Insuring Alpine Grasslands against Drought-Related Yield Losses Using Sentinel-2 Satellite Data

Abstract

This work estimates yield losses due to drought events in the mountain grasslands in north-eastern Italy, laying the groundwork for index-based insurance. Given the high correlation between the leaf area index (LAI) and grassland yield, we exploit the LAI as a proxy for yield. We estimate the LAI by using the Sentinel-2 biophysical processor and compare different gap-filling methods, including time series interpolation and fusion with Sentinel-1 SAR data. We derive the grassland production index (GPI) as the growing season cumulate of the daily product between the LAI and a meteorological water stress coefficient. Finally, we calculate the drought index as an anomaly of the GPI. The validation of the Sentinel-2 LAI with ground measurements showed an RMSE of 0.92 [m² m⁻²] and an R² of 0.81 over all the measurement sites. A comparison between the GPI and yield showed, on average, an R² of 0.56 at the pixel scale and an R² of 0.74 at the parcel scale. The developed prototype GPI index was used at the end of the growing season of the year 2022 to calculate the payments of an experimental insurance scheme which was proposed to a group of farmers in Trentino-South Tyrol.

1. Introduction

The mountain grasslands in the European Alps have an important economic function because they are the major source of forage for livestock farming and a place of recreation for tourism [1,2]. In addition, they play a crucial role in climate regulation, biodiversity safeguarding, landscape conservation, soil quality preservation, and soil stabilization and protection from erosion [3,4,5].

Alpine grasslands are more exposed to climate change compared to other parts of Europe [2,6,7,8]. Despite water being abundant in the past in the Alps, its scarcity has started to raise concerns because drought events are becoming more and more frequent [9,10,11,12], mainly altering grassland productivity [13] and thus endangering the income stability of mountain farmers. According to climate observations, the mean annual temperature in the Alps has increased by about 1.8 °C since 1880, which is almost twice that of the mean of global warming, with strong acceleration over recent decades [14,15]. Regional climate projections predict a further temperature increase by the end of the 21st century of up to +4 °C with respect to 1981–2010 under the worst-case emission scenario [15,16]. Precipitation, in contrast, is predicted to decrease in summer, especially in the southern Alps, and increase in winter during the 21st century [15].

These changes in climate will not only cause changes in species composition and warming-related upward shifts in many grass communities [17] but will also affect the productivity of grasslands [18] and their predictability, with more economic instability and possible losses for the farmers [19]. This is especially likely in the case of combined extreme events like drought and heat waves [20]. Accordingly, it is urgent to plan the necessary strategies and adaptation pathways.

Risk management instruments, like insurance, can help grassland-based mountain agricultural systems to overcome production shortcomings and thus increase their economic robustness and resilience, prevent land abandonment, and maintain their functioning over time. Traditional insurance schemes, such as the so-called indemnity insurance, require physical inspections by insurance appraisers to assess the damage, but this approach is not economically sustainable for grassland due to its high cost compared to the low value of the production. In addition, it is difficult to obtain reliable site-specific reference values for the expected normal yield to estimate forage losses because the yield of a specific year depends on management (e.g., fertilization and cut frequency), which can differ from that of the previous year. Furthermore, damage evaluations based on physical inspections are subjective as they depend on an expert’s experience and can lead to unfair assessments. Index-based insurances are not exposed to these issues because the payoffs depend on an index related to grassland production that does not require physical checks and provides site-specific reference values for yield over time. In this context, freely available high-resolution satellite data from the Sentinel constellations of the Copernicus Program allow for the development of accurate and low-cost tools to support risk management. Despite the fact that some high-resolution large-scale products are now freely available, for example, the Copernicus land monitoring service [21], they lack the adaptation of processing chains to local conditions, including land use, topography, meteorology, and near-real-time production and dissemination, which are crucial to facilitate the exploitation of satellite-based products for agricultural insurance. In addition, they are not developed considering the needs and knowledge of the local agricultural stakeholders, who are, in our case study, the final users of such products.

It is worth mentioning that index insurances based on satellite data are subject to basis risk [22] due to the possible discrepancy between the index-derived payout and the actual losses. Basis risk can hinder the attractiveness of these new insurance schemes for the farmers or the insurance companies. One important challenge to minimizing the basis risk associated with index insurance is the proper selection of the parameters that are related to drought occurrence and drought impact. Currently, operational index insurances for pasture and rangeland use a parameter for estimating yield loss greenness indices, like the normalized difference vegetation index (NDVI) [23,24,25], or meteorological drought indices derived from weather station data [26,27]. These parameters present some criticalities. Regarding NDVI, it can be affected by soil effects, geometry of illumination, and atmospheric conditions. Furthermore, it is subjected to saturation under high production values [28]. Regarding weather station data, they often lack the spatial density that is needed to correctly represent extreme weather conditions like drought. In addition, the impact of meteorological drought conditions on yield depends on land management, topography, and soil type.

In this work, we develop an index of grassland production (grassland production index: GPI) based on the leaf area index (LAI), a biophysical parameter that is directly related to grassland yield. In the formulation of the GPI, we include a coefficient of water stress derived from gridded meteorological data at a high spatial resolution. The aim is to lay the foundation for index-based insurance for mountain grasslands, in which the insurance payments depend on yield losses estimated from the GPI anomalies at the parcel scale. One noteworthy part of the work is the involvement of the local agricultural stakeholders in the design of the index.

The idea of an index of grassland production based on biophysical variables derived from satellite data was originally proposed by Roumiguié et al. [29,30], who developed a forage production index (FPI) using fractional cover (fCover) estimated from moderate-resolution satellite data. The authors also showed [31] that the inclusion of a coefficient for quantifying water stress allowed them to improve the performance of the index. In contrast to FPI, we used the LAI instead of the other variables related to grassland productivity used for FPI, namely fCover and the fraction of absorbed photosynthetically active radiation (fAPAR). This was carried out in order to be able to ensure the availability of ground measurements to assess how accurately we can derive the LAI from Sentinel-2 and to study the relationship between the LAI and yield.

In order to verify that the S2 LAI is a reliable parameter for the derivation of index-based insurance, we performed a systematic validation of the LAI and an evaluation of the derived GPI. To this aim, we collected ground measurements for the LAI and yield at selected plots within grassland farms, which are representative of the target of the new insurance scheme under development.

We assume that the LAI is a valid parameter for estimating the impact of drought on grassland productivity, considering the high correlation that we observed between the LAI and yield and the simple retrieval algorithms, as compared to other variables related to the carbon and water cycle, like biomass and evapotranspiration. The LAI can be estimated from Sentinel-2A/B (S2) multispectral data at a spatial resolution of up to 10 m. This resolution allows for a proper characterization of the management practices and of the vegetation conditions within the single parcels [32,33,34,35].

