Prototyping Sentinel-2 green LAI and brown LAI products for cropland monitoring

Abstract

For agricultural applications, identification of non-photosynthetic above-ground vegetation is of great interest as it contributes to assess harvest practices, detecting crop residues or drought events, as well as to better predict the carbon, water and nutrients uptake. While the mapping of green Leaf Area Index (LAI) is well established, current operational retrieval models are not calibrated for LAI estimation over senescent, brown vegetation. This not only leads to an underestimation of LAI when crops are ripening, but is also a missed monitoring opportunity. The high spatial and temporal resolution of Sentinel-2 (S2) satellites constellation offers the possibility to estimate brown LAI (LAI_G) next to green LAI (LAI_G). By using LAI ground measurements from multiple campaigns associated with airborne or satellite spectra, Gaussian processes regression (GPR) models were developed for both LAI_G and LAI_B, providing alongside associated uncertainty estimates, which allows to mask out unreliable LAI retrievals with higher uncertainties. A processing chain was implemented to apply both models to S2 images, generating a multiband LAI product at 20 m spatial resolution. The models were adequately validated with in-situ data from various European study sites (LAI_G: R² = 0.7, RMSE = 0.67 m²/m²; LAI_B: R² = 0.62, RMSE = 0.43 m²/m²). Thanks to the S2 bands in the red edge (B5: 705 nm and B6: 740 nm) and in the shortwave infrared (B12: 2190 nm) a distinction between LAI_G and LAI_B can be achieved. To demonstrate the capability of LAI_B to identify when crops start senescing, S2 time series were processed over multiple European study sites and seasonal maps were produced, which show the onset of crop senescence after the green vegetation peak. Particularly, the LAI_B product permits the detection of harvest (i.e., sudden drop in LAI_B) and the determination of crop residues (i.e., remaining LAI_B), although a better spectral sampling in the shortwave infrared would have been desirable to disentangle brown LAI from soil variability and its perturbing effects. Finally, a single total LAI product was created by merging LAI_G and LAI_B estimates, and then compared to the LAI derived from S2 L2B biophysical processor integrated in SNAP. The spatiotemporal analysis results confirmed the improvement of the proposed descriptors with respect to the standard SNAP LAI product accounting only for photosynthetically active green vegetation.

1. Introduction

Satellite-derived leaf area index (LAI) products are widely used in regional-to-global vegetation monitoring services such as global models of biosphere-atmosphere exchanges of energy, CO2, water vapour (e.g. Asner et al., 2003; Ganguly et al., 2008; Levis et al., 2012; Mahowald et al., 2016), crop growth modelling and yield prediction (e.g. Ines et al., 2013; Liu et al., 2014; Li et al., 2017; Weiss et al., 2020). Current operational services of global LAI products are mainly generated from satellite observations of SPOT/VGT, MODIS and NOAA/AVHRR sensors (Baret et al., 2013; Yan et al., 2016; Xiao et al., 2016). When reviewing the core retrieval algorithms behind these LAI products, they can be categorized as either based on: (1) the inversion of a vegetation radiative transfer model (RTM), or (2) a machine learning method, typically trained by RTM data. See Verrelst et al. (2015a) for a systematic review on retrieval methods, and Jiang et al. (2017); Liu et al. (2018) for inter-annual comparisons of these global LAI products. Among the global LAI retrieval algorithms routinely processing satellite imagery, the machine learning family of artificial neural networks turned out to be popular because of their fast processing and ability to deliver accurate estimates. Neural networks are currently used as core algorithms driving the rendering of global LAI product such as LAI3g (Zhu et al., 2013), TCDR (Claverie et al., 2016) and GLASS (Xiao et al., 2016). At the same time - but then for user-initiated applications - a neural network model has been implemented as Sentinel-2 toolbox processor to enable individual S2 image processing into LAI maps (Weiss and Baret, 2016).

Regardless of whether the LAI retrieval algorithm is based on RTM inversion or on a trained machine learning algorithm, these methods quantify LAI only for green vegetation. Although LAI is defined as one half the total leaf area per unit ground surface area (Jonckheere et al., 2004; Yan et al., 2019), the term green LAI (hereafter referred to as LAI_G is more commonly used in the context of remote sensing studies (Baret et al., 2013; Nguy-Robertson et al., 2014; Verrelst et al., 2014). LAI_G represents half the total area of solely green elements per ground unit area. For crop fields, this thus only accounts for functioning above-ground parts of the plants that are green and photosynthetically active during a significant fraction of the growth cycle (Boegh et al., 2002; Duveiller et al., 2012). However, at the end of the growing season foliage remains on the plant during ripening and senescence phase until falling off or being harvested, meaning LAI stays high, as senescent leaves still intercept radiation (Steduto and Hsiao, 1998; Li and Guo, 2016), even when chlorophyll content in photosynthesizing tissues degrades to zero and the plants loses its greenness, ending with the death of the leaf (Gregersen et al., 2013).

This non-photosynthetic vegetation, such as crop residues and senescence foliage, accounts for a significant amount of above-ground biomass and represents a key factor in the carbon cycle (Vuichard et al., 2007; Li and Guo, 2016), as well as controlling the uptake of carbon, water and nutrients, serving therefore for ecosystem managing (Li and Guo, 2016). Accordingly, the spatially-explicit quantification of LAI over senescent vegetation, hereafter referred to as brown LAI or LAI_B, can thus potentially help closing the above-ground carbon budget. Moreover, routine quantification of LAI_B might be of interest for managing agroecosysems; it can indicate when the crop is ready for harvesting or whether crop residues are left on the field (Daughtry et al., 2006). Also, regular monitoring of LAI_B can contribute to detect seasonal drought events (Hank et al., 2019) or to improve fire risk assessments, since senescent croplands are more fire-prone than green fields (Guerschman et al., 2009). Consequently, it would thus not only be erroneous but also a missed monitoring opportunity to interpret senescent vegetation with a LAI close to zero, as is typically occurring with LAI models that rely on RTM simulations of green vegetation (Houborg and Boegh, 2008; Verger et al., 2011). Generally, LAI_G models tend to systematically underestimate LAI when crops are ripening, because they are not calibrated for senescent vegetation (Haboudane et al., 2004; Wang et al., 2005; Delegido et al., 2013) and even can cause confusion with bare soil (Houborg et al., 2007).

Historically, apart from the importance of quantifying the LAI of green photosynthesizing vegetation for multiple crop monitoring reasons (e.g. Frank and Dugas, 2001; Gitelson et al., 2003b), another reason why emphasis has mostly been on LAI_G is technical. Traditional silicon-based optical optical sensors only covered the spectral region sensitive to green vegetation in the visible and near infrared (VNIR, 400–1100 nm) part of the electromagnetic spectrum. The amount of green vegetation drives the spectral variation in the VNIR, because photosynthetic active plant tissues lead to strong absorption in the visible and to high reflectance in the near infrared (Gitelson et al., 2003a), making VNIR sensors successful for the estimation of LAI_G (see Fang et al. (2019) and Yan et al. (2019) for reviews on methods and instruments). Instead, when moving towards the shortwave infrared (SWIR, ~ 1100-2500 nm) other leaf components absorb and scatter light, such as water, cellulose, lignin and several other biochemical compounds (Curran, 1989; Kokaly et al., 2009), causing changes in the vegetation reflectance spectra that permits the identification of the senescent process. For this reason, the SWIR region shows great potential for distinguishing green from senescent vegetation and also from bare soil (Delegido et al., 2015), and bands in the SWIR can thus serve for the estimating of LAI over sensescent vegetation, thus LAI_B (Brown et al., 2000; Kokaly et al., 2009).

