Physics-constrained deep learning for biophysical parameter retrieval from Sentinel-2 images: Inversion of the PROSAIL model

In this era of global warming, the regular and accurate mapping of vegetation conditions is essential for monitoring ecosystems, climate sustainability and biodiversity. In this context, this work proposes a physics-guided data-driven methodology to invert radiative transfer models (RTM) for the retrieval of vegetation biophysical variables. A hybrid paradigm is proposed by incorporating the physical model to be inverted into the design of a neural network architecture, which is trained by exploiting unlabeled satellite images. In this study, we show how the proposed strategy allows the simultaneous probabilistic inversion of all input PROSAIL model parameters by exploiting Sentinel-2 images. The interest of the proposed self-supervised learning strategy is corroborated by showing the limitations of existing simulation-trained machine learning algorithms. Results are assessed on leaf area index (LAI) and canopy chlorophyll content (CCC) in-situ measurements collected on four different field campaigns over three European tests sites. Prediction accuracies are compared with performances reached by the well-established Biophysical Processor (BP) of the Sentinel Application Platform (SNAP). Obtained overall accuracies corroborate that the proposed methodology achieves performances equivalent to or better than the state-of-the-art methods.