On the other side, the LAI from S2 might not be sufficient enough to characterize the temporal evolution of vegetation because there is an S2 overpass every 5 days and because, within each acquisition, entire images or portions of them can be unusable due to the presence of clouds. It is thus necessary to implement gap-filling methods to increase the frequency of the available LAI estimates [36,37]. Here, we investigate different gap-filling strategies.

After providing a description of the study area and of the ground data collection (Section 2), we summarize the methods used to derive the GPI and its anomalies (Section 3.1) and to evaluate the LAI and GPI against the ground measurements of the LAI and yield (Section 3.2). In Section 3.3., we describe the fusion experiment between S1 SAR and S2 optical data. The results of the comparison between the in situ LAI and yield (Section 4.1) and the validation of the S2 LAI (Section 4.2) support the choice of the LAI as a proxy of grassland productivity, while the comparison between different LAI interpolation techniques (Section 4.3) justifies the choice of the method adopted in the first version of the GPI. The comparison between the GPI and the yield measurements helps to identify the main factors affecting the ability of the GPI to predict yield variations (Section 4.4). The yield losses estimated for the year 2022 from the GPI anomalies are summarized in Section 4.5. The preliminary results of the S1–S2 data fusion (Section 4.6) set the basis for the future enhancement of the GPI. The strengths and limitations of the proposed index are discussed in Section 5.

2. Study Area and Ground Data Collection

2.1. Study Area

The region Trentino-South Tyrol, including the autonomous provinces of Trento and Bolzano/Bozen, is located in north-eastern Italy and covers an area of 13,605 km². Its mountainous territory, covering a large portion of the Dolomites and southern Alps, is characterized by strong topographic complexity. Its climate is highly variable, being influenced by humid air masses from the Atlantic, dry air masses from the continental east, and warm air from the Mediterranean region [38], and it is characterized by cold and humid winters, and warm and dry summers.

In South Tyrol, 20.83% of the area is covered by agricultural areas [39], 70.02% of which is covered by grasslands [40]. Similarly, 15.32% of the area of Trentino is used as an agricultural area [39], 83.06% of which consists of meadows and pastures [40].

The index-based insurance, which is the target of the present study, is potentially applicable to all the meadows primarily aimed at forage production and natural pastures, which cover around 1870 km² (Figure 1). The management intensity of meadows in this region mostly ranges between one and four cuts per year. They are mainly fertilized with organic manures (farmyard manure combined with manure effluent, slurry, and biogas slurry). The botanical composition varies according to the site characteristics and management intensity [41,42].

2.2. Ground Data Collection

The main criteria for test site selection were (i) to be representative of the most frequent situations in the local agriculture being the subject of the new insurance system, (ii) to minimize logistic efforts for sampling, and (iii) to guarantee good reliability of the communication with the farm owners.

The chosen test sites are eight differently managed meadows at four farms located in three different municipalities: Ritten (R1: 46.5309473N, 11.432009E; R2: 46.5290155N, 11.4327394E; R3: 46.5340003N, 11.4011517E; R4: 46.5413533N, 11.4057968E), Laurein (L1: 46.4536094N, 11.0790985E; L2: 46.4551238N, 11.0754604E) and Fondo (F1: 46.4336386N, 11.1381509E; F2: 46.4290127N, 11.1324068E) (Figure 2,

Table 1. Information on the sites where the ground measurements were collected.

Site	Parcels Codes	Elevation [m asl]	Years	Type of Measurement	Frequency of Measurement
Laurein/Lauregno (BZ)	L1, L2	1330–1340	2021	LAI and yield	Every 1–2 weeks
Ritten/Renon (BZ)	R1, R2, R3, R4	1250–1270	2021	LAI and yield	Every 1–2 weeks
Muntatschinig/ Monteschino (BZ)	V1, P2	1490, 1549	2017–2021	LAI	1 to 3 per month
Fondo (TN)	F1, F2	970	2021	LAI and yield	Every 1–2 weeks

). They cover an elevation range of about 600 m (970 to 1550 m asl.), have a cut frequency of two to four cuts per year, with additional grazing on the last regrowth, and total N-inputs ranging between 112 and 189 kg ha⁻¹ year⁻¹, exclusively provided by means of organic manures. Information about the botanical composition was obtained prior to the first cut by means of visual estimates of the yield proportion [43] of species and of the functional groups of grasses (also including graminoids), legumes, and forbs in two sampling areas of 5 × 5 m in each meadow (mean values are reported). The investigated parcels are extremely to quite poor in species (9 to 19 species), except for L1 and L2 (27 and 24 species, respectively). All the meadows are rich in grasses (69 to 99% yield proportion), with the dominant species being typical of intensive management, such as Alopecurus pratensis, Dactylis glomerata, Lolium perenne, and Elymus repens (see Table S1).

The sampling campaign started on the 9 April 2021 and continued until the 29 October 2021. At each sampling event and for each sampling site, four plots corresponding to different pixels of the S2 grid were sampled. In each plot, three LAI values were collected by a LAI 2200C Plant Canopy Analyzer (LI-COR, Lincoln, NE, USA) [44]. The lens of the LAI 2200C was partially disclosed with a 90° view cap to reduce the potential disturbance that could derive from the operator holding the instrument. Moreover, the lens was held in the shadow of the operator to keep the light conditions as stable as possible during each measurement cycle. Each LAI value consisted of one measurement above the vegetation canopy and five measurements below the canopy just above the ground level, set apart by some centimeters from one another. The lens was held parallel to the ground. Afterward, one grass sample was taken in each plot by means of electric scissors within a metal frame of 0.25 m² at a stubble height of 4–5 cm in order to obtain a realistic estimate of the forage yield. The respective LAI value had been previously obtained within the same area that was going to be harvested afterward. The samples were dried at 60 °C until the weight reached constancy, and they were weighted to determine the dry matter yield. In addition, complementary visual assessments were taken of the lodging (estimated on a four-level scale: no lodging, light lodging, medium lodging, and heavy lodging).

The samples were collected every two weeks for the first growth cycle until the beginning of stem elongation and then once per week until the first cut, taking care to perform the last measurement as close as possible to the mowing event. For the regrowth, sampling occurred about every two weeks, taking care also, in this case, to perform the last measurement as close as possible to the mowing event. Over the whole 2021 field campaign season, 1615 LAI measurements and 558 yield samples were taken.