Nowadays, an attractive operational Earth observer with capability to exploit the SWIR spectral region for vegetation mapping involves the Sentinel-2 (S2) mission. The S2 constellation of two satellites enables a weekly global revisit time, and their core optical sensor – the Multi-Spectral Instrument (MSI) – has been configured with 13 bands covering the VNIR to SWIR spectral range and is, among others, optimized for agroecosystems monitoring applications (Drusch et al., 2012; Li and Roy, 2017). This opens opportunities to new or improved quantification of vegetation products, e.g. enabling an explicit distinction between photosynthesizing (i.e. green) and senescent (i.e. brown) vegetation (Delegido et al., 2015). The latter study had earlier demonstrated to separate and quantify brown from green vegetation with a two-step approach based on band-pairs spectral indices with spectral regions closely positioned to the S2 spectral bands in the red edge (B5: 705 nm and B6: 740 nm) and in the SWIR (Bl 1: 1610 nm and Bl 2: 2190 nm). When it comes to operational processing, however, relying on spectral indices may not be the best strategy because of well-known limitations: (1) underexploitation of the available multispectral information content, (2) poorly transferability of rules to other images in space and time, and more importantly, (3) the lack of uncertainty estimates that serve as a quality indicator (Verrelst et al., 2015b,a). Therefore, alternative retrieval methods must be sought that not only enable an explicit separation between LAI_G and LAI_B by fully exploiting the available spectral data, but are also generally applicable over most regions without requiring site-specific calibration, and at the same time provide associated uncertainty estimates.

With the purpose of opening new opportunities to monitor crop biophysical variables at the field scale into an operational context, the recently finished HORIZON 2020 Sentinels Synergy for Agriculture (SENSAGRI) project expressed the ambition to retrieve independently LAI over photosynthesizing (green) and senescent (brown) agroecosystem surfaces from S2 imagery (http://sensagri.eu/). The project aimed to deliver a portfolio of crop monitoring products, including both LAI_G and LAI_B, implementable into operational S2 data processing chains. This requires the development of retrieval models that are fast, robust and easily applicable. To this end, based on a systematic comparison of parametric, nonparametric and RTM-inversion retrieval methods taking both accuracies and run-time into account, it was concluded that most accurate and fast estimates from S2 can be achieved with novel machine learning methods (Verrelst et al., 2015b). Particularly, Gaussian processes regression (GPR) was evaluated as promising to serve as core retrieval algorithm (Verrelst et al., 2012b, 2015a). GPR is a machine learning regression algorithm framed in a Bayesian context, which has the attractive feature that it provides along with the variable estimates associated uncertainty estimates (Rasmussen and Williams, 2006).

Altogether, given the general framework sketched above, this work can be translated into the following main objectives: (1) to develop independent GPR models for an explicit quantification of LAI_G and LAI_B based on S2 reflectance data; (2) to implement these models in an automated processing chain, thereby delivering S2 image-based LAI_G,B maps; (3) to generate LAI_G and LAI_B time series and identify LAI seasonal trends over multiple crops types across several European study sites.

The content of this study is organized as follows: Section 2 provides the rationale and formulation of independent LAI_G and LAI_B estimation. Section 3 presents the methodology, describes the data used for developing and validating the models and outlines the processing chain. Section 4 evaluates the performances of the obtained LAI_G and LAI_B models and analyzes the spatiotemporal consistency of the LAI estimates by identifying patterns corresponding to crop senescence, harvesting and crop residues. In Section 5 the utility of the results for crop monitoring systems is discussed, and Section 6 finally concludes the LAI_G,B mapping approach.

2. Theoretical framework LAI_G and LAI_B mapping

In the VNIR, green vegetation is spectrally easily distinguishable from brown vegetation or bare soil due to the strong light absorption of chlorophyll leaf pigment, whereas in the SWIR, around 1400 nm and 1900 nm, strong decreases in the reflectance spectrum are caused by the leaf water content absorption (Kokaly et al., 2009). The challenge therefore rather lies in differentiate senescence vegetation from bare soil, which show a more similar and featureless spectral signatures in the VNIR, where most S2 bands are positioned. This can be observed in Fig. 1, where the reflectance spectra for these three land cover types are plotted along with the Sentinel-2A spectral bands detailed in Table 1. In the SWIR, a clear distinction primarily associated with the absorption of cellulose and lignin appears near 2100 nm and 2300 nm (Daughtry et al., 2004; Kokaly et al., 2009), which are not perceptible in the spectra of bare soil, especially at the region around Bl 2 (2190 nm), where dry vegetation presents some subtle absorption feature with respect to bare soil. The absorption effect of water content in green vegetation prevents the identification of these absorption features (Daughtry et al., 2004). The two SWIR bands, together with the other S2 bands, should thus enable to discriminate dry vegetation from bare soil. However, given the rather slight local decreases in the reflectance spectra of dry vegetation in the SWIR, only two bands in this spectral region may not suffice for an accurate characterization of LAI_B, especially in situations of sparse vegetation cover, when the soil spectral signal influence could hinder the correct estimation of dry plant matter (Serbin et al., 2009). That open question is yet to be resolved in this work.

Fig. 1.

Illustration of the spectral response function of S2A bands used in this study and the spectral signature of green vegetation (alfalfa), diy vegetation (barley) and bare soil as collected from the HyMap sensor during SPARC campaign (see Section 3.1). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Table 1. Sentinel-2 MSI band settings.

Band #	B1	B2	B3	B4	B5	B6	B7	B8	B8a	B9	B10	B11	B12
Band center (nm)	443	490	560	665	705	740	783	842	865	945	1375	1610	2190
Band width (nm)	20	65	35	30	15	15	20	115	20	20	30	90	180
Spatial resolution (m)	60	10	10	10	20	20	20	10	20	60	60	20	20

While the quantification of LAI_G and LAI_B has been earlier experimentally demonstrated by means of spectral indices (Delegido et al., 2015), the SENSAGRI project aimed to move away from parametric methods towards a more general machine learning method by exploiting the full spectral information of S2 data at 10 and 20 m. In view of operational mapping requirements, i.e. determined by both accuracy and run-time, GPR emerged as an excellent choice (Verrelst et al., 2015b), based on a probabilistic (Bayesian) approach for learning generic regression problems with kernels (Rasmussen and Williams, 2006). Its mathematical formulation is briefly outlined below.

2.1. Gaussian processes regression (GPR)

The GPR model establishes a relation between the input (B-bands spectra) x ∈ ℝ^B and the output variable (canopy parameter) y ∈ ℝ. Assuming that the observed variable y is formed by noisy observations of the true underlying function f(x), i.e. y = f(x) + ε, and that the noise ε is additive band-independent Gaussian distributed with zero mean and variance σ_n, the estimation of y provided by GPR can be expressed as:

$$\widehat{y}={\displaystyle \sum _{i=1}^N{\alpha }_{i}K(x_i,x)},$$ (1)

where ${\left\{x_i\right\}}_{i=1}^N$ are the training spectra, α_i ∈ ℝ is the weight assigned to each one of them, and K is the Kernel function evaluating the similarity between the test spectrum x and the generic training spectra x_i, i = 1,…,N. The kernel selected for this work is the standard scaled radial basis function (RBF), defined as:

$$K(x_i,x_j)=\nu \exp \left(-{\displaystyle \sum _{b=1}^B\frac{{(x_i^{(b)}-x_j^{(b)})}^2}{2{\sigma }_b^2}}\right)+{\delta }_{ij}{\sigma }_n^2,$$ (2)

where ν is a scaling factor, σ_b is a dedicated parameter controlling the spread of the training information for each particular spectral band b, σ_n is the noise standard deviation and δ_ij is the Kronecker’s symbol. The kernel is thus parametrized by signal (ν, σ, σ_n) and noise (σ_n) hyperparameters, collectively denoted as θ = {ν, σ,σ_n}, where σ=(σ₁, …,σ_B).

