Super typhoon Haiyan makes landfall
Super typhoon Haiyan
Subscribe here to receive special images


Lake Maggiore - Italy, 3-6 April 2000

Validation of neural network techniques to estimate canopy biophysical variables from VEGETATION data

M.Weiss(1) , F.Baret(1), M.Leroy(2), O.HautecĹ“ur(2), L.Prévot(1), and N.Bruguier(1)
INRA Bioclimatologie, Domaine Saint-Paul, 84914 Avignon Cédex 9, France
(2) CESBIO, 18 avenue E.Belin, BP 2801, 31041 Toulouse Cédex 4, France

Paper (pdf file, 856 k)

The objective of this study is to develop a global algorithm to monitor the vegetation, applicable to cultivated as well as natural vegetation areas. The monitoring is performed through the estimation of vegetation biophysical variables from 26-day VEGETATION data (fraction cover Fc, leaf area index LAI and fraction of absorbed photosynthetically active radiation fAPAR). Those variables are thus closely linked to the radiative transfer within the canopy and pertinent with regards to possible applications such as canopy primary production modeling or prediction of flux transfer of mass and at the soil-vegetation-atmosphere interface.

The learning phase of the neural networks is achieved by using a synthetic catalog of VEGETATION BRDF. The latter is built thanks to well-known radiative transfer models and a wide range of model parameters for different dates and latitudes. In this study, we do not take into account atmospheric effects and work only with top of canopy reflectance data. As the number of VEGETATION reflectance data during 26 days depends mainly from the latitude and cloud occurrence, it is necessary to pre-process these data to get a constant number of inputs required by neural networks. This is achieved by inverting a linear BRDF model to estimate the nadir and hemispherical reflectances in the 4 VEGETATION wavebands. Three neural networks are then calibrated using these inputs to estimate Fc, LAI, and fAPAR. The optimal architecture is found to be one layer with four sigmoid neurons and one output layer with one linear neuron.

A first validation is performed using a synthetic BRDF catalog of homogeneous and mixed pixels. Results show good performances on Fc and fAPAR. The LAI estimation is less satisfactory for dense canopies due to the saturation of the canopy reflectance. Moreover, LAI estimation is sensitive to the pixel heterogeneity.

A second validation is then performed on experimental data sets provided by the ReSeDA (Remote Sensing Data Assimilation, 1997). The ReSeDA site is a 4km*5km agricultural area (mainly wheat, sunflower, alfalfa and maize) near Avignon (France). Ground measurements (LAI, Fc) were performed during the whole crop cycles. Reflectance data were acquired with the airborne POLDER sensor at 15 dates during the year. The neural network technique is first modified to be consistent with POLDER measurements and then applied to retrieve biophysical variables. The comparison between estimated variables and in situ measurements is quite consistent with the results obtained with synthetic data