In addition to the field campaigns described above, a further dataset was used for LAI validation, including a time series of the LAI collected at the long-term socio-ecological research site installed by the Institute for Alpine Environment of Eurac Research at Muntatschinig in the Vinschgau valley in the Autonomous Province of Bolzano/Bozen (Muntatschinig site, Figure 2 and Table 1). At this site, a meadow (V1: 46.6864155N, 10.579762E) and a pasture (P2: 46.68486167N, 10.58519267E) parcel were sampled from 2017 to 2021 throughout the vegetation-growing season. Multi-annual observations of the biophysical variables are important to understand the dynamics of vegetation growth regarding two sites with very different forms of management [45,46]. The meadow is mown twice a year (end of June and the beginning of September), and its dominant species is Trisetum flavescens. The pasture is lightly grazed by livestock from mid-June to mid-October, and its dominant species is Festuca valesiaca. At each measurement event, three replicates of the LAI measurements were acquired by LAI2200C at three to five plots corresponding to the different pixels of the S2 grid. The sampling campaign over the whole period resulted in 910 LAI measurements.

3. Methods

3.1. Estimation of Yield Losses for the Year 2022

3.1.1. Estimation of the LAI from S2

The LAI was estimated by the SNAP S2 Toolbox Biophysical processor (s2tbx) by using the Graph Processing Tool (GPT) for bulk processing [47,48] (Figure 3). All the images available in the years 2017–2022, including the five S2 tiles covering Trentino-South Tyrol (T32TPR, T32TPS, T32TPT, T32TQS, and T32TQT) were processed. The S2 imagery was preprocessed to L2A by Sen2Cor [49], and specific revisions to the processing chain were implemented to adapt the retrieval to the high heterogeneity of the alpine environment. In particular:

• S2 L2 reflectance in the 10 m bands B3 and B4 and in the 20 m bands B5, B6, B7, B8a, B11, and B12 were used after resampling 20 m imagery to a 10 m spatial resolution to fully exploit the highest spatial resolution available from the MSI sensor and maintain (at the same time) the full spectral information required by the LAI biophysical processor to perform robust estimates. A preliminary comparison with ground measurements motivated this choice, as it showed that using only the 10 m bands, as in [50], performs worse over our study area;
• An additional cloud mask layer was generated by sen2cloudless [51] and was integrated, together with LAI biophysical processor quality flags, in the LAI maps masking process to minimize the occurrence of bad quality LAI values and their impact on the final drought index calculation.

3.1.2. Gaussian Process Regression (GPR) Models

GPR models were explored to interpolate the cloudy LAI retrievals at the parcel level (Section 3.1.3) and to enrich the S2 LAI time series with S1 SAR imagery (Section 3.3). The GPR models [52] allow for nonparametric regression and function approximation, also providing an estimation of the uncertainty in the prediction. In recent years, GPR models have often been used to estimate the biophysical parameters from optical and radar remote-sensing data [53,54,55]. Hereafter follows a short summary of how GPR models can be used for solving regression problems, while a complete mathematical description can be found in [52].

When given a training set, $D,$ of $N$ observations, where $D=\{\left({\mathit{x}}_{\mathit{i}}, y_i\right) |i=1, \dots ,N\}$, ${\mathit{x}}_{\mathit{i}}$ represents the input vectors, and $y_i$ represents the output scalars, the purpose is to estimate a function that is able to predict $y$ for the new inputs, ${\mathit{x}}_{*}$, for which there are no observations. In order to estimate this function, GPR uses a Bayesian approach. $f\left(\mathit{x}\right)$ is assumed to be a Gaussian-distributed random vector with a zero mean and covariance matrix $K\left(\mathit{x}, \mathit{x}\right)$, $f\left(\mathit{x}\right)=N\left(0, K\right)$. The elements of the covariance matrix are calculated by a kernel function [52], $k\left({\mathit{x}}_i,{\mathit{x}}_j\right)$, which describes the similarity between vectors ${\mathit{x}}_i$ and ${\mathit{x}}_j$. According to Bayes’ theorem, the posterior distribution of $f$ at the test point ${\mathit{x}}_{*}$, conditioned on training data and hyperparameters, can be expressed as

$$\begin{array}{c}p\left(y_{*}|{\mathit{x}}_{*}, D\right)=N(y_{*}|{\mu }_{GP}\left({\mathit{x}}_{*}\right), {\sigma }_{GP}^2\left({\mathit{x}}_{*}\right))\end{array}$$

(1)

with the predictive mean and predictive variance given by

$$\begin{array}{c}f\left({\mathit{x}}_{*}\right)={\mu }_{GP}\left({\mathit{x}}_{*}\right)={\mathit{k}}_{*}^T\left(K+{\sigma }_n^2 I_N\right) \mathit{y}\\ {\sigma }_f^2\left({\mathit{x}}_{*}\right)={\sigma }_{GP}^2\left({\mathit{x}}_{*}\right)=k\left({\mathit{x}}_{*}, {\mathit{x}}_{*}\right)+{\sigma }_n^2-{\mathit{k}}_{*}^T{\left(K+{\sigma }_n^2 I_N\right)}^{-1}{\mathit{k}}_{*}\end{array}$$

(2)

where ${\mathit{k}}_{*}$ = [$k\left({\mathit{x}}_{*}, {\mathit{x}}_1\right), \dots , k\left({\mathit{x}}_{*}, {\mathit{x}}_N\right)$]^T contains the similarities between the test point and the training points, $K$ is the $N\times N$ kernel covariance matrix containing the similarities between the training points, $\mathit{y}=\left[y_1,\dots , y_N\right]$^T, $k\left({\mathit{x}}_{*}, {\mathit{x}}_{*}\right)$ is a scalar with the self-similarity of ${\mathit{x}}_{*}$, ${\sigma }_n^2$ is a hyperparameter describing noise variance, and $I_N$ is the $N\times N$ identity matrix. The covariance function has an associated set of hyperparameters that are estimated during the model training phase by maximizing the marginal likelihood of the model.

3.1.3. Interpolation of the S2 LAI at Pixel and Parcel Level

The simplest gap-filling method that was tested is the linear interpolation of the LAI for each S2 MSI pixel. As an alternative to pixel-based gap-filling, two spatial-temporal interpolation techniques were assessed to exploit all the information available for each parcel while minimizing data storage:

• Parcel-based linear interpolation consists of extrapolating the mean values from the good-quality pixels within a given parcel followed by a linear interpolation between two subsequent dates. The interpolated LAI values can, thus, be easily computed from the parcel-specific linear functions at the desired temporal sampling interval. This approach is appealing due to its simplicity and fast interpretability. At the same time, it could be affected by noisy data or large temporal gaps between the available input LAI maps. The average LAI for a parcel p at the time t_i, when there is no observation available, is estimated as

  $$\begin{array}{c}{\mathrm{LAI}}_{p, t_i}={\mathrm{LAI}}_{p, t_1}+\frac{t_i-t_1}{t_2-t_1} \left({\mathrm{LAI}}_{p, t_2}-{\mathrm{LAI}}_{p, t_1}\right)\end{array}$$

