Implementing a neural network emulation of a satellite retrieval algorithm

Research output: Chapter in Book/Report/Conference proceedingChapter

3 Citations (Scopus)

Abstract

As shown in Stogryn et al. (1994), Krasnopolsky et al. (1995), Thiria et al. (1993), Cornford et al. (2001), and many other studies summarized in Chapter 9, neural networks (NN) can be used to emulate the physically-based retrieval algorithms traditionally used to estimate geophysical parameters from satellite measurements. The tradeoff involved is a minor sacrifice in accuracy for a major gain in speed, an important factor in operational data analysis. This chapter will cover the design and development of such networks, illustrating the process by means of an extended example. The focus will be on the practical issues of network design and troubleshooting. Two topics in particular are of concern to the NN developer: computational complexity and performance shortfalls. This chapter will explore how to determine the computational complexity required for solving a particular problem, how to determine if the network design being validated supports that degree of complexity, and how to catch and correct problems in the network design and developmental data set. As discussed in Chapter 9, geophysical remote sensing satellites measure either radiances using passive radiometers or backscatter using a transmitter/receiver pair. The challenge is then to estimate the geophysical parameters of interest from these measured quantities. The physics-based forward problem (equation 9.4) captures the cause and effect relationship between the geophysical parameters and the satellite-measured quantities. Thus, the forward problem must be a single-valued function (i.e. have a single possible output value for each set of input values) if we have access to all of its input parameters. As a result, the forward problem is generally well-posed, i.e. variations in the input parameters are not grossly amplified in the output. One could, however, imagine some geophysical processes for which the forward problem was ill-posed for some parameter values as a result of a sudden transition from one regime of behavior to another (e.g. the onset of fog formation producing a sharp change in shortwave albedo in response to a minor change in air temperature).

Original languageEnglish (US)
Title of host publicationArtificial Intelligence Methods in the Environmental Sciences
PublisherSpringer Netherlands
Pages207-216
Number of pages10
ISBN (Print)9781402091179
DOIs
StatePublished - Dec 1 2009

Fingerprint

network design
fog
radiance
radiometer
backscatter
albedo
parameter
physics
air temperature
remote sensing

All Science Journal Classification (ASJC) codes

  • Environmental Science(all)
  • Earth and Planetary Sciences(all)

Cite this

Young, G. S. (2009). Implementing a neural network emulation of a satellite retrieval algorithm. In Artificial Intelligence Methods in the Environmental Sciences (pp. 207-216). Springer Netherlands. https://doi.org/10.1007/978-1-4020-9119-3_10
Young, George S. / Implementing a neural network emulation of a satellite retrieval algorithm. Artificial Intelligence Methods in the Environmental Sciences. Springer Netherlands, 2009. pp. 207-216
@inbook{776b0446aeb043c884499f7921ea32f0,
title = "Implementing a neural network emulation of a satellite retrieval algorithm",
abstract = "As shown in Stogryn et al. (1994), Krasnopolsky et al. (1995), Thiria et al. (1993), Cornford et al. (2001), and many other studies summarized in Chapter 9, neural networks (NN) can be used to emulate the physically-based retrieval algorithms traditionally used to estimate geophysical parameters from satellite measurements. The tradeoff involved is a minor sacrifice in accuracy for a major gain in speed, an important factor in operational data analysis. This chapter will cover the design and development of such networks, illustrating the process by means of an extended example. The focus will be on the practical issues of network design and troubleshooting. Two topics in particular are of concern to the NN developer: computational complexity and performance shortfalls. This chapter will explore how to determine the computational complexity required for solving a particular problem, how to determine if the network design being validated supports that degree of complexity, and how to catch and correct problems in the network design and developmental data set. As discussed in Chapter 9, geophysical remote sensing satellites measure either radiances using passive radiometers or backscatter using a transmitter/receiver pair. The challenge is then to estimate the geophysical parameters of interest from these measured quantities. The physics-based forward problem (equation 9.4) captures the cause and effect relationship between the geophysical parameters and the satellite-measured quantities. Thus, the forward problem must be a single-valued function (i.e. have a single possible output value for each set of input values) if we have access to all of its input parameters. As a result, the forward problem is generally well-posed, i.e. variations in the input parameters are not grossly amplified in the output. One could, however, imagine some geophysical processes for which the forward problem was ill-posed for some parameter values as a result of a sudden transition from one regime of behavior to another (e.g. the onset of fog formation producing a sharp change in shortwave albedo in response to a minor change in air temperature).",
author = "Young, {George S.}",
year = "2009",
month = "12",
day = "1",
doi = "10.1007/978-1-4020-9119-3_10",
language = "English (US)",
isbn = "9781402091179",
pages = "207--216",
booktitle = "Artificial Intelligence Methods in the Environmental Sciences",
publisher = "Springer Netherlands",
address = "Netherlands",

}

Young, GS 2009, Implementing a neural network emulation of a satellite retrieval algorithm. in Artificial Intelligence Methods in the Environmental Sciences. Springer Netherlands, pp. 207-216. https://doi.org/10.1007/978-1-4020-9119-3_10

Implementing a neural network emulation of a satellite retrieval algorithm. / Young, George S.

