2008 IEEE International Geoscience & Remote Sensing Symposium
July 6-11, 2008 | Boston, Massachusetts, U.S.A.

HD-2: Feedforward Neural Networks: Theory & Applications in Atmospheric Remote Sensing

Sunday Afternoon, July 6, 13:30 - 17:30

Presented by

Frederick W. Chen

Abstract

This tutorial will provide an introduction to neural networks and describe their application in current atmospheric remote sensing problems. Neural networks are a computational intelligence method inspired by biological neural networks. They were developed to learn variations in data in a way similar to how biological organisms learn from their environments. The ability of neural networks to learn nonlinear variations in data without any assumptions can be very useful in problems where variations can be extremely difficult to express in closed form. In this tutorial, applications to satellitebased estimation of atmospheric temperature and water vapor profile and precipitation rate will be described.

This first part of the tutorial will describe how to use neural networks. Important issues include data for training and evaluation and topology selection. Common mistakes will also be covered. While neural networks can be powerful, they are also frequently misused due to a widespread lack of understanding of their capabilities. Another important issue is preprocessing and postprocessing. Measurements (e.g. from satellites) are rarely in a form that neural networks can learn with, and frequently preprocessing is needed to make the measurements and auxiliary data most relevant to a particular problem. Specific examples of preprocessing will include methods based on principal component analysis (PCA) and methods that consider the topology of relevant variables. Preprocessing can also be used to reduce the dimensionality of the data set which can improve computational efficiency in both training and simulation. Postprocessing can also be useful in some situations, e.g. one in which the output varies over a wide range of scales.

Neural networks have been used to develop computationally efficient atmospheric retrieval algorithms. Blackwell (IEEE TGRS, 11/2005) has developed an algorithm for estimating temperature and water vapor profile that uses projection principal components to preprocess hyspectral infrared data from the Atmospheric Infrared Sounder (AIRS), and Chen and Staelin (IEEE TGRS, 2/2003) have developed a method for estimating precipitation rate using data from the Advanced Microwave Sounding Unit (AMSU). Applications in other areas of geoscience and remote sensing are possible.

Outline

  1. Introduction (1-1.5 hrs)
    1. Inspiration from biological systems
    2. Brief descriptions of NN (including self-organizing maps, feedback NN, ...)
    3. In-depth description of feedforward NN
    4. Training NN (training algorithms; training, validation, and testing sets)
    5. Choosing appropriate topology
    6. Simple examples to provide intuition
  2. Preprocessing & postprocessing (1-2 hours)
    1. PCA & variations (NAPC, PPC, constrained PCA)
    2. Topological preprocessing (circular data)
    3. Preprocessing based on a priori knowledge of system
    4. Postprocessing
  3. Applications in atmospheric remote sensing (1 hour)
    1. AMSU precipitation
    2. SCC/NN

Speaker Biography

Frederick W. Chen was most recently a technical staff member at the Massachusetts Institute of Technology (MIT) Lincoln Laboratory (Lexington, MA) where he worked on problems in satellite-based atmospheric remote sensing using microwave and infrared data. His research interests also include signal separation, transform coding, data compression, neural nets, and other areas in signal processing and information theory. In 1996 he worked at Argonne National Laboratory (Argonne, IL) in the Reactor Analysis Division on failure detection using neural nets and control system theory, and from 1997 to 2004 he worked at the MIT Research Laboratory of Electronics (Cambridge, MA) in the Remote Sensing and Estimation Group on precipitation estimation using satellitebased passive opaque microwave radiometry, and compression of 2-D geophysical data. He received the S.B., M.Eng., and Ph.D. degrees in electrical engineering from MIT in 1998, 1998, and 2004, respectively. He received 1st prize in the student paper competition at the 2002 IEEE International Geoscience and Remote Sensing Symposium.