Let us define the stacked output training samples y = (y₁,..., y_N)^T, the covariance terms of the test point k∗ = (k(x∗,x₁), …,k(x∗,x_N))^T, and k∗∗ = k(x∗,x∗). For prediction purposes, the GPR is obtained by computing the posterior distribution over the unknown output y∗,p(y∗|x∗,D), where 𝒟 ≡ {x_n,y_n|n= l,…,N} is the training dataset. This posterior can be shown to be a Gaussian distribution, for which one can estimate the predictive mean (point-wise predictions):

$${\mu }_{y^{\ast }}=k_{\ast }^T{\left(K+{\sigma }_n^{2}I\right)}^{-1}y,$$ (3)

and the predictive variance (confidence intervals):

$${\sigma }_{y^{\ast }}^2=k_{\text{∗∗ }}-k_{\ast }^T(K+{\sigma }_n^{2}I{)}^{-1}{k}_{\ast }.$$ (4)

The corresponding hyperparameters θ are typically obtained by maximizing the marginal likelihood (also called evidence) of the observations. See Rasmussen and Williams (2006) for further details. A more detailed survey on GPR properties and alternative kernel versions is also provided in Camps-Valls et al. (2016).

On the application side, three important properties of the method are worth stressing here. First, the obtained weights α_i, after hyperparameter optimization gives the relevance of each training spectrum x_i. The predictive mean is essentially a weighted average of the canopy parameter values associated to the training samples: the closer to the test sample, the higher the weight. Second, the inverse of σ_b represents the relevance of band b. Intuitively, high values of σ_b mean that relations largely extend along that band hence suggesting a lower informative content. These features have been extensively studied in (Verrelst et al., 2012a, 2012b; Verrelst et al., 2013) and proved to be valuable for gaining insight in band relevance. In this respect, unlike other ML techniques such as neural networks (Bentez et al., 1997), with GPR the physical consistency of the functioning of the model can be assessed (Camps-Valls et al., 2016). Finally, and probably most interesting for mapping applications, the GPR model provides not only a per-pixel prediction, but also an uncertainty (or confidence) level for the prediction, which enables to evaluate the model transferability in space and time (Verrelst et al., 2013).

3. Material and methods

Statistical retrieval of any variable requires a training database of input-output data pairs and a validation dataset. Accordingly, the pursued approach is to train two distinct GPR models; one for LAI_G and one for LAI_B estimation. The input refers to spectral information corresponding to a certain LAI value as output. In contrast to similar LAI retrieval approaches with GPR (e.g. Campos-Tabemer et al., 2015; García-Haro et al., 2018) and earlier operational neural network models trained by simulated reflectance data generated by the turbid medium canopy RTM SAIL (Bacour et al., 2006; Verger et al., 2008, 2011), in this prototype version only experimental data and thus no simulated data have been used for model development. The reasons to do so are as follows: (1) although simulated training data may work well at lower spatial resolution (see developed models for SPOT/VGT, MODIS, AVHRR), relying on a turbid medium RTM faces limitations to represent heterogeneous surfaces at a higher spatial resolution (Widlowski et al., 2005), and thus undermines the validity of the retrieval model when it comes to image processing of heterogeneous landscapes (Widlowski et al., 2008); (2) simulation of LAI_B, i.e. simulation of brown leaves by the leaf optical model PROSPECT-5 coupled with SAIL, did not lead to a satisfactory description of senescent vegetation. The drawback of relying on experimental data, however, is that they may be insufficiently representative to achieve globally applicable models, as being constrained by the acquisition and crop state conditions (Weiss et al., 2020). Therefore, by striving to develop models as robust as possible to ensure their spatiotemporal extrapolation capacity, ground based measurements were collected from several field campaigns covering a wide range of vegetation types and thus variability of LAI_G and LAI_B.

3.1. LAI data collection

Using experimental data requires the spectral information to match the S2 band settings. Consequently, only ground measurements with either simultaneous S2 acquisitions or hyperspectral data covering all S2 spectral bands can be exploited. In the latter case, these spectral data were subsequently integrated to reconstruct the S2 band settings.

The data used for this study was partly obtained from the earlier ESA (European Space Agency) SPARC (2003) and DAISEX (1999) field campaigns carried out in Barrax agricultural site (Spain). During these campaigns, LAI was measured over a large variability of crops in vegetated and senescence state. Airbone hyperspectral data were acquired by the HyMap sensor. More details can be found in Delegido et al. (2013, 2015). At the same time, new campaigns were conducted in 2017 and 2018 during the SENSAGRI project. For the most part, LAI_G data collection covers diverse croplands at different European agricultural sites. To account also for forested areas, LAI measurements acquired over a spruce forest in eastern Czech Republic were included in the LAI_G training dataset. In contrast, as LAI_B has rarely been measured in past campaigns, LAI_B training dataset comprises only LAI measurements obtained from sampling fully senescent crops twice in Barrax and in Valladolid (Spain).

The common instruments used for performing the LAI measurements were LAI-2000 and LAI-2200 (LI-COR, Lincoln, NE, USA). Other commercial passive optical instruments include AccuPAR (Decagon Devices, Pullman, WA, USA) and imaging devices such as the digital hemispherical photography (DHP). The possible non-random positioning of the leaves or clumping and the contribution of woody materials were only experimentally corrected in the measurements conducted over forest (Homolová et al., 2007). Conventional optical methods are not capable of distinguish solely green plant elements from other compounds such as stems or senescent leaves (Jonckheere et al., 2004). Hence, for the remaining ground measurements used in this study, for which the mentioned effects were neglected, the term LAI refers to effective Plant Area Index (PAI) (Fang et al., 2014; Yan et al., 2019). Although the leaf surface is usually the most prominent physiological part visible to the sensor (Jonckheere et al., 2004; Hank et al., 2019), larger deviations of measured and observed LAI are expected during flowering or later crop growth stages, when the presence of nongreen plant elements is more noteworthy (Richter et al., 2012) and due to senescent leaves rolling and bending (Fang et al., 2014).

The sampling strategy followed on field was based on the VALERI protocol (Fernandes et al., 2014), where the extent of the Elementary Sampling Unit (ESU) is defined as the minimum area compatible with the resolution of the satellite product to be validated (30 m × 30 m). The objective of the sampling strategy is to locate the ESU in order to capture the variability across the study site, and to repeat measurements within the ESU to capture the variability within the sampled crop. Thus, following a systematic non-destructive sampling scheme, LAI was derived from 5 canopy measurements collected within each ESU, so that the final LAI value for each ESU corresponds to the mean value of these measurements. Over forest, measurements were taken every 2.5 m along two 50 m long perpendicular transects. For these field campaigns, LAI data were directly linked to S2 spectral information from the image closest to measurement date. All S2 images used for this purpose were resampled to 20 m spatial resolution; 60 m resolution spectral bands were discarded, i.e. Bl, B9 and B10 (see Table 1). Finally, each LAI model was trained with the remaining 10 bands. When available in the ESA Scientific Data Hub, S2 Level-2A products (BOA reflectance) were used. Otherwise, S2 corrected images were processed from Level-1C products (TOA reflectance) using the Sen2Cor tool (Louis et al., 2016). An overview of the LAI_G and LAI_B training datasets is given in Table 2.

