Inverting the PROSAIL canopy reflectance model using neural nets trained on streamlined databases

Abstract

The widely used PROSAIL radiative transfer model was coupled with a simple soil reflectance parameterisation to estimate the leaf area index (LAI) of winter wheat (Triticum aestivum) from ground-based spectrometer data. To avoid time-consuming numerical optimisations, a neural net (NN) was used for model inversion. The NN was trained on 3000 spectral patterns generated by the reflectance model. The training database was previously streamlined to provide good approximation of the response surface while keeping the net compact. Streamlining was achieved by retaining only those synthetic spectra that belong both to the simulated and actual measurement spaces. The estimated LAI (nobs = 15) compared well with completely independent reference measurements taken four times during the 2000 growing season in four commercial winter wheat fields (1.8 ≤ LAI ≤  8.1). The coefficient of determination (R2) between measured and estimated LAI was 0.87 with a root mean squared error (RMSE) of 0.89 (m2 m–2). Even for LAIs exceeding 3–4, saturation effects were low. Three measurement dates yielded RMSE lower than 0.8. Only during stem elongation did RMSE exceed 1. Higher errors for this time period were attributed to abrupt changes in the canopy structure (i.e. average leaf angle) not taken into account. Compared to the normalised difference vegetation index (NDVI), the inversion of PROSAIL using hyperspectral reflectances performed well, with errors reduced by more than 50% as compared to the NDVI model (RMSE: 1.91 m2 m–2).