  (3)

  where ${\mathrm{LAI}}_{p, t_1}$ and ${\mathrm{LAI}}_{p, t_2}$ are the average values of the LAI over the parcel at time t₁ and t₂, before and after t_i;
• Nonparametric, nonlinear interpolation is based on the use of GPR models to infer the most suitable continuous function from the available time series data. In this case, the output model is not constrained to a given set of points, as is the case of the linear method; thus, this possibly limits the effect of noisy data. According to this method, the mean and variance of the function estimating the LAI for a parcel at the time $t_{*}$, when there is no S2 imagery available, are estimated as

  $$\begin{array}{c}f\left(t_{*}\right)={\mathit{k}}_{t_{*}}^T{\left(K_t+{\sigma }_{nt}^2{I}_N\right)}^{-1}\mathit{y}\\ {\sigma }_f^2\left(t_{*}\right)=k\left(t_{*}, t_{*}\right)+{\sigma }_n^2-{\mathit{k}}_{t*}^T{\left(K_t+{\sigma }_n^2 I_N\right)}^{-1}{\mathit{k}}_{t*}\end{array}$$

  (4)

  where ${\mathit{k}}_{t*}={\left[k\left(t_{*}, t_1\right), \dots , k\left(t_{*}, t_N\right)\right]}^T$ is the array of the similarities of the test point $t_{*}$ with the training points, $K_t$ is the covariance matrix with the similarities between the training points, $\mathit{y}$ is the LAI time series derived from S2, and ${\sigma }_{nt}^2$ is the variance of the time series additive noise. The covariance matrix $K_t$ is calculated by the Matérn 3/2 kernel function [52,56]:

  $$\begin{array}{c}{K}_t={\sigma }_{st}^2\left(1+\frac{\sqrt{3} \left(\mathit{t}{J}_{1,N}-J_{N,1}{\mathit{t}}^T\right)}{{\sigma }_t}\right)\mathrm{e}\mathrm{x}\mathrm{p}\left(-\frac{\sqrt{3} \left(\mathit{t}{J}_{1,N}-J_{N,1}{\mathit{t}}^T\right)}{{\sigma }_t}\right)\end{array}$$

  (5)

  where ${\sigma }_t$ is the length scale along the time dimension, ${\sigma }_{st}^2$ is the time series signal variance, $\mathit{t}=\left[t_1,\dots , t_N\right]$^T is the vector of $N$ sample dates, and $J_{1,N}$ and $J_{N,1}$ are the $1\times N$ and $N\times 1$ unit matrices introduced for the dimensional equality of the expression.

3.1.4. Filling in the Missing LAI Values at the Beginning and End of the Growing Season

A further gap-filling technique was applied in all those instances showing LAI gaps at the head and tail of the time series, e.g., for those pixels or parcels where the time series showed missing values at the beginning and end of the growing season. Such situations prevented us from linearly gap-filling the values at the extremes of the growing season, as there were no LAI values available to be used for the interpolation. An extreme gradient boosting (XGB) regression model [57] was trained to predict any missing LAI value within the first 15 days and the last 15 days of the growing season. Elevation, daily mean temperature, month, and year were considered predictors. After performing Bayesian optimization hyperparameter fine-tuning [58], the LAI was estimated with an average RMSE of 0.33 [m² m⁻²], assessed by a five-fold cross-validation.

3.1.5. Estimation of the GPI and of Its Anomalies

Based on the daily LAI gap filled by pixel-based linear interpolation, the GPI was calculated as the growing season cumulate of the daily product between the LAI and a coefficient of water stress, $Cws$:

$$\begin{array}{c}{\mathrm{GPI}}_n={\displaystyle \sum _{i=SOS}^{i=EOS}}({\mathrm{LAI}}_{grassland i}\times Cw{s}_i) \end{array}$$

(6)

where $Cws$ is a meteorological water stress coefficient. The motivation behind including a meteorological water stress coefficient is derived from the authors of [31], who tested different formulations for a forage productivity index based on satellite data and demonstrated that the water stress coefficient significantly improved the performance of the index. Those results were based on the application of moderate-resolution satellite data over homogeneous grasslands in France; thus, their validity over our heterogeneous Alpine study area remains to be verified.

$Cws$ is defined as

$$\begin{array}{c}Cws=0.5+0.5 \mathrm{A}\mathrm{W}\end{array}$$

(7)

where $\mathrm{AW}$ is the available water, i.e., the ratio between the cumulated precipitation, P [mm day⁻¹], and the potential evapotranspiration, ${\mathrm{ET}}_0$ [mm day⁻¹], over a certain time interval. The daily $Cws$ was computed considering a running window of 30 days, over which the daily values of P and ${\mathrm{ET}}_0$ were cumulated. $\mathrm{AW}$ was set to 1 when P exceeded ${\mathrm{ET}}_0$ so that $Cws$ can only vary between 0.5 when the water shortage is maximum and 1 when no water stress conditions occur.

$Cws$ was computed on a 250 m resolution grid covering Trentino-South Tyrol from the year 2004 to the year 2021. ${\mathrm{ET}}_0$ was estimated by means of the Jensen-Haise equation [59], based on mean temperature T (°C) and solar radiation Rs (MJ m⁻²):

$$\begin{array}{c}{\mathrm{ET}}_0=\left(0.025 T+0.078\right) 0.408 Rs\end{array}$$

(8)

The daily time series of T and P were derived from the 250 m gridded dataset obtained by interpolating the quality-checked and homogenized database from the station observations of the regional meteorological networks [60].

Daily Rs grids at 250 m resolution were derived by applying a geostatistical downscaling to the daily Downward Surface Shortwave Flux (DSSF) product available on the LSA-SAF system (https://landsaf.ipma.pt (accessed on 10 November 2022)). In particular, the sharpening of daily DSSF was performed by Regression Kriging (RK) based on elevation, slope steepness and its orientation [61].

The start (SOS) and end (EOS) of the growing season were assumed to correspond to the 1 April and the 31 October, or to vary with elevation (

Table 2. Definition of the growing season length according to elevation. SOS and EOS are specified by day of year (DOY).

Elevation [m asl]	DOY of Start of the Growing Season (SOS)	DOY of End of the Growing Season (EOS)
<500	79	273
500–699	84	263
700–899	91	258
900–1099	100	253
1100–1299	105	248
1300–1500	121	242

). This corresponds to the specifications of the insurance system currently implemented in Trentino-South Tyrol and was determined by the agricultural consortia based on knowledge of the local agricultural practices. In this way, the insurance systems aim to account for the reduction in the growing season length occurring with increasing altitude [2].