Table 2. Overview of field campaigns for LAI_G and LAI_B collection used for training the GPR retrieval models.

Location	Date	#Points	Range	Instrument	Vegetation type	Spectral data
LAIG
Barrax, Spain	July 2003	102	0.4–6.2	LAI-2000	Alfalfa, corn, garlic, onion, potato, sugar beet, wheat	HyMap
Valencia, Spain	May/Nov 2017	34	0.41–5.41	LAI-2200	Alfalfa, artichoke, lettuce, onion, potato	S2
Biely Kríž, Czech Republic	August 2016	7	5.3–9.3	LAI-2200	Spruce forest	S2
Foggia, Italy	March 2017	6	3.08–4.23	LAI-2200	Wheat	S2
Poznan, Poland	July 2017	6	2.69–4.2	LAI-2200	Maize, triticale, wheat	S2
Kiev Oblast, Ukraine	June 2018	3	0.27–0.56	DHP	Maize, soybean	S2
Toulouse, France	August 2018	1	1.77	DHP	Maize	S2
LAIB
Barrax, Spain	June 1999/July 2003	13	2.16–3.74	LAI-2000	Barley, wheat	HyMap
Valladolid, Spain	June 2017/2019	38	0.38–4.01	LAI-2200	Barley, chickpeas, oat, rye, triticale, vetch, wheat	S2

As a final step, since S2 satellite imagery does not consist of vegetated surfaces alone, more realistic LAI maps can be achieved by adding spectra of non-vegetated samples to the training dataset (Verrelst et al., 2013). To do so, a set of multiple S2 cloud-free images from different locations were selected and analyzed for extracting spectral profiles of distinct non-vegetated pixels for which LAI_G and LAI_B values are set to 0. The extracted spectra, which correspond to bare soil, man-made surfaces such as build-up areas and roads, and water bodies, were added to both training databases. Because of the resemblance between dry vegetation and bare soil spectra, a higher portion of non-vegetated spectra had to be included in the LAI_B training database to avoid incorrect retrievals over this surface type. Specifically, the total number of data pairs used in the final training dataset were 218 for LAI_G (159 non-zero LAI and 59 zero LAI) and 94 for LAI_B (51 non-zero LAI and 43 zero LAI).

Field data kept for validation were collected in several field campaigns within the SENSAGRI project life, over Spain, France, Poland and Ukraine. The ground sampling methodology was the same explained previously. The validation dataset for LAI_G comprises similar crop types as used in the training dataset. Regarding LAI_B, main crops sampled were cereals, such as wheat, barely or oat, and rapeseed. Table 3 summarizes the validation data collection for both LAI_G and LAI_B.

Table 3. Overview of campaigns for LAI_G and LAI_B collection used for validation.

Location	Date	#Points	Range	Instrument	Vegetation type	Spectral data
LAIG
Toulouse, France	Nov 2017/March/May/Jul/Aug 2018	52	0.03–3.84	DHP	Maize, soybean, sunflower	S2
Poznan, Poland	Apr/Jun/Aug 2018	50	0.96–4.23	LAI-2200	Beetroot, maize, triticale, wheat	S2
Kiev Oblast, Ukraine	May/Jun/Aug 2018	40	0.04–4.81	DHP	Maize, soybean, sunflower, wheat	S2
LAIB
Valladolid, Spain	Jun-2019 - Jul-2019	40	0.8–3.61	AccuPAR	Barley, oat, rapeseed, wheat	S2

3.2. GPR models training

The GPR models’ development was carried out within an in-house developed MATLAB package named ARTMO (Automated Radiative Transfer Models Operator) (Verrelst et al., 2012c).¹ ARTMO embodies a suite of leaf and canopy radiative transfer models (RTMs) and several retrieval toolboxes, i.e. a spectral indices toolbox (Rivera et al., 2014), an inversion toolbox (Rivera et al., 2013), and a machine learning regression algorithms (MLRA) toolbox (Caicedo et al., 2014; Camps-Valls et al., 2013), which includes GPR and was used here.

Given the data-driven nature of GPR, multiple trainings were evaluated before reaching robust regression models. To do so, LAI_G and LAI_B final training datasets (Table 2) were previously iteratively extended and improved by adding data from multiple campaigns. This ensured that a large variability of vegetation types was covered by the training dataset, as well as consistent seasonal performances.

With the purpose of using the collected field data to the fullest, a 4-fold cross-validation (CV) sampling scheme was applied. The models performance was evaluated both in terms of accuracy and error by means of calculating several goodness-of-fit metrics, namely: Coefficient of determination (R²_cv) Mean Absolute Error (MAE_cv), Root Mean Square Error (RMSE_cv) and Normalized RMSE (NRMSE_cv (%))

3.3. Development of a processing chain

Within the framework of SENSAGRI, a processing chain was developed to automate the S2 image processing into LAI products based on imported GPR models. A general outline of the processing chain is displayed in Fig. 2. In short, S2 L2A images are downloaded from the Copernicus data hub and then passed as input to the processing chain. The following steps are consecutively executed: (1) loading only the S2 bands at 20 m, the atmospheric bands with 60 m resolution spectral bands were discarded, i.e. Bl, B9 and B10. All 10 m bands are by default also provided at 20 m, except the B8 band at 10 m (SPOT-5 heritage with a broad bandwidth of 115 nm) that needs to be upscaled. Although there is spectral overlap with B8a (at 865 nm) both bands are sensitive towards LAI and therefore included; (2) loading the scene classification image generated by Sen2Cor (Louis et al., 2016). Particularly, the shadowing and cloud classes are used as mask, so that, by default, the corresponding pixels are not processed for the retrieval, which greatly helps to speed up the overall processing. Likewise, other non-vegetated pixels such as water bodies can be excluded from processing; (3) importing the GPR models and applying them to the S2 processed image. Also, other processing parameters are set to avoid out-of-memory problems; (4) writing a multi-band LAI image in a specified output format at 20 m spatial resolution.

Fig. 2. Flowchart of retrieval processing chain; from a S2 L2A image to subsequent LAI product.

3.4. LAI product: Uncertainties, composite map and LAI_Total

An essential part of new-generation retrieval algorithms is the provision of uncertainty estimates (Section 2.1). Consequently, each GPR LAI model exploits this information by generating three output layers:

1. LAI estimate: mean estimates μ [m²/m²];

2. LAI SD: associated absolute uncertainties expressed as standard deviation (SD) [m²/m²] around μ;

3. LAI Cv: relative uncertainties expressed as coefficient of variation (Cv=SD/μ ∗100 [%]).

From the theoretical point of view, low relative uncertainty values indicate higher retrieval fidelity with respect to the available training information. The per-pixel generation of uncertainty estimates provides insight into the retrieval outcome throughout an image. Hence, the use of an uncertainty threshold enables the identification of pixels that lead to LAI estimates with acceptable quality. Because the associated uncertainty SD is expressed around the mean estimate, relative uncertainties, expressed in %, may provide a more intuitive interpretation than absolute uncertainty estimate (Verrelst et al., 2013). Thus, by applying a relative uncertainty threshold a map of LAI values fulfilling minimum accuracy requirements is provided. The advantage of this approach is that non-vegetated areas are usually automatically masked out, as the LAI models were hardly calibrated for those surfaces. Consequently, their close-to-zero LAI values lead to rather high relative uncertainties that are easily detected.

