector components from satellite scatterometer data. A hybrid mixture density network is implemented to incorporate prior knowledge of the predominantly bimodal function branches. An advantage of a fully probabilistic model is that more sophisticated and principled methods can be used to resolve ambiguities.",
keywords = "wind vector retrieval, ERS-1 satellite, probabilistic models, mixture density networks, neural networks",
author = "Evans, {David J.} and Dan Cornford and Nabney, {Ian T.}",
year = "1998",
month = "10",
day = "22",
language = "English",
isbn = "NCRG/98/022",
publisher = "Aston University",
type = "WorkingPaper",
institution = "Aston University",

}

Structured neural network modelling of multi-valued functions for wind retrieval from scatterometer measurements. / Evans, David J.; Cornford, Dan; Nabney, Ian T.

Birmingham : Aston University, 1998.

Research output: Working paperTechnical report

TY - UNPB

T1 - Structured neural network modelling of multi-valued functions for wind retrieval from scatterometer measurements

AU - Evans, David J.

AU - Cornford, Dan

AU - Nabney, Ian T.

PY - 1998/10/22

Y1 - 1998/10/22

N2 - A conventional neural network approach to regression problems approximates the conditional mean of the output vector. For mappings which are multi-valued this approach breaks down, since the average of two solutions is not necessarily a valid solution. In this article mixture density networks, a principled method to model conditional probability density functions, are applied to retrieving Cartesian wind vector components from satellite scatterometer data. A hybrid mixture density network is implemented to incorporate prior knowledge of the predominantly bimodal function branches. An advantage of a fully probabilistic model is that more sophisticated and principled methods can be used to resolve ambiguities.

AB - A conventional neural network approach to regression problems approximates the conditional mean of the output vector. For mappings which are multi-valued this approach breaks down, since the average of two solutions is not necessarily a valid solution. In this article mixture density networks, a principled method to model conditional probability density functions, are applied to retrieving Cartesian wind vector components from satellite scatterometer data. A hybrid mixture density network is implemented to incorporate prior knowledge of the predominantly bimodal function branches. An advantage of a fully probabilistic model is that more sophisticated and principled methods can be used to resolve ambiguities.

KW - wind vector retrieval

KW - ERS-1 satellite

KW - probabilistic models

KW - mixture density networks

KW - neural networks

M3 - Technical report

SN - NCRG/98/022

BT - Structured neural network modelling of multi-valued functions for wind retrieval from scatterometer measurements

PB - Aston University

CY - Birmingham

ER -

Evans DJ, Cornford D, Nabney IT. Structured neural network modelling of multi-valued functions for wind retrieval from scatterometer measurements. Birmingham: Aston University. 1998 Oct 22.