Finally, the drought index was calculated as the anomaly of the GPI based on the preceding five years [62]:

$$\Delta {\mathrm{GPI}}_n=\frac{{\mathrm{GPI}}_n}{Olympic average\left({\mathrm{GPI}}_{n-1};\dots ;{ \mathrm{GPI}}_{n-5}\right)}$$

(9)

where the Olympic average is calculated by excluding the maximum and minimum values. Based on the indications of the agricultural consortia and on the conditions for the facilitation of the insurance policy [63], the final drought index that is used to estimate yield losses and insurance payments was averaged among the parcels with the same land use and belonging to the same farm and municipality. This reduces the risk of conflicts due to unequal payments within the same administrative unit. Furthermore, this spatial aggregation procedure excludes the GPI calculation polygons with low availability to high-quality S2 data, which could prevent a robust gap-filling.

The calculation of the GPI and ΔGPI was performed in Python 3 programming language [64].

3.2. Evaluation of the LAI and GPI as Proxies for Forage Yield

The measurement campaigns described in Section 2.2 were specifically designed to cover the spatial heterogeneity of the study area and the temporal variability of the LAI and yield and were used for the following analyses:

• In order to verify whether the LAI is a good proxy for grassland yield in our study area, we compared the ground measurements of the LAI and yield by regression analysis, also considering the effect of lodging;
• In order to evaluate the accuracy of the S2 LAI, we validated it against ground measurements of the LAI;
• We compared the performances of different interpolation methods against parcel-scale LAI measurements;
• We evaluated the GPI against yield measurements.

3.2.1. S2 LAI Validation

The S2 LAI dataset was compared to the ground observations of the LAI via the mean absolute error (MAE), mean bias (MB), root mean squared error (RMSE), and the coefficient of determination (R²), with the MAE, MB, and RMSE calculated as follows:

$$\begin{array}{c}\mathrm{M}\mathrm{A}\mathrm{E}=\frac{1}{N}{\displaystyle \sum _{i=1}^N}\left|Y_i-X_i\right| \end{array}$$

(10)

$$\begin{array}{c}\mathrm{M}\mathrm{B}=\frac{1}{N}{\displaystyle \sum _{\mathrm{i}=1}^N}{Y}_i-X_i \end{array}$$

(11)

$$\begin{array}{c}\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{\frac{1}{N}{\displaystyle \sum _{i=1}^N}{\left(Y_i-X_i\right)}^2} \end{array}$$

(12)

where X_i and Y_i are the observed and estimated LAI at the ith time step, and the estimated LAI consists of the original retrievals before the gap-filling.

A 5 m buffer was applied to select the pixels corresponding to each sampled plot to account for the absolute geolocation uncertainty of S2 and of the handheld global navigation satellite system (GNSS) used during the field campaigns. The standard deviation and weighted predicted mean for each plot were calculated based on the pixels intersecting the buffer.

3.2.2. Relationship between the Daily S2 LAI and GPI and the Physical Yield Measurements

The relationship between the remote sensing estimates of the LAI or GPI and the ground measurements of the forage yield was performed by ANOVA by using different subsets of data, depending on (i) the LAI interpolation method; (ii) the use of the LAI or GPI, being the daily value of the LAI corrected by means of Cws; (iii) the use of remote-sensing estimates obtained at the S2-pixel or parcel scale with the respective yield measurements (single measurements obtained within a single S2-pixel or their mean at the parcel level); (iv) the use of the exact yield sampling date or a conventional date of two days before as a reference date for extracting the LAI or GPI. This aims at preventing large discrepancies between the LAI/GPI and yield due to satellite overpass occurring on the mowing date, with the cut being performed after yield sampling and before the satellite overpass; (v) the inclusion or exclusion of the ground measurements performed just before or after the mowing dates. The analysis was performed for each possible combination of the levels of these factors. An adjusted R² was used to assess the fit of each model. The effect of these factors on R² was separately analyzed for each factor by means of linear mixed models, also accounting for the paired observations given by the combination of the levels of all other factors and differing only by the level of the factor in question, which were considered to be repeated measurements. The covariance structure was adjusted using the Akaike Information criterion as an indicator of model fit. The most suitable interpolation method (factor i) was identified first, and then all the remaining analysis (factors ii to v) was performed with the values obtained by this interpolation method. For all the formal analyses of this chapter, IBM^® SPSS^® Statistics version 27 was used. The respective scatterplots were produced with the package ggplot2 and the function lm of R, Version 4.3.1 [65].

3.3. Pilot Experiment of S2 LAI Time Series Enrichment by S1 SAR Data

SAR data are not affected by cloud coverage; therefore, they can be used to reduce the time gap between two subsequent LAI estimates. SAR imagery is influenced by many factors, including vegetation water content, undercover soil moisture, and terrain roughness, and can thus be difficult to interpret, especially over grasslands [36,66]. Consequently, the relationship between SAR data and the target biophysical parameter is strongly nonlinear and depends on external ancillary information. In order to overcome these issues, here, nonlinear machine learning regression was exploited. The rationale behind such techniques is the inference of a generic nonparametric and potentially nonlinear relationship between the input features (the SAR signal and possibly ancillary data) and the output target variable (LAI), exploiting a set of representative training examples. In particular, the use of the GPR technique was investigated because of its good generalization ability when the number of available training data is limited and the promising performance achieved in similar application contexts [54,66].

3.3.1. S1 Data Preprocessing

The S1 SAR imagery time series consists of 34 scenes acquired over the Muntatschinig test site (relative orbit 168) in the period April 2017–October 2017. This interval of time was chosen based on the availability of ready-to-use soil moisture data derived from S1, which are needed to perform the experiments described in Section 3.3.2. Only the descending pass direction was used because the ascending orbit was affected by layover effects over the study area and, thus, was not usable. The S1 images were downloaded from the Copernicus Open Access Hub as Level 1 Interferometric Wide Swath Ground Range Detected (GRD) High-Resolution mode, VV + VH dual polarization mode. Level 1 GRD consists of focused SAR data that have been detected, multi-looked, and projected to ground range using an Earth ellipsoid model. Pixel spacing is 10 × 10 m in High-Resolution mode. An additional preprocessing of S1 VH and VV bands was performed by the SNAP S1 toolbox [48], consisting of spatial speckle filtering with a 7 × 7-pixel window Lee-Sigma spatial filter, bad acquisition pixel masking by means of extremely low and high backscattering value exclusion according to a statistic performed over the test area, and co-registering with the S2 LAI time series. Finally, for each S1 acquisition, two indices that are representative of the sensitivity of SAR backscattering to vegetation development were generated, including the VH/VV simple ratio and the dual-pol radar vegetation index (RVI, [67]), calculated as

$$\mathrm{RVI}=\frac{4\mathrm{VH}}{\mathrm{VH}+\mathrm{VV}}$$

(13)

3.3.2. S1 and S2 Data Fusion by GPR

A pilot assessment of GPR to fuse the S1 and S2 time series data was performed to enrich the LAI time series retrieved from the S2 optical data with the S1 SAR data. This test was performed for the growing season of the year 2017 over a subset of the study area, including 118 parcels (Figure S1,

Table 3. Characteristics of the parcels used to train the GPR model to estimate the LAI from S1, aggregated by land use.

