Archetypal crop trait dynamics for enhanced retrieval of biophysical parameters from Sentinel-2 MSI

Abstract

We present a new method for estimating biophysical parameters from Earth Observation (EO) data using a crop-specific empirical model based on the PROSAIL Radiative Transfer (RT) model, called an ‘archetype’ model. The first-order model presented uses maximum biophysical parameter magnitude, phenological and soil parameters to describe the spectral reflectance (400–2500 nm) of vegetation over time. The approach assumes smooth variation and archetypical coordination of crop biophysical parameters over time for a given crop. The form of coordination is learned from a large sample of observations. Using Sentinel-2 observations of maize from Northeast China in 2019, we map reflectance to biophysical parameters using an inverse model operator, synchronise the parameters to a consistent time frame using a double logistic model of $LAI$, then derive the model archetypes as the median value of the synchronised samples. We apply the model to estimate time series of biophysical parameters for different cereal crops using an ensemble framework with a weighted K-nearest neighbour solution, and validate the results with ground measurements of different crops collected near Munich, Germany in 2017 and 2018. The results show $R$ values greater than 0.8 for leaf area index ($LAI$) and leaf brown pigment content ($C_{brown}$), with an RMSE of 0.94 ${\text{m}}^2/{\text{m}}^2$ for $LAI$ and 0.15 for $C_{brown}$. The chlorophyll content ($C_{ab}$) and canopy water content ($C{C}_w$) were retrieved at a higher level of accuracy, with $R$ values around 0.9 and an RMSE of $6.59\mu {\text{g/cm}}^2$ for $C_{ab}$ and 0.03 ${\text{g/cm}}^2$ for $C{C}_w$. Comparison of forward-modelled hyperspectral reflectance with independent ground measures shows that the retrieved parameters account for 90% of the variation in canopy reflectance, with an overall RMSE of around 0.05 in reflectance units. The retrievals for all terms are mostly within $1\sigma$ when measurement and prediction uncertainty are taken into account, except for some early and late season issues in leaf and canopy water due to the complexity of canopy structure and understory during these periods. The approach provides a new form of constraint for the simultaneous estimation of biophysical parameters from EO and greatly reduces the rank of the problem. It is suitable for monitoring crop conditions where biophysical parameters vary smoothly over time consistently with each archetype form. The approach can be refined for other canopy types and canopy representations and could provide strong constraints on expected smoothly-varying canopy features to aid in the interpretation of EO signals across different regions of the electromagnetic spectrum.