Artificial Intelligence Methods in the Environmental Sciences. Springer Netherlands, 2009. p. 207-216.

Research output: Chapter in Book/Report/Conference proceedingChapter

TY - CHAP

T1 - Implementing a neural network emulation of a satellite retrieval algorithm

AU - Young, George S.

PY - 2009/12/1

Y1 - 2009/12/1

N2 - As shown in Stogryn et al. (1994), Krasnopolsky et al. (1995), Thiria et al. (1993), Cornford et al. (2001), and many other studies summarized in Chapter 9, neural networks (NN) can be used to emulate the physically-based retrieval algorithms traditionally used to estimate geophysical parameters from satellite measurements. The tradeoff involved is a minor sacrifice in accuracy for a major gain in speed, an important factor in operational data analysis. This chapter will cover the design and development of such networks, illustrating the process by means of an extended example. The focus will be on the practical issues of network design and troubleshooting. Two topics in particular are of concern to the NN developer: computational complexity and performance shortfalls. This chapter will explore how to determine the computational complexity required for solving a particular problem, how to determine if the network design being validated supports that degree of complexity, and how to catch and correct problems in the network design and developmental data set. As discussed in Chapter 9, geophysical remote sensing satellites measure either radiances using passive radiometers or backscatter using a transmitter/receiver pair. The challenge is then to estimate the geophysical parameters of interest from these measured quantities. The physics-based forward problem (equation 9.4) captures the cause and effect relationship between the geophysical parameters and the satellite-measured quantities. Thus, the forward problem must be a single-valued function (i.e. have a single possible output value for each set of input values) if we have access to all of its input parameters. As a result, the forward problem is generally well-posed, i.e. variations in the input parameters are not grossly amplified in the output. One could, however, imagine some geophysical processes for which the forward problem was ill-posed for some parameter values as a result of a sudden transition from one regime of behavior to another (e.g. the onset of fog formation producing a sharp change in shortwave albedo in response to a minor change in air temperature).

AB - As shown in Stogryn et al. (1994), Krasnopolsky et al. (1995), Thiria et al. (1993), Cornford et al. (2001), and many other studies summarized in Chapter 9, neural networks (NN) can be used to emulate the physically-based retrieval algorithms traditionally used to estimate geophysical parameters from satellite measurements. The tradeoff involved is a minor sacrifice in accuracy for a major gain in speed, an important factor in operational data analysis. This chapter will cover the design and development of such networks, illustrating the process by means of an extended example. The focus will be on the practical issues of network design and troubleshooting. Two topics in particular are of concern to the NN developer: computational complexity and performance shortfalls. This chapter will explore how to determine the computational complexity required for solving a particular problem, how to determine if the network design being validated supports that degree of complexity, and how to catch and correct problems in the network design and developmental data set. As discussed in Chapter 9, geophysical remote sensing satellites measure either radiances using passive radiometers or backscatter using a transmitter/receiver pair. The challenge is then to estimate the geophysical parameters of interest from these measured quantities. The physics-based forward problem (equation 9.4) captures the cause and effect relationship between the geophysical parameters and the satellite-measured quantities. Thus, the forward problem must be a single-valued function (i.e. have a single possible output value for each set of input values) if we have access to all of its input parameters. As a result, the forward problem is generally well-posed, i.e. variations in the input parameters are not grossly amplified in the output. One could, however, imagine some geophysical processes for which the forward problem was ill-posed for some parameter values as a result of a sudden transition from one regime of behavior to another (e.g. the onset of fog formation producing a sharp change in shortwave albedo in response to a minor change in air temperature).

UR - http://www.scopus.com/inward/record.url?scp=79951762504&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=79951762504&partnerID=8YFLogxK

U2 - 10.1007/978-1-4020-9119-3_10

DO - 10.1007/978-1-4020-9119-3_10

M3 - Chapter

AN - SCOPUS:79951762504

SN - 9781402091179

SP - 207

EP - 216

BT - Artificial Intelligence Methods in the Environmental Sciences

PB - Springer Netherlands

ER -

Young GS. Implementing a neural network emulation of a satellite retrieval algorithm. In Artificial Intelligence Methods in the Environmental Sciences. Springer Netherlands. 2009. p. 207-216 https://doi.org/10.1007/978-1-4020-9119-3_10