Land Use	Number of Parcels	Minimum Parcel Area [m2]	Maximum Parcel Area [m2]
Wooded summer pasture, tare 20%	21	64	154,924
Wooded summer pasture, tare 50%	12	397	42,743
Summer pasture	4	4602	11,346
Pasture	5	958	16,428
Wooded pasture, tare 20%	8	507	2131
Permanent meadow	64	28	72,467
Meadow, special area	4	981	4844

). The temporal extent is based on the availability of ready-to-use input features, while the spatial extent includes samples of all land uses where the proposed drought index in Trentino-South Tyrol is potentially applicable.

The target variable was the LAI derived from S2. For the input features, different combinations of the following variables were considered, which were normalized by scaling between 0 and 1:

• VV and VH polarization bands from the S1 dual polarization StrimMap product;
• Simple ratio VH/VV and RVI, which have been proven to enhance the sensitivity of SAR data to vegetation characteristics [67,68,69];
• Soil moisture content maps (SM), produced by Eurac Research from S1 imagery [70], with the aim of providing information on the soil water content status to the regression algorithm to help disentangle the intrinsic ambiguity of SAR data;
• Day of year (DOY).

The selection of the input features was performed based on a five-fold cross-validation, testing all the possible combinations of the input features across the 118 parcels.

The training of the GPR model, consisting of the optimization of the hyperparameters through the maximization of the marginal likelihood, was performed at different scales: (a) land use-specific and (b) parcel-specific. The trained models were used to estimate the LAI and its uncertainty from the selected input features at the test point ${\mathit{x}}_{*}\in {ℝ}^D$, where $D$ is the number of input features. For this application, the squared exponential covariance function was used [52]:

$$k\left({\mathit{x}}_i,{\mathit{x}}_j\right)={\sigma }_f^2\exp \left[-\frac{1}{2}\frac{{\left({\mathit{x}}_i-{\mathit{x}}_j\right)}^T\left({\mathit{x}}_i-{\mathit{x}}_j\right)}{{\mathsf{\sigma }}_l^2}\right]$$

(14)

where ${\sigma }_f^2$ is the signal variance, and ${\sigma }_l$ is the scale length.

After feature selection, two different experiments were performed, i.e., “spatial sampling” to verify the ability of the GPR model to predict missing pixel values due to cloudiness and “temporal sampling” to verify the ability of the GPR model to predict pixel values on missing acquisition dates. In the first case, 30%, 50%, or 70% of the randomly selected pixels within each S1 acquisition date were excluded from the training set and were used as the test set. In the “temporal sampling” experiment, 3, 6, or 9 acquisition dates within the entire growing season were excluded from the training set and were used as the test set. Both experiments were performed on individual parcels and on clusters of parcels with the same land cover type. To avoid biased results, random sampling was performed five times, and then the GPR model performances were averaged.

4. Results

4.1. Comparison between the Ground Measurements of the LAI and Yield

As the first part of the GPI assessment, we verified the relationship between the LAI, which was the variable selected for the calculation of the index, and yield, which was the target variable, for the year 2021, when the ground measurements of both variables were collected (Table 1). Based on all the available observations, an R² = 0.66 between the ground measurements of the LAI and yield was found. The LAI tends to increase faster than the yield for low yield values, with an overestimation of lower-yield values (Figure 4), and most of the strongly underestimated yield values are observations affected by the medium or heavy lodging of the vegetation at the time of the measurement (Figure 4a). Indeed, dropping these observations (40 cases, 20 of which were affected by heavy lodging) from the investigated dataset resulted in an improvement of R² to 0.71, with a slight reduction in the slope, and its standard error from 0.64 ± 0.020 to 0.59 ± 0.017 (Figure 4b).

4.2. Validation of the LAI Derived from S2

As the second part of the GPI assessment, we verified the performance of the LAI estimated from satellite imagery against the ground measurements of the LAI performed in the year 2021 over four sites and in the years 2017–2021 over one site (Table 1). The S2 LAI showed an R² between 0.48 and 0.88 and an RMSE between 0.71 and 1.2 [m² m⁻²] with respect to the ground measurements of the LAI for all the monitored meadows (

Table 4. Metrics of the validation of the LAI estimated from S2, with ground measurements of the LAI. n is the number of pairs used for the comparison.

Parcel	Years	MAE [m2 m−2]	RMSE [m2 m−2]	MB [m2 m−2]	R2	n
F1	2021	0.58	0.71	−0.39	0.88	21
F2	2021	0.69	0.87	−0.69	0.85	8
L1	2021	0.98	1.10	0.18	0.83	11
L2	2021	0.58	0.81	−0.11	0.83	11
R1	2021	0.82	1.01	0.08	0.85	25
R2	2021	0.96	1.20	0.81	0.74	19
R3	2021	0.72	0.98	0.20	0.48	25
R4	2021	0.86	0.95	0.67	0.83	20
V1	2017–2021	0.92	1.10	0.30	0.69	60
P2	2017–2021	0.59	0.67	0.53	0.12	89
V1, P2 (Figure 6)	2017–2021	0.72	0.87	0.44	0.80	149
All (Figure 5)	2021	0.78	0.98	−0.19	0.78	163
All	2017–2021	0.80	0.92	−0.31	0.81	289

). Only at the pasture site, P2, were the performances worse (Table 4), with an R² = 0.12. A tendency to overestimate low LAI values is observable (Figure 5 and Figure 6). In the year 2021 (Figure 5), R² = 0.78 and RMSE = 0.98 [m² m⁻²], based on the measurements from all the sites. In the period 2017–2021 (Figure 6), R² = 0.80 and RMSE = 0.87 [m² m⁻²], based on the measurements from the two Muntatschinig parcels.

4.3. Assessment of Three LAI Interpolation Methods

The performance of the different interpolation approaches was assessed via a comparison of the field measurements of the LAI aggregated at the parcel level (Figure 7). The comparison was performed when and where the ground data were available, i.e., in the year 2021 over four sites and in the years 2017–2021 over one site (Table 1). On average, when considering all the stations and all the available data, for the parcel-based linear interpolation, RMSE = 1.2 [m² m⁻²] and R² = 0.72; for the pixel-based interpolation, RMSE = 1.35 [m² m⁻²] and R² = 0.67, and for GPR interpolation, RMSE = 1.28 [m² m⁻²] and R² = 0.61. When considering the higher R², only the first two methods were applied for a comparison between the LAI/GPI and yield (Section 4.4).

