Skip all navigation and jump to content Jump to site navigation
About MODIS News Data Tools /images2 Science Team Science Team Science Team

   + Home
ABOUT MODIS
MODIS Publications Link
MODIS Presentations Link
MODIS Biographies Link
MODIS Science Team Meetings Link
 

 

 

Kimes, D, Gastellu-Etchegorry, J, Esteve, P (2002). Recovery of forest canopy characteristics through inversion of a complex 3D model. REMOTE SENSING OF ENVIRONMENT, 79(3-Feb), 320-328.

Abstract
Radiative transfer models for vegetation serve as a basis for extracting vegetation variables using directional/spectral data from modern-borne sensors (e.g., MODIS. MISR, POLDER, SeaWiFS). Only recently have significant efforts been made to provide operational algorithms to invert these models. These efforts have exposed a need to significantly improve the efficiency and accuracy of traditional methods for inverting these physically based models. In an effort to overcome the limitations of traditional inversion methods. a neural network method was designed and tested. In this study, a complex 3D model (Discrete Anisotropic Radiative Transfer, DART) was inverted for a wide range of simulated forest canopies using POLDER-like data. The model was inverted to recover three forest canopy variables: forest cover, leaf area index, and a soil reflectance parameter. The ranges of these variables were 0.4-1.0, 0.8-9.3, and 0.0-1.0, respectively. Two inversion methods were used - a traditional inversion technique using a modified simplex method, and a neural network method in combination with an exhaustive variable selection technique. A comparison of the methods' efficiency, accuracy, and stability was made. The neural network method gave relatively accurate solutions to the inversion problem given a small subset of directional/spectral data using only one to five view angles. Using only nadir data, the root mean squared error (RMSE) for the forest cover, leaf area index. and the soil reflectance parameter were 0.025, 0.23, and 0.15, respectively, and using the best view directions (2-5) were 0.021, 0.21, and 0.11, respectively. In general, the neural network method was more accurate than the simplex method. The results from both methods showed that the addition of directional view angles, as opposed to only a nadir view, can significantly improve the accuracy of recovering forest canopy characteristics. The traditional simplex method is computationally intensive and may not be appropriate for many operational applications on a per-pixel basis for regional and global data. The neural network method was computationally efficient and can be applied on a per-pixel basis. In general, the neural network technique had significantly lower RMSE values at the low noise levels. However. at moderate noise levels. the simplex method was equal to the neural network method in RMSE values. At high noise levels. the simplex method had significantly lower RMSE values than the neural network method. The neural network approach can provide an accurate, efficient, and stable inversion method for radiative transfer models using directional/spectral data from modern-borne sensors. (C) 2002 Elsevier Science Inc. All rights reserved.

DOI:

ISSN:
0034-4257

NASA Home Page Goddard Space Flight Center Home Page