In addition, the multi-band LAI image includes an additional product called LAI_Total, which consists in the combination of LAI_G and LAI_B into a single LAI variable. Anticipating to the temporal evolution of the variable and since no information about the heterogeneity within a pixel is available, the LAI_Total value of a certain pixel is given by the highest and thus the expected predominant LAI estimation. Therefore, LAI_Total en-compasses in a single variable the whole phenological evolution of LAI and could serve as an alternative of existing LAI products, thereby avoiding underestimation of LAI over senescent crops.

After analyzing multiple maps over the selected study sites, the relative uncertainty applied for mapping purposes in this study is set to 40% for both LAI_G and LAI_B. Although merely serving for visualization purposes, this threshold strives to preserve a maximal coverage of vegetated surfaces: a lower threshold led to discarding vegetated areas, especially for LAI_B, which is more difficult to distinguish from bare soil. Subsequently, a synergistic LAI composite map can be created by representing the LAI_Total and using the uncertainty threshold to serve as mask, i.e. it only shows LAI_G,B that fall within the defined threshold, using green and brown colorscales for visualization (Fig. 3).

Fig. 3. Flowchart of LAI_G and LAI_B processing into a LAI_Total composite map.

3.5. S2 time series processing

As a final step, in order to analyze the spatial and temporal distribution of LAI retrievals over four selected European test sites, a collection of atmospherically corrected S2 images ranging from early 2017 to later 2018 was processed into LAI products according to Fig. 2. These test sites cover diverse European regions: Valladolid (Spain), Toulouse (France), Foggia (Italy) and Poznan (Poland). The amount of S2 images, their corresponding tile and the time span covered are detailed in Table 4.

Table 4. LAI product time series over the selected study sites.

Study site	# S2 Images	Tile	Period
Spain	66	T3OTUM	02/01/2017-29/09/2018
France	53	T31TCJ	17/12/2016-28/09/2018
Italy	59	T33TWF	09/12/2016-10/10/2018
Poland	53	T33UXT	01/01/2017-28/09/2018

To further assess the spatiotemporal LAI product consistency, the inter-comparison with an existing LAI product compatible in terms of spatial and temporal resolution was subsequently performed. ESA provides a S2 Toolbox integrated in the Sentinel Application Platform (SNAP) software. Among others, the toolbox includes the L2B biophysical processor, which can be used to map widely several canopy biophysical variables, including LAI (Vuolo et al., 2016). In SNAP, the LAI product can be derived at 20 m resolution from a L2A BOA reflectance product. The implemented retrieval algorithm is tailored for S2 and based on a neural network, which is trained with PROSPECT÷SAIL simulations. More details of the algorithm outline can be found in Weiss and Baret (2016). Accordingly, the same S2 images collection listed in Table 4 were processed into LAI using the L2B biophysical processor embedded in the SNAP software (LAI_SNAP).

4. Results

4.1. GPR models assessment

During the course of the SENSAGRI project, the GPR algorithm was retrained multiple times when more data became available, and the subsequent retrieval models tested. The CV goodness-of-fit results of the final GPR models are presented in Table 5. Both MAE_CV and RMSE_CV indicators show that the LAI_B associated error was lower than the one from LAI_G model. On the other hand, the NRMSE_CV, calculated to compare the performances across the two models, suggests a higher accuracy for LAI_G rather than LAI_B. Moreover, the LAI_G model achieves a better correlation, with a R²_CV close to 0.9.

Table 5. Cross-validation (CV) statistics of GPR models LAI_G and LAI_B training.

Model	MAECV (m2/ m2)	RMSECV (m2/ m2)	NRMSECV (%)	R2CV	biasCV
LAIG	0.45	0.66	7.12	0.89	−0.021
LAIB	0.39	0.55	13.75	0.76	0.062

The scatterplot of estimated against measured LAI values is shown in Fig. 4. In agreement with the error indices previously described, the deviation of model estimations from the l:l-line was more remarkable for LAI_B than LAI_G. In fact, apart from a few outliers, most of LAI_G estimates fell relatively close upon the 1:1 line, although estimates are inaccurately predicted at very high LAI values where the model tend to underestimate. Conversely, the correlation observed in LAI_B modelling indicates a less accurate performance, including more discrepancy and a greater dispersion. The LAI_B versus bare soil confusion can be observed in the LAI_B scatterplot, where for various samples bare soil is wrongly retrieved as LAI_B. It is worth noting that for both models few points are estimated as non-vegetated pixels, especially in the LAI_B model, where most likely low vegetated pixels are confused with bare soil.

Fig. 4.

Above: Measured vs. estimated LAI values resulting from the GPR training phase along the 1:1-line. Below: σ_b of the generated LAI_G (left) and LAI_B (right) GPR models. The lower the sigma the more relevant is the S2 band.

One of the advantages of GPR is that it provides insight in the contributions of the bands when developing the regression model. Differences between samples are given for each band typically using covariance function, where is related to the relevance of each band b. High values of σ suggest a lower informative content, whereas most relevant bands with a higher impact on the predictive function correspond to lower σ. A distribution of σ_b calculated for each GPR model is presented in Fig. 4.

The most relevant S2 bands (lowest σ_b) for the LAI_G model were those located in the visible (B2: 490 nm, B3: 560 nm, B4: 665 nm), the red edge (B5: 706 nm, B6: 740 nm, B7: 783 nm) and the NIR (B8a: 865 nm). Bands with little impact (high σ_b) are further in NIR (B8: 842 nm) and in the SWIR (Bl 2: 2100 nm). Nevertheless, that does not mean these bands do not contribute at all. In fact, any attempt to remove these bands worsened the accuracy results. Conversely, the distribution of σ_b, for the LAI_B model suggests that bands throughout the VNIR spectral range are needed for model prediction. Particularly, the bands playing a key role in the development of the LAI_B model are located in the visible (B2: 490, B4: 665 nm), red edge (B5: 706 nm, B6: 740 nm) and SWIR (Bl2: 2190 nm). The relevance of the SWIR band, as compared to the LAI_G model, confirms its importance for senescence vegetation detection. Although the remaining bands showed less relevance, also here, band removal did not lead to improved performances.

4.2. Direct validation

To assess their respective accuracy and robustness, the performances of the LAI_G and LAI_B retrieval models were evaluated under realistic conditions using independent validation data. The ESU LAI ground measurements were compared against the LAI_G and LAI_B values corresponding to the closest pixel to the centre coordinates of the sampled ESU. The dataset used for validation has been already described in Section 3.

Fig. 5 shows estimated against observed LAI values for both LAI_G and LAI_B. A linear fit was used to quantify the models’ predictive capability. In addition, a set of statistical indicators were also calculated, including error indices and correlation-based measures. For the French test site, LAI_G tends to overestimate the measured LAI, especially at low values, while at mid values, LAI estimation is closer to observed value. Over the Polish test site, for LAI values higher than 3, the retrievals mainly underestimate the corresponding in-situ reference data, leading to a worse correlation. Finally, Ukrainian validation results show that LAI_G generally overestimates the measured LAI, while highest values are not accurately retrieved.

Fig. 5.

Comparison of estimated LAI_G (left) and LAI_B (right) versus the LAI ground measured data considering all study sites. Dashed black line corresponds to 1:1 line and the solid black line represents the linear fit.

