Assessment of vegetation biochemical and biophysical variables is useful when developing indicators for biodiversity monitoring and climate change studies. Here, we compared a radiative transfer model (RTM) inversion by merit function and five machine learning algorithms trained on an RTM simulated dataset predicting the three plant traits leaf chlorophyll content (LCC), canopy chlorophyll content (CCC), and leaf area index (LAI), in a mixed temperate forest. The accuracy of the retrieval methods in predicting these three plant traits with spectral data from Sentinel-2 acquired on 13 July 2017 over Bavarian Forest National Park, Germany, was evaluated using in situ measurements collected contemporaneously. The RTM inversion using merit function resulted in estimations of LCC (R ² = 0.26, RMSE = 3.9 µg/cm²), CCC (R ² = 0.65, RMSE = 0.33 g/m²), and LAI (R ² = 0.47, RMSE = 0.73 m²/m²), comparable to the estimations based on the machine learning method Random forest regression of LCC (R ² = 0.34, RMSE = 4.06 µg/cm²), CCC (R ² = 0.65, RMSE = 0.34 g/m²), and LAI (R ² = 0.47, RMSE = 0.75 m²/m²). Several of the machine learning algorithms also yielded accuracies and robustness similar to the RTM inversion using merit function. The performance of regression methods trained on synthetic datasets showed promise for fast and accurate mapping of plant traits accross different plant functional types from remote sensing data.

The Sentinel satellite fleet of the Copernicus Programme offers new potential to map and monitor plant traits at fine spatial and temporal resolutions. Among these traits, leaf area index (LAI) is a crucial indicator of vegetation growth and an essential variable in biodiversity studies. Numerous studies have shown that the radiative transfer approach has been a successful method to retrieve LAI from remote-sensing data. However, the suitability and adaptability of this approach largely depend on the type of remote-sensing data, vegetation cover and the ecosystem studied. Saltmarshes are important wetland ecosystems threatened by sea level rise among other human- and animal-induced changes. Therefore, monitoring their vegetation status is crucial for their conservation, yet few LAI assessments exist for these ecosystems. In this study, the retrieval of LAI in a saltmarsh ecosystem is examined using Sentinel-2 and RapidEye data through inversion of the PROSAIL radiative transfer model. Field measurements of LAI and some other plant traits were obtained during two succeeding field campaigns in July 2015 and 2016 on the saltmarsh of Schiermonnikoog, a barrier island of the Netherlands. RapidEye (2015) and Sentinel-2 (2016) data were acquired concurrent to the time of the field campaigns. The broadly employed PROSAIL model was inverted using two look-up tables (LUTs) generated in the spectral band’s settings of the two sensors and in which each contained 500,000 records. Different solutions from the LUTs, as well as, different Sentinel-2 spectral subsets were considered to examine the LAI retrieval. Our results showed that generally the LAI retrieved from Sentinel-2 had higher accuracy compared to RapidEye-retrieved LAI. Utilising the mean of the first 10 best solutions from the LUTs resulted in higher R² (0.51 and 0.59) and lower normalised root means square error (NRMSE) (0.24 and 0.16) for both RapidEye and Sentinel-2 data respectively. Among different Sentinel-2 spectral subsets, the one comprised of the four near-infrared (NIR) and shortwave infrared (SWIR) spectral bands resulted in higher estimation accuracy (R² = 0.44, NRMSE = 0.21) in comparison to using other studied spectral subsets. The results demonstrated the feasibility of broadband multispectral sensors, particularly Sentinel-2 for retrieval of LAI in the saltmarsh ecosystem via inversion of PROSAIL. Our results highlight the importance of proper parameterisation of radiative transfer models and capacity of Sentinel-2 spectral range and resolution, with impending high-quality global observation aptitude, for retrieval of plant traits at a global scale.