4.4. Comparison between the LAI/GPI from the Remote Sensing and Ground Yield Measurements

As the third and last part of the GPI assessment, we compared the LAI and GPI estimated from satellite imagery against the ground measurements of yield performed in the year 2021 over four sites, including eight parcels (Table 1).

The use of the LAI gap-filled by pixel-based linear interpolation (method 1) resulted in higher mean R² values when compared to parcel-based linear interpolation (method 2), with a mean difference of about 0.1; this was apparently caused by the notably higher R² values obtained by this method at the sites F1, R1, and R4, while similar results were observed using the two methods at the remaining sites (Figure 8). For this reason, all further analyses were performed using method 1.

The analysis of the relationship between the remote sensing-based LAI or GPI and yield showed that model fit (assessed by the R² of the linear regression) greatly varied depending on the combination of factor levels used in the models, and it ranged from a minimum of 0.45 to a maximum of 0.75 (Figure 9, Figures S5–S7).

When averaged over the levels of all other factors and considering the effect of the paired observations given by the combination of the levels of all other factors (differing only by the level of the factor in question), systematic patterns regarding model fit were detected (Figure 10). The use of data averaged at the parcel level instead of at the pixel level resulted in the most relevant improvement in model fit, with a mean difference in R² of 0.17 (Figure 10b). The inclusion of the measurements taken on dates close in time to the mowing event (the last one preceding it and the first one following it) was proven to negatively affect model fit. Indeed, the removal of these measurements increased R² from 0.57 to 0.66 (Figure 10d). A significant but small improvement in R² from 0.60 to 0.62 resulted from the use of satellite imagery from two days prior to the sampling event when compared to the use of that of the real sampling events (Figure 10c). In contrast, it was observed that the use of the GPI instead of the LAI did not result in a significant improvement in model fit (Figure 10a).

The impact of operating at the parcel vs. pixel level and of excluding observations close to the mowing events is exemplified in Figure 9 for the relationship between the remote sensing-estimated LAI and the respective ground measurements of yield by using the real mowing dates as a reference. Operating at the plot level improved R² from 0.45 to 0.64 when using all data and from 0.55 to 0.74 when excluding the data close to the mowing date. At both the pixel and parcel levels, the exclusion of the observations close to the mowing events led to an improvement in R² of 0.10.

The same pattern was detected for all other analyses when using a date that occurred two days before the mowing event for LAI extraction and all for other analyses addressing the GPI as a dependent variable (Figures S5–S7).

Only eight yearly yield values at the parcel scale were available from the test sites. The analysis of their correlation with yearly cumulated LAI and GPI showed a strong linear relationship between the LAI and yield (Figure 11a) and a moderate correlation between the GPI and yield (Figure 11b) when a fixed length of the growing season was assumed. When the growing season length was assumed to vary with elevation, the R² between the LAI and yield decreased from 0.56 to 0.43 (Figure 12a), and that of GPI decreased from 0.42 to 0.31 (Figure 12b). Moreover, the p-value of the relationship between the GPI and yield increased from 0.049 to 0.087.

4.5. Estimated Yield Losses for the Year 2022

In the year 2022, yield losses were estimated using ΔGPI for eight farms. ΔGPI was calculated for all the insured parcels of each farm and was averaged for all the parcels of a farm located in the same municipality and with the same land use. The damage was calculated as (ΔGPI-1) × 100, with negative values representing a reduction in production with respect to the average of the preceding years.

For several parcels, the estimated damage was high (Figure 13); however, when averaged by farm, municipality, and land use, only in one case (Figure 13l, Farm 8, located in Brentonico, used as a permanent meadow) did the average damage exceed the threshold of 30%, which is set by the Italian agricultural risk management plan [71] as the minimum value of damage that can be refunded for index-based insurance policies.

4.6. Evaluation of the GPR Model’s Potential to Enrich the S2 LAI Time Series by Using S1 SAR Data

S1–S2 data fusion via the GPR model was tested over a subset of the study area for the year 2017, as described in Section 3.3. Here, the results of the feature selection and of the model validation are summarized.

4.6.1. Feature Selection

The five-fold cross-validation, encompassing all the possible combinations of the input features, showed, on average, that the best predictors are VH, SM, and DOY. From this combination, the LAI was estimated from S1, with an R² = 0.83 and an RMSE = 0.25 [m² m⁻²], with respect to the LAI estimated from S2. The performance depends on land use (Figure S2, which includes only the best combinations of the features). Training using the individual parcels performed better than the training using clusters of parcels for the same land use, especially for “permanent meadow” (AP2).

The combination of VH, SM, and DOY provided (globally) the best performance and was used for the analyses presented hereafter.

4.6.2. Temporal and Spatial Gap-Filling Experiments

For all the land-use types, when the dimension of the temporal gaps in the LAI time series increased, the RMSE increased (Figure S3). R² decreased with increasing numbers of missing acquisition dates for the parcel-based training, while the experiment with three missing dates provided the worst results for the training based on land-use classes (Figure S3d).

For the parcel-based models, on average, when considering all the parcels, R² = 0.44, 0.40, and 0.28, and RMSE = 0.55, 0.65, and 0.91 [m² m⁻²] for 3, 6, and 9 missing acquisition dates, respectively. For the land-use based models, on average, based on all the land use, R² = 0.20, 0.26, and 0.19, and RMSE = 0.69, 0.87, and 0.98 [m² m⁻²] for 3, 6, and 9 missing acquisition dates, respectively.

The dimension of the spatial gaps in the LAI images derived from S2 did not influence the average results (Figure S4). Specifically, when artificial gaps of 30%, 50%, and 70% of the pixels were introduced, the comparison between the S1 and S2 LAIs showed, respectively, R² = 0.53, 0.53, and 0.52, and RMSE = 0.66, 0.67, and 0.68 [m² m⁻²].

5. Discussion

This work derives a forage production index for mountain grasslands, known as the GPI, based on the S2 LAI. The choice of the LAI as a biophysical indicator of grassland productivity is motivated by its strong correlation with the measured yield over our study area. In addition, the calculation of the LAI is relatively simple, thus providing the GPI immediately after the end of the growing season, which is a requirement for the timely calculation of insurance payments.

