Canopy nitrogen content retrieval from hyperspectral satellite data through spectral band selection with Gaussian processes

Advanced retrieval models allow us to make inferences from the signals acquired remotely by satellites to a set of variables, to better understand and describe the states and dynamics of croplands. One essential variable is canopy nitrogen content (CNC), being one of the most relevant traits for agricultural monitoring applications. In the next coming years, there will be an increasing amount of available data acquired by a new generation of hyperspectral satellites (image spectrometer missions), such as PRISMA, and upcoming EnMAP and CHIME missions. When dealing with hyperspectral satellite data, the curse of dimensionality and the effects of noise can be successfully alleviated through feature (band) selection procedures. In our proposed setting, most meaningful spectral bands for the retrieval of CNC were selected, providing a lower spectral subset of the original data but maintaining the physical meaning of each spectral band. Radiative transfer models (RTM) simulate bi-derectional reflectance as a function of diverse biochemical and biophysical input parameters. In this way, RTMs allow to build upon new methods and prepare future missions due to its capability of simulating real scenarios based on their physical consistent definition. In this work, we focus on the leaf optical properties model PROSPECT-PRO coupled with the canopy reflectance 4SAIL model to establish a training database for Gaussian process (GP) regression algorithms. The proposed methodology performs regression from input values, the reflectance, to the output values, the biophysical parameters or traits of interest. In this work, we explored a spectral band selection tool (GPR-BAT) embedded in the ARTMO toolbox (https://artmotoolbox.com/), dedicated to the transformation of optical remote sensing images into biophysical vegetation products and maps. GPR-BAT is based on a sequential backward band removal (SBBR) algorithm that iteratively removes the spectral bands which contribute less to the regression model. This procedure is repeated until only one relevant band is left over. GPR-BAT allows to: i) identify the most informative or relevant bands to estimate one specific biophysical or biochemical variable, and ii) find a smaller set of bands preserving the optimal predictions. The optimal set of 15 bands achieved a coefficient of determination (R²) of 0.6 and a normalised root mean squared error (NRMSE) of 19 % to retrieve canopy nitrogen content sampled over maize and winter wheat during a field campaign in the North of Munich, Germany (MMNI site), during 2017 and 2018 growing seasons. Furthermore, a variance-based global sensitivity analysis of the PROSAIL-PRO model confirmed the optimal position of the identified band setting within the nitrogen (protein) sensitive wavelength domain. The optimal set of bands were to be found in the near infrared and in the short wave infrared, especially in the 1700-1800 nm region. Applying the established models on acquired PRISMA images revealed the adequacy of the proposed method for mapping applications. We conclude that our proposed methodology achieved promising results both in accuracy of estimates and mapping quality over different geographical regions.