Understanding the vertical pattern of leaf traits across plant canopies provide critical information on plant physiology, ecosystem functioning and structure and vegetation response to climate change. However, the impact of vertical canopy position on leaf spectral properties and subsequently leaf traits across the entire spectrum for multiple species is poorly understood. In this study, we examined the ability of leaf optical properties to track variability in leaf traits across the vertical canopy profile using Partial Least Square Discriminatory Analysis (PLS-DA). Leaf spectral measurements together with leaf traits (nitrogen, carbon, chlorophyll, equivalent water thickness and specific leaf area) were studied at three vertical canopy positions along the plant stem: lower, middle and upper. We observed that foliar nitrogen (N), chlorophyll (C_ab), carbon (C), and equivalent water thickness (EWT) were higher in the upper canopy leaves compared with lower shaded leaves, while specific leaf area (SLA) increased from upper to lower canopy leaves. We found that leaf spectral reflectance significantly (P ≤ 0.05) shifted to longer wavelengths in the ‘red edge’ spectrum (685–701 nm) in the order of lower > middle > upper for the pooled dataset. We report that spectral bands that are influential in the discrimination of leaf samples into the three groups of canopy position, based on the PLS-DA variable importance projection (VIP) score, match with wavelength regions of foliar traits observed to vary across the canopy vertical profile. This observation demonstrated that both leaf traits and leaf reflectance co-vary across the vertical canopy profile in multiple species. We conclude that canopy vertical position has a significant impact on leaf spectral properties of an individual plant’s traits, and this finding holds for multiple species. These findings have important implications on field sampling protocols, upscaling leaf traits to canopy level, canopy reflectance modelling, and subsequent leaf trait retrieval, especially for studies that aimed to integrate hyperspectral measurements and LiDAR data.

In remote sensing applications, leaf traits are often upscaled to canopy level using sunlit leaf samples collected from the upper canopy. The implicit assumption is that the top of canopy foliage material dominates canopy reflectance and the variability in leaf traits across the canopy is very small. However, the effect of different approaches of upscaling leaf traits to canopy level on model performance and estimation accuracy remains poorly understood. This is especially important in short or sparse canopies where foliage material from the lower canopy potentially contributes to the canopy reflectance. The principal aim of this study is to examine the effect of different approaches when upscaling leaf traits to canopy level on model performance and estimation accuracy using spectral measurements (in-situ canopy hyperspectral and simulated Sentinel-2 data) in short woody vegetation. To achieve this, we measured foliar nitrogen (N), leaf mass per area (LMA), foliar chlorophyll and carbon together with leaf area index (LAI) at three vertical canopy layers (lower, middle and upper) along the plant stem in a controlled laboratory environment. We then upscaled the leaf traits to canopy level by multiplying leaf traits by LAI based on different combinations of the three canopy layers. Concurrently, in-situ canopy reflectance was measured using an ASD FieldSpec-3 Pro FR spectrometer, and the canopy traits were related to in-situ spectral measurements using partial least square regression (PLSR). The PLSR models were cross-validated based on repeated k-fold, and the normalized root mean square errors (nRMSE_cv) obtained from each upscaling approach were compared using one-way analysis of variance (ANOVA) followed by Tukey’s post hoc test. Results of the study showed that leaf-to-canopy upscaling approaches that consider the contribution of leaf traits from the exposed upper canopy layer together with the shaded middle canopy layer yield significantly (p < 0.05) lower error (nRMSE_cv < 0.2 for canopy N, LMA and carbon) as well as high explained variance (R² > 0.71) for both in-situ hyperspectral and simulated Sentinel-2 data. The widely-used upscaling approach that considers only leaf traits from the upper illuminated canopy layer yielded a relatively high error (nRMSE_cv>0.2) and lower explained variance (R² < 0.71) for canopy N, LMA and carbon. In contrast, canopy chlorophyll upscaled based on leaf samples collected from the upper canopy and total canopy LAI exhibited a more accurate relationship with spectral measurements compared with other upscaling approaches. Results of this study demonstrate that leaf to canopy upscaling approaches have a profound effect on canopy traits estimation for both in-situ hyperspectral measurements and simulated Sentinel-2 data in short woody vegetation. These findings have implications for field sampling protocols of leaf traits measurement as well as upscaling leaf traits to canopy level especially in short and less foliated vegetation where leaves from the lower canopy contribute to the canopy reflectance.