5.1. Design of the GPI

For the definition and development of the GPI, we took the appropriate measures to minimize basis risk, which is usually associated with index insurance systems, i.e., the difference between the index value and the derived payoffs and the actual yield variations [72]. Specifically, we considered the three main components of basis risk identified by [73]:

• Design risk, as derived from the possibility that the index does not correctly describe the variable of interest, i.e., yield, was minimized by studying the correlation between the LAI and yield, verifying the accuracy of the LAI derived from satellite data, and assessing the ability of the GPI to explain yield variability;
• Temporal basis risk, emerging from the mismatch between the temporal aggregation of the index and the temporal scale of interest for grassland growth, was reduced by estimating the GPI as the growing season cumulate of the daily values of the LAI;
• Spatial risk, as the distance between the insured farm and the point where the index is estimated, is partially solved thanks to the high spatial resolution of S2 data, which allows for an estimation of the GPI by aggregating the pixels within the parcels belonging to the insured farms. However, when the parcels are aggregated by municipality, as foreseen by the Italian regulations [71], the variability from one parcel to another is lost, and the farmers are not rewarded for the parcels where high losses are estimated. Yet, through aggregation, it is possible to exclude (from the calculation of the damage) the parcels that lack the requirements for the calculation of ΔGPI, i.e., dense tree cover, recent changes in land use, and extremely small parcel size.

In this work, we accumulated the LAI values over the entire growing season to obtain an estimate of the yield. One alternative method that might be tested is the accumulation of LAI estimations immediately before mowing events. In the near future, an approach to identifying mowing events from S1 and S2 satellite data [74,75] can be incorporated into our algorithm for the calculation of the GPI. In particular, there is one method under development focusing on the Alpine region, which is where our study area is located [76]. In the use of these approaches, on the one hand, it is still challenging to obtain the exact date of the mowing over heterogeneous agricultural areas from earth observation sensors, and on the other hand, it is difficult to obtain good LAI retrievals from S2 in correspondence to mowing due to cloudiness. So far, the yearly cumulation is, in our study setup, the best solution to get a reasonable quantitative assessment of production over the growing season. Nevertheless, the correct identification of the mowing dates might be useful for refining and constraining S2 LAI gap-filling approaches. For example, it would avoid the use of images acquired before and after a mowing event to estimate a missing LAI value in between.

5.2. Ground Measurements of the LAI and Yield

In the analysis of the relationship between the in situ LAI and yield, the fit improvement obtained by disregarding the yield observations affected by lodging suggest that lodging can be regarded as a source of inaccuracy and a possible critical point if LAI retrievals from remote sensing are used in insurance systems to assess the yield in the field. The impact of lodging on grassland reflectance is rarely mentioned in the literature, but a few studies suggest that lodging can alter the leaf area of sunlight interception, which reduces the LAI [77,78]. As lodging is favored by the occurrence of large amounts of yield, this noise is mainly expected to occur in situations of high forage availability, thus leading to an underestimation of very high yield values.

5.3. S2 LAI Validation and Its Relationship with Yield

The validation of the S2 LAI showed that the S2 biophysical processor could estimate the LAI at a satisfactory level of accuracy for the grasslands of Trentino/South Tyrol. Despite the fact that our study area is rather challenging due to land cover heterogeneity and topographic complexity, the results are in line with other large-scale validation studies performed over homogeneous sites [79,80]. We observed a tendency to overestimate low yield values when using the LAI. Indeed, at the pasture site, characterized by very low productivity and LAI values, the worst performances were observed, suggesting that the index we developed is appropriate for estimating yield losses in meadows, but further improvements are needed for pastures or, in general, for extensively managed grassland with low yield potential.

In the comparison between the daily S2 LAI/GPI and the in situ yield, the improvement in model fit observed when moving from the pixel to the parcel level suggests that averaging data at the parcel level potentially reduces the effect of single outliers. This is a positive outcome because the insurance payments are estimated based on the GPI at the parcel scale.

In contrast to what was observed in [31], in which moderate resolution remote sensing data were used, no improvement in model fit could be achieved by taking into account the meteorological component (i.e., the water stress coefficient, Cws).

In the comparison between the S2 LAI/GPI cumulated during the growing season and the yearly in situ yield, the correlation was slightly reduced when we assumed the length of the growing season, which varies with elevation (Figure 11 and Figure 12). Although an ultimate evaluation of this approach is very difficult based on only eight observations, some considerations are possible. On the one hand, the discrepancies might be attributable to the fact that there can be yield accumulation beyond the empirically determined growing season start and end. On the other hand, because the LAI was found to overestimate low yield values, this approach could prevent the incorrect accumulation of yield in times when plant growth had not really started. Moreover, it is a common practice to let young cattle return to the home farms from the summer pastures to graze the last regrowth of the meadows from late Summer to Autumn. This prevents a reliable estimation of productivity during the last part of the growing season based on remote sensing observations. Considering a shorter length for the growing season can help to avoid these cases and to include only the period in which grassland is mown and there is no grazing.

5.4. Enrichment of the S2 LAI with SAR Data

We preliminarily assessed the potential of the fusion of optical and radar data for estimating grassland LAI. This exploratory analysis showed promising results, especially for spatial gap-filling, suggesting that the fusion with S1 could help to reduce gaps due to cloudiness over the LAI maps obtained from S2. In particular, VH polarization (in combination with ancillary data) showed better predictive capacity than the VV or SAR vegetation indices in our specific test case. This could be ascribed to the increased sensitivity of VH polarization to grassland structures and to less pronounced saturation tendencies with increasing grassland height (and, thus, LAI) [81]. The GPR model training performed by aggregating all the parcels with the same land use gave worse results for the permanent meadows. This class includes parcels with very different sizes and agricultural management. A different way of aggregating the parcels for model training should be tested for this land use class, considering the heterogeneity of agricultural practices. In the future, we plan to perform GPR model training based on a more extended dataset. In addition, a more thorough exploration, with a comparison of different machine learning models, is required before including this technique in the production of the GPI.

6. Conclusions and Outlook

This study supports the development of an insurance instrument to mitigate the consequences of drought for grassland farming in the Alps. An innovative index-based insurance system was developed in co-operation with local agricultural stakeholders, accounting for the characteristics of the region in terms of meteorology, topography, and land use and management. A thorough comparison of the satellite-based LAI and GPI with extensive ground measurements of the LAI and yield proved the robustness of the drought index.

For the growing season of the year 2022, the agricultural consortia of Trentino and South Tyrol saw estimated yield losses due to droughts at eight farms, including around 700 agricultural parcels, according to the anomalies of the GPI developed in this paper. The first version of the GPI included the pixel-based linear interpolation of the LAI, the Cws meteorological coefficient, and an adjustment to the length of the growing season depending on elevation.

The accuracy of the GPI can be improved by enriching the LAI time series based on optical S2 data with radar S1 data because the gaps in the S2 data due to cloudiness can last several days. The synergistic exploitation of S1 and S2 was tested in this paper and will be developed as the next step.

The agricultural consortia of Trentino and South Tyrol, which are cutting-edge in terms of offering innovative solutions for the management of agricultural risk to their associate farmers and to their partner insurance companies, have greatly appreciated the developed GPI index and are interested in proposing GPI-based insurance to several grassland farmers starting from the year 2024, possibly based on an enhanced GPI using optical and SAR data fusion.