Estimating Forest Leaf Area Index and Canopy Chlorophyll Content with Sentinel-2: An Evaluation of Two Hybrid Retrieval Algorithms

Abstract

Estimates of biophysical and biochemical variables such as leaf area index (LAI) and canopy chlorophyll content (CCC) are a fundamental requirement for effectively monitoring and managing forest environments. With its red-edge bands and high spatial resolution, the Multispectral Instrument (MSI) on board the Sentinel-2 missions is particularly well-suited to LAI and CCC retrieval. Using field data collected throughout the growing season at a deciduous broadleaf forest site in Southern England, we evaluated the performance of two hybrid retrieval algorithms for estimating LAI and CCC from MSI data: the Scattering by Arbitrarily Inclined Leaves (SAIL)-based L2B retrieval algorithm made available to users in the Sentinel Application Platform (SNAP), and an alternative retrieval algorithm optimised for forest environments, trained using the Invertible Forest Reflectance Model (INFORM). Moderate performance was associated with the SNAP L2B retrieval algorithm for both LAI (r2 = 0.54, RMSE = 1.55, NRMSE = 43%) and CCC (r2 = 0.52, RMSE = 0.79 g m−2, NRMSE = 45%), while improvements were obtained using the INFORM-based retrieval algorithm, particularly in the case of LAI (r2 = 0.79, RMSE = 0.47, NRMSE = 13%), but also in the case of CCC (r2 = 0.69, RMSE = 0.52 g m−2, NRMSE = 29%). Forward modelling experiments confirmed INFORM was better able to reproduce observed MSI spectra than SAIL. Based on our results, for forest-related applications using MSI data, we recommend users seek retrieval algorithms optimised for forest environments.

Evaluation of Global Decametric-Resolution LAI, FAPAR and FVC Estimates Derived from Sentinel-2 Imagery

“…According to previous studies, the uncertainty of Sentinel-2 LAI estimates was generally 0.54-1.16 m 2 /m 2 for crops [36,37,39,46,47,[50][51][52] and 1.55 m 2 /m 2 for forest [49]. In terms of Sentinel-2 FAPAR estimates, the uncertainty was 0.11 for crops [39] and 0.16-0.24 for forests [38].…”

“…and land cover types (vegetation, not-vegetated and water). The pixel accuracy of cloud identification was approximate 90% according to the previous study [59], and this layer has been widely used to select the high-quality vegetation pixels in validation activities of Sentinel-2 biophysical estimates [37,38,49]. Moreover, the accuracy of vegetation and non-vegetated identifications from this layer will be further evaluated in Section 4.1 to indicate the reliability of this layer in the validation of Sentinel-2 biophysical estimates.…”

Deriving Leaf Area Index Reference Maps Using Temporally Continuous In Situ Data: A Comparison of Upscaling Approaches

“…Nevertheless, as an extension of the SAVI with additional terms to supress atmospheric effects, the EVI is highly correlated to the SAVI [39], and would be expected to be similarly insensitive to the soil background. In the case of the RTM-based approach, reduced sensitivity to the soil background (and thus reduced bias at lower LAI values) is to be expected, since a variety of soil background reflectance spectra were explicitly incorporated in the INFORM simulations used to train the ANNs [26].…”

“…To upscale in situ estimates of LAI using an RTM-based approach, a hybrid retrieval algorithm making use of leaf/canopy RTMs and machine learning techniques was applied. We adopted the Invertible Forest Reflectance Model (INFORM)-based retrieval algorithm presented by [26], as it was shown to provide improved retrieval accuracy over forest environments when compared to methods involving onedimensional RTMs. The algorithm consists of an artificial neural network (ANN) trained with INFORM simulations designed to reflect the deciduous broadleaf forest site it was developed for.…”

“…In addition to the standard algorithm, we also trained a further ANN using a subset of bands (B3, B4, B8A, B11 and B12) to assess whether a reduction in performance would occur if MSI's red-edge bands were not used as inputs (hereafter referred to as INFORMbroadband). For further information on the ANN training database and procedure, we refer the reader to [26]. Because the ANN training data were optimised for a different study site, some degree of bias in the INFORM-based LAI retrievals was expected.…”