Towards a massive sentinel-2 LAI time-series production using 2-D convolutional networks

Biophysical parameters and more specifically the leaf area index provide an absolute quantification of biomass vegetation allowing overview development status a plant. However, estimation requires sophisticated complex algorithms. This paper proposes new procedure to estimate using Sentinel-2 data. The proposed relies on 2-D convolutional network known as UNet algorithm for regression. architecture is adapted account processing large chunks Moreover, adopted makes use dropout Bayesian approximation at inference step in order allow estimating confidence interval, which very important quality indicator production biophysical parameters. validated multiple tiles years compared multilayer perceptron Sentinel Application Platform European Space Agency, also SNAP. algorithms coherent results when obtained SNAP software with average correlation 0.99 both provides better terms Euclidean distance, mean squared error R2 score. One main advantage vast reduction time regressor. tile 20 m 18 s, 13.5 min 15 UNet, SNAP, respectively. allows massive temporal sequences based images. Furthermore, experiments conducted crop types prove that approach can serve generic regardless type.