Biophysical variables estimation from Sentinel 2 images: emphasizing the importance of vertical heterogeneity, solar and viewing geometry, and associated biophysical properties

Abstract

Climate change disrupts ecosystems and increases extreme weather events. Modeling ecosystem functioning is crucial for effective adaptation strategies. This study focuses on quantifiable vegetation properties, including leaf area index (LAI), chlorophyll content (CHL), leaf mass per area (LMA), and equivalent water thickness (EWT). One of the most effective ways to retrieve plant biophysical and structural properties at a large scale is by using multispectral satellite images. However, accurately quantifying plant biophysical variables from such datasets presents several challenges, including understanding the influence of the vertical distribution of such traits within the canopy on the corresponding signal; the impact of sun-view geometry during image acquisition, the use of genotype-specific relationship or instead the use of genotype biophysical traits in the model. This study estimates biophysical variables from a large measurement dataset obtained on forest plantations in Sao Paulo, Brazil, and analyzes the effect of vertical heterogeneity, solar and viewing geometry, and associated biophysical properties. The dataset comprises in-situ and remote sensing data. In-situ measurements of LAI, CHL, LMA, and EWT were collected from 2019 to 2021 on 25 eucalypt genotypes. Biomass measurements were conducted in 1323 trees, and CHL, LMA, and EWT were measured per vertical third of the canopy. To evaluate the influence of the vertical heterogeneity, we defined a weighted expression of the top and middle thirds of the canopy to average these biophysical variables and relate them to the remote sensing data, which includes Sentinel-2 images acquired from 2019 to 2021, at dates close to the field measurements.  First, 22 vegetation indices (VIs) were used to build regression models, each regressing the weighted target variable to a VI. After determining the optimal weight, we tested the accuracy of the linear models by accounting for sun-sensor geometry and vegetation traits. Both 10-fold cross-validation and an independent test dataset were used to assess model performance together with root mean square (RMSE) and coefficient of determination (R2). Results show that most models using 70% top and 30% medium canopy produced the best performances in estimating CHL. For both LMA and EWT, the optimal percentage was 50%-50%. These outcomes indicate that shaded parts of the canopy play a significant role in the above-canopy reflectance, especially for LMA and EWT, which are particularly sensitive to spectral domains ranging from 1700 to 2400 nm, which has a higher transmittance rate towards the canopy. Concerning the inclusion of sun-sensor geometry, most models, generated with different VIs, benefitted from these variables to predict the target variable, resulting in lower RMSEs. The use of several canopy traits in the model reduced the error but would require to have previous knowledge of them. These empirical results underline the influence of sensor geometry and other biophysical properties on the prediction of LAI, CHL, EWT, and LMA. We believe that these results advocate for further investigation using radiative transfer model inversion.