Overall, the LAI_G model was validated with an RMSE of 0.67 m²/m² and a correlation of R² = 0.70. Low LAI_G values are occasionally overestimated, mainly in the French site. From 2 to higher LAI_G values, points’ dispersion increases and therefore the model is less precise. The LAI_B model was validated using a single dataset from Spain with an RMSE of 0.43 m²/m² and a correlation of R² = 0.62. While at lower and higher LAI_B values the matching is good, more disagreement between estimates and measurements appears within the LAI range between 2 and 3. The error metrics suggest a superior performance of the LAI_B model as opposed to the LAI_G model, although less validation measurements were available and they came from only one study site. Yet the linear fit applied to both scatterplots reveal a similar correlation between estimated-measured LAI values for both retrieval models.

4.3. LAI_G and LAI_B mapping

For each study site, an area of interest was selected from cloud-free S2 images to generate a composite map that combines both LAI_G and LAI_B retrievals fulfilling the maximum allowed uncertainty threshold of 40%. The obtained maps are displayed in Fig. 6. The maps indicate the LAI mapping variability over time for agricultural sites along an increasing latitudinal gradient. LAI_G reaches at all locations a maximum cover in spring. In Northwest Spain most croplands, mainly cereal fields under rainfed conditions, are already senescent in late June and harvest time is approaching, whereas LAI_G pixels are mostly associated to newly planted crops or irrigated parcels. When moving to South France, a similar pattern can be observed: croplands reached the senescent stage during June. Conversely, in a Mediterranean region over Southeast Italy, the senescence particularly started already in May. By late spring, less croplands show pronounced senescence and fields with low LAI_B suggest they have already been harvested. In the Polish region LAI_G clearly predominates over LAI_B. Also this region is more forested, which explains the patches with high LAI_G > 5. Ripening and senescing took place later in time, so that golden-brown croplands are observed from July onwards. Overall, LAI_G is abundant in northern and in irrigated areas, while LAI_B is more pronounced in southern, dry agricultural areas. On all maps, water bodies, buildings, roads, and in general also bare soil areas and undetected cloudy pixels are automatically masked out as they fell outside the uncertainty threshold.

Fig. 6. Seasonal evolution of LAI_Total(LAI_G + LAI_B) over the different European study sites.

4.4. Time series analysis

For each study test site, several crop types were selected to illustrate the temporal evolution of LAI_G and LAI_B along one year. Fig. 7 shows the annual seasonal patterns observed in Fig. 6. The temporal profiles represent the mean value of LAI_G and LAI_B estimations computed for a selected crop area. The standard deviation is also plotted as vertical bars to provide a description of the variation of LAI values within the crop. LAI temporal gaps correspond to cloudy acquisitions. Missing LAI_G,B values correspond to exceptional unreliable GPR estimates such as negative values.

Fig. 7. Temporal evolution of several crop types over the different study sites described by LAI_G and LAI_B mean values of all pixels within the crop limits and the associated standard deviation, which is plotted as vertical bars.

The tracked LAI_G and LAI_B trends reproduce the typical phenological stages of winter crops, such as rapeseed and cereals (wheat, barley, rye). In this case, LAI_G increases progressively, reaching a maximum peak in spring. By that time also LAI_B emerges and senescence starts to take place, until LAI_B overtakes LAI_G by late summer. Abrupt drops of LAI_B suggest a harvest event (Fig. 7(a,b,c,h)). Occasionally, some traces of dry vegetation can remain longer, and therefore LAI_B is not completely extinct (Fig. 7(f)). Particularly, the appearance of green vegetation on some sites after summer indicates a new crop growing cycle. For example, the tomato crop in Italy shows two consecutive crop cycles. For irrigated summer crops, (Fig. 7(d,h)), LAI_G starts increasing from April on, with a maximum coverage around midsummer.

Concerning the general dynamics of LAI_B, it can be expected that crop senescence leads to a gradual conversion of LAI_G into LAI_B, meaning that LAI_B would stay on the same order as LAI_G. However, this was not observed because of the following two reasons. In the first place, during the senescence stage plant water content drops significantly, leading to leaves lower down in the canopy and thus causing LAI to decrease (Steduto and Hsiao, 1998; Inman-Bamber, 2004; Smit and Singels, 2006). In the second place, after reaching the maturity stage, the clumping index gradually declines, as related to the canopy development and crown thickening, resulting in an increased degree of foliage clumping and more pronounced gap fraction (Ryu et al., 2012; Pisek et al., 2013; Fang et al., 2014).

The same analysis was conducted for several crop types on an agricultural region in Valladolid (Spain), adding moreover available information about harvest date, in order to explicitly identify the LAI transition from green to senescent vegetation (Fig. 8). It can be observed that both LAI_G and LAI_B are able to capture the expected LAI seasonal variation of annual winter crops. After reaching the typical LAI_G peak value, the senescence period starts until crops are finally harvested throughout summer, after most of crop vegetation has turned into dry vegetation. Summer crops, such as sunflower (Fig. 8(g)), exhibit the same LAI_G and LAI_B trends but from July onwards. Instead, being an irrigated summer crop, beetroot (Fig. 8(e)) keeps a regular LAI_G coverage over most of time and is harvested by the end of the year. Altogether, for all crop types it can be observed that the harvest date agrees with a subsequent sudden decrease of the LAI type more predominant in that moment, either LAI_G or LAI_B. In case of winter crops, harvest takes place when LAI_B is maximum, or shortly after this moment.

Fig. 8.

Seasonal variation of LAI_G and LAI_B of different crop types over the region of Valladolid (Spain). Solid lines represent the mean value of all pixels within the crop limits, while the vertical bars show the associated standaid deviation. Finally, the dashed line represents the haivest date.

4.5. Product inter-comparison: LAI_Total and LAI_SNAP

Both LAI_Total and LAI_SNAP products were compared over several crop types across the four study sites by examining a time series, covering a 2 years period from 2017 to 2018 (Fig. 9). For each crop type, the respective LAI_Total and LAI_SNAP mean and the associated standard deviation of all pixels within the crop limits were computed. In case of LAI_Total each pixel was assigned the corresponding LAI_G or LAI_B value, according to the criterion explained in Section 3.4. When computing the LAI_SNAP mean value, those pixels labeled as not reliable by the associated quality flag band were discarded.

Fig. 9.

Temporal evolution of several crop types across four study sites described by LAI_Total and LAI_SNAP (crop mean value and standard deviation, which is plotted as vertical bars). Shaded areas and crop type correspond to a time series period previously represented in Figs. 7 and 8.

Overall, the LAI temporal distribution is relatively consistent between the two algorithms, as both show a realistic temporal LAI dynamic and their respective estimations are close. Also, no systematic under- or overestimation of one model with respect to the other one is observed. Initially, crop seasonal evolution is mostly in agreement, reaching about the same time the peaks of season. Yet, some discrepancies are observed at the onset of the senescence period (Fig. 9(a,c,f,g,h)). Because LAI_SNAP refers only to green LAI, it seems not able to capture LAI of senescent vegetation, contrary to LAI_Total, which assimilates LAI_B too. Temporal discontinuity and strong variability (sharp peaks), are generally coincident for both products, presumably due to residual errors (atmospheric correction, geocoding, etc.) from S2 reflectance information. As a final remark, it is also worth noting that the LAI_SNAP pattern exhibits a generally noisier evolution as compared to LAI_Total, which automatically masks out clouds (Fig. 9(g,h)).

