Retrieval of olive tree biophysical properties from Sentinel-2 time series based on physical modelling and machine learning technique

Abstract

ABSTRACT The purpose of this work is to build a model that enables to estimate and monitor the biophysical properties of olive trees planted in super-intensive groves, using innovative forward/backward radiative transfer modelling. In the forward step, the Discrete Anisotropic Radiative Transfer (DART) model simulates a large dataset of satellite image (here, Sentinel-2, S2) reflectance values of realistic olive tree mock-ups with high accuracy. The backward model is composed by two successive regression neural networks trained with the DART simulated dataset. These networks are optimized respectively to estimate biophysical properties of the understory and of the olive trees from the reflectance images. To ensure accurate property retrieval, the dataset covers all possible biophysical and structural properties of the olive trees and understory in addition to the soil optical properties. Therefore, it is composed of trees with various sizes, leaf area index (LAI), tree covers, and leaf biophysical properties associated with ranges of chlorophyll content (Cab), carotenoid content (Ccar), leaf mesophyll structure (N), and leaf equivalent water thickness (EWT) that are compatible to standard properties of olive tree leaves. Particular attention is paid to soil signature estimation, as it is very bright and so greatly influences the canopy signature. Soil reflectance is estimated from the satellite images. Different soil classes with different moisture contents are considered. Inversion results from time series of S2 images are consistent with the olive tree phenological stages. For instance, Cab increases during winter and decreases during summer. The sharp decrease coincides to tree stress caused by lack of water and long-time exposure to the sun. Likewise, an increase in water content in winter is noted, especially in images taken after rainy events. Besides, olive tree sizes estimation in the areas crossed by water alleys shows the highest values. Furthermore, validation is done by comparing our inverted properties to in-situ measurements. For example, LAI, Cab, and EWT RMSE (Root Mean Squared Error) are 0.58 m2m−2, 2.5 µg cm−2 and 0.003 cm, respectively.