Verger, A; Baret, F; Camacho, F (2011). Optimal modalities for radiative transfer-neural network estimation of canopy biophysical characteristics: Evaluation over an agricultural area with CHRIS/PROBA observations. REMOTE SENSING OF ENVIRONMENT, 115(2), 415-426.
Neural networks trained over radiative transfer simulations constitute the basis of several operational algorithms to estimate canopy biophysical variables from satellite reflectance measurements. However, only little attention was paid to the training process which has a major impact on retrieval performances. This study focused on the several modalities of the training process within neural network estimation of LAI, FCOVER and FAPAR biophysical variables. Performances were evaluated over both actual experimental observations and model simulations. The SAIL and PROSPECT radiative transfer models were used here to simulate the training and the synthetic test datasets. Measurements of LAI. FCOVER and FAPAR were achieved over the Barrax (Spain) agricultural site for a range of crop types concurrently to CHRIS/PROBA satellite image acquisition. Results showed that the spectral band selection was specific to LAI, FCOVER and FAPAR variables. The optimal band set provided significantly improved performances for LAI, while only small differences were observed for the other variables. Gaussian distributions of the radiative transfer model input variables performed better than uniform distributions for which no prior information was exploited. Including moderate uncertainties in the reflectance simulations used in the training process improved the flexibility of the neural network in cases where simulations departed slightly from observations. Simple neural network architecture with a single hidden layer of five tangent sigmoid transfer functions was performing as good as more complex architectures if the training dataset was larger than ten times the number of coefficients to tune. Small sensitivity of performances was observed depending on the way the solution was selected when several networks were trained in parallel. Finally, comparison with a NDVI based approach showed the generally better retrieval accuracy of neural networks. (C) 2010 Elsevier Inc. All rights reserved.