5. Discussion

This work presented a prototype processing chain to map realistically LAI of senescing croplands with potential for operational usage. The processing chain consists of two GPR retrieval models that explicitly quantify LAI over photosynthetically active vegetation (LAI_G) and senescent vegetation (LAI_B) from S2 images. The key aspects and limitations of the developed method are now discussed, with emphasis on: (1) LAI_G find LAI_B GPR models training, (2) models validation, and (3) S2 imagery spatiotemporal processing. Finally, suggestions for improved LAI_B mapping beyond S2 are given.

5.1. LAI_G and LAI_B GPR model training

Just as for any machine learning regression algorithm, the performances of the two GPR models depend entirely on the quality of the training data, which makes this step of utmost importance (Verrelst et al., 2012b, 2015a). To make the models generally applicable, data from several field campaigns and vegetation types were brought together. In the case of LAI_B, less reference data were available and due to the spectral similarity with bare soil, the proportion of non-vegetated LAI values included in the model training was higher in LAI_B than in LAI_G database. Otherwise, more bare soil pixel and dark areas were mapped as LAI_B. For instance, CV results of the LAI_B model indicated some overestimation of retrieved low-mid LAI, meaning in some cases bare soil is retrieved as senescent vegetation. Partly, this can be attributed by the limited training dataset for senescent vegetation; very few historic campaigns focused on brown vegetation and moreover on few crop types only. Alternatively, it was explored whether senescent leaf spectra can be generated by PROSPECT-5 or later versions (Feret et al., 2008). The model parametrizes the brown pigments or denatured proteins that appear in senescent and dry leaves inducing light absorption. Ideally, variation of this particular variable allows to simulate from fully green to fully brown leaves (e.g. Houborg et al., 2009; Danner et al., 2019), however, the input variable brown pigments is not expressed in quantifiable units and is thus not measurable by field instruments (Danner et al., 2019). Also, simulations of senescent canopies by coupling the leaf model with SAIL were unsatisfactory, and were ineffective for training the LAI_B model. Another point is that ideally the model should be able to properly quantify senescent vegetation given all types of soil surfaces, e.g. varying soil types and soil moisture. Although it is possible to simulate such variability with canopy models such as SAIL, it makes that the amount of simulations can become overly big, which can deteriorate the performance of the retrieval model. When relying on simulations, a well-balanced sampling scheme is required (Verrelst et al., 2016a). Exploiting experimental data, when available, seems therefore more straightforward. However, when aiming to further improve the LAI_B algorithm, it is recommended to incorporate more representative data into the training dataset, thereby taking into account a greater variability of not only crop types at multiple stages of senescing, but also variability of soil moisture or soil type (Quemada et al., 2018).

An appealing aspect of GPR model training is the provision of band ranking, as it gives some physical meaning to the developed models. The S2 bands positioned in the red edge and SWIR were ranked as more relevant for spectrally distinguishing dry vegetation from bare soil. Particularly the SWIR band Bl2 (2190 nm) is justified by the subtle absorption feature at 2100-2300 nm in dry plants caused by lignin and cellulose content (Kokaly et al., 2009), and its exploitation to identify brown vegetation was suggested before (Delegido et al., 2015). In contrast, the Bl 1 band (1610 nm) was evaluated with the least relevance on the LAI_B model, meaning this band is poorly sensitive to brown vegetation. Regarding LAI_G it was no surprise that the VNIR spectral region with the red edge bands plays an important role in the model development. The VNIR has long been successfully exploited for LAI_G retrieval (Nguy-Robertson et al., 2014), including S2-based LAI studies (Pasqualotto et al., 2019; Xie et al., 2019). Given the data-driven nature of GPR, however, it must be remarked that the band ranking strongly depends on the quality of the training data, and is thus not free from anomalies (Verrelst et al., 2016b).

Apart from the training data, another aspect that can lead to improved performances is on the design of the GPR algorithm. The following modifications are worth exploring: (1) introducing alternative kernel versions. A standard RBF kernel was used, but alternative kernel flavours may be worth to explore (Camps-Valls et al., 2016). (2) An attractive alternative is to develop multi-output GPR models (García-Haro et al., 2018; Pipia et al., 2019). With multi-output models various variables can be predicted with the same model, thereby taking covariance relationships into account. Even though earlier efforts with neural networks did not demonstrate improvements as opposed to single-output models (Bacour et al., 2006; Baret et al., 2007), it can be interesting to explore the multi-output alternative given that LAI_G and LAI_B are strongly related. Nevertheless, it should be noted that then for the same spectra both LAI_G and LAI_B samples need to be collected, which is currently not the case, and may be impossible to measure at the same time. (3) Apart from the standard GPR formulation, various alternative versions have been recently developed, including heteroscedastic GPR (Lazaro-Gredilla et al., 2014), which tends to lead to lower uncertainties, because of additionally exploiting the noise in the data. See also Camps-Valls et al. (2016) for a review on these GPR developments.

5.2. LAI_G and LAI_B GPR model validation

Optimized LAI_G and LAI_B GPR models were subsequently validated against ground data across various European sites. Validation results of the LAI_G model manifest generally a good agreement across the whole range of represented LAI values and over cropland across three distinct study sites, with a subtle tendency of underestimation of highest values, due to saturation of LAI_G around 4–5, similarly as observed in Camacho et al. (2013); Houborg et al. (2015); Xiao et al. (2016). It should be noted that only a few LAI values corresponding to coniferous forest were included in the training dataset and consequently, further validation is required to determine the LAI_G model validity over both coniferous and deciduous forests. Likewise, the LAI_B validation data faced limitations because of a reduced number samples; coming from a single study site. In this respect, future campaigns are encouraged to collect LAI of senenscent vegetation, although in contrary to LAI_G, for LAI_B sampling protocols are yet to be established.

Regarding field data collection, even though the same sampled scheme was systematically applied, it is worth noting that ground data used were taken using different instruments, which may contribute to instrument-specific measurement deviations. Consistency between instruments depends on the vegetation type, illumination conditions and LAI measured range (Garrigues et al., 2008). Over spatially homogeneous canopies, both LAI-2000 and LAI-2200 can provide reliable and closer estimates to true LAI (Ryu et al., 2010; Fang et al., 2014). At the same time, some studies found a good agreement between LAI-2000 or LAI-2200 and DHP measurements (e.g. Jonckheere et al., 2004; Fang et al., 2014). In contrast, AccuPAR is known to underestimate in comparison to the latter instalments (Fang et al., 2014). Additionally, larger transmittance of senescence leaves may result in lower LAI estimates from LAI-2000 and AccuPAR (Garrigues et al., 2008), which could explain the more remarkable overestimation noticed in LAI_B model as compared to LAI_G (Fig. 5), since the training data of the former were exclusively collected with LAI-2000 and LAI-2200. Though the ESUs are aimed to be located in homogeneous crop fields, within-crop variability conditions and other registrations errors might cause some mismatch between the actual LAI measurement and the corresponding image pixel. Nevertheless, the overall uncertainty and accuracy metrics, resulting from a pixelwise comparison, are within the range of other similar studies of green LAI products derived from medium to coarse resolution satellites (Camacho et al., 2013; Li et al., 2015; Pasqualotto et al., 2019).

5.3. LAI_G and LAI_B S2 imagery spatiotemporal processing

The developed GPR models were eventually used to map simultaneously LAI_G and LAI_B at European scale, performing consistently in a variety of agricultural landscapes, suggesting a good model portability to different regions from the ones where training data were collected. Spatiotemporal analysis of LAI_G and LAI_B showed realistic temporal profiles, in terms of seasonality and range. Crop temporal profiles proved to be able describing the LAI yearly evolution as well as the transition from green to senescent vegetation, at a S2 highest spatial resolution. It was also possible to illustrate the life cycle of multiple crop types along a wide dynamic LAI range, as this variable is related to the phenological stages of vegetation. Nevertheless, the LAI_B model still faces some difficulties in distinguishing properly true senescent vegetation pixels from bare soil areas and fallow land, as sometimes LAI_B is estimated long after harvest time and without any previous green vegetation quantification.

When computing the LAI_Total during the transition stage of vegetation browning, independent LAI_G and LAI_B estimations are obtained for the same pixel. Considering that each model is optimized for canopies of exclusively green or senescent leaves, assigning one of the two estimations undoubtedly lead to an underestimation in quantification of the actual amount of present vegetation. LAI_Totalrepresents either LAI_G or LAI_B without providing the real contribution of the intermixing of green and senescent leaf material. Yet, the underestimation is aimed to be minimized by favoring the the highest calculated LAI, which leads to a meaningful temporal pattern. Another approach to tackle the browning during crop ripening was by Houborg et al. (2015). The authors developed a regularized physicallly-based model to retrieve total LAI from Landsat multispectral sensor reflectances, where leaf spectra for green and senescent were simulated using the PROSPECT model. Intermixed green and senescent leaves reflectance was then obtained weighting the respective spectra with an input parameter, the canopy fraction of senescent leaves. This approach allowed to describe a growing season time series of green LAI over maize and soybean fields, along with the fraction of senescent vegetation parameter. Alternatively, Houborg et al. (2009) contrasted the spatial variation of the canopy fraction of senescent material with the spatial pattern of leaf chlorophyll content. Despite providing valuable canopy status information, the inversion approaches used in those studies, however, rely on ancillary data constraints, scene specific LUT generation, solely used standard broad spectral bands (green, red and near-infrared) and were only validated with green LAI in-situ measurements.

Conversely, the here proposed method provides LAI pixelwise retrievals through an image-based automated processing chain, following multiple consecutive steps without requiring any additional scene information or iterative processing; as the retrieval models are already trained and are subsequently applied to higher spectral resolution S2 images. When comparing time series of LAI_Total with the analogous S2 derived LAI_SNAP both temporal profiles revealed the expected plant development stages (green-up, peak LAI and senescence) without any larger discrepancies for most of the crops analyzed, except some pronounced differences observed in Fig. 9(b,h), where LAI_SNAP surpasses LAIrotαi during the vegetative stage. After crop maturation, LAI_SNAP decays as it only represents green vegetation, while LAI_Total integrates LAI_B estimations and is able to quantify the amount of senescence vegetation remaining until harvesting or falling off. Moreover, LAI_Total proved to be more stable and smoother than the LAI_SNAP product as with the inclusion of the uncertainty threshold some form of cloud masking is part of the retrieval processing chain. Altogether, optimizing the simultaneous quantification of LAI_G and LAI_B thus represents a significant improvement as opposed to the conventional LAI_G products. The explicit separation of LAI_G and LAI_B can eventually become an important remote sensing vegetation product since quantification of LAI_B can play an important role in detecting crop stress and the onset of drought.

5.4. Suggestions for improved LAI_B mapping beyond S2

Finally, although this work demonstrated the first streamlined production of LAI_B from S2 images, it must be remarked that quantification of senescent vegetation remains a challenging exercise. Not only because of little LAI_B data available for training and validation, but more importantly because of the spectral similarity between LAI_B and bare soil (Delegido et al., 2015; Kokaly et al., 2009). Earlier studies proposed spectral indices for distinguishing senescent vegetation from bare soil by using bands from the SWIR domain available in hyperspectral data (Delegido et al., 2015; Danner et al., 2017). Alternatively, the fractional covers of non-photosynthetic vegetation and bare soil can also be obtained from optical sensors through the linear spectral unmixing approach (Guerschman et al., 2015; Li and Guo, 2016). This method performance is restricted to the spectral resolution of the sensor to differentiate the spectra of each cover type and still seasonal changes, when green and senescent material concur, result in a lower estimations accuracy (Wang et al., 2019). This study demonstrated that also the S2 multispectral sensor is able to discriminate photosynthetic vegetation, non-photosynthetic vegetation and bare soil thanks to their bands in the red edge and SWIR. S2 is arguably the best suited Land mission from what is currently available to map LAI_B of individual fields in an operation context. Yet, the GPR band ranking suggested that the Bll band (1610 nm) appeared to be poorly positioned or too broad (90 nm) to take part in discriminating vegetation brownness from bare soil. In this respect, given the specific absorption features of cellulose and lignin in the SWIR (Curran, 1989; Kokaly et al., 2009), as well the spectral responses in the water absorption regions due to a reduction of leaf water content related to plant sensescing (Elvidge, 1990; Nagler et al., 2000), it has been argued that more bands are required for a proper quantification of LAI_B (e.g. Li and Guo, 2016; Hank et al., 2019). Evidently, the subtle spectral differences between both land cover types can be better captured with a multitude of narrow bands in the VNIR and SWIR domains, and there is no doubt that LAI_B can be quantified with improved accuracy with the upcoming operational imaging spectrometer missions. Promising new-generation imaging spectrometer missions include PRISMA (Loizzo et al., 2018), EnMAP (Guanter et al., 2015), SBG (Hulley et al., 2019), and the ESA high-priority candidate mission CHIME (Nieke and Rast, 2018). These missions cover the full VNIR and SWIR range with contiguous narrow bands and will deliver a spatial resolution similar as S2 (30-60 m). It is expected that these hyper-spectral missions will enable an improved derivation of crop and soil variables as well as allowing a more precise quantification of the main crop characteristics (Verrelst et al., 2019; Berger et al., 2020), this giving the potential to improve farm management and field productivity.

6. Conclusions

Thanks to the Sentinel-2 (S2) bands in VNIR and the red edge (B4: 665 nm, B5: 705 nm, B6: 740 nm) and in the shortwave infrared (Bl 2: 2190 nm) an explicit spectral distinction between photosynthetic (green) and senescent (brown) vegetation can be achieved, and thus LAI for both vegetation types can be quantified. Two GPR models, one for LAI_G and another for LAI_B, solely trained with experimental data, have been successfully implemented into a processing chain for independent LAI retrieval from S2 imagery. Both models were quantitatively assessed across multiple European study sites through direct validation using in-situ data. Both LAI_G and LAI_B estimates were merged into a single LAI_Total product, which ensures that LAI over senescent vegetation is not underestimated. Associated uncertainty estimates were used as a mask to produce composite maps where only estimates with sufficient confidence (below 40% uncertainty) are represented. A comparison of LAI_Total with LAI derived from the S2 processor revealed reliable and consistent spatial and temporal distribution of retrievals. The robustness and portability of the trained GPR models were positively evaluated by means of mapping LAI_Total composites over multiple European regions of Spain, France, Italy and Poland. All produced maps show a consistent performance. Also, multiple crop types across the European test sits were selected to illustrate annual temporal performance of the GPR models, suggesting a consistent LAI_G and LAI_B seasonal evolution. With the simultaneous production of LAI_G and LAI_B layers, S2 data has been exploited to the fullest for agroecosystem monitoring applications. Being fully automated and image-based, the developed retrieval scheme has potential for operational usage if implemented into a cloud computing facility, which would be more appropriate for large-scale mapping operations. Routine delivery of these products opens opportunities to monitor crop ripening and senescing, harvest time and crop residues. These are key applications of remote sensing for agriculture, which could contribute to monitor sustainable agricultural production and safeguarding the future of food security.