Two Datasets over South Tyrol and Tyrol Areas to Understand and Characterize Water Resource Dynamics in Mountain Regions

Abstract

In this work, we present two datasets for specific areas located on the Alpine arc that can be exploited to monitor and understand water resource dynamics in mountain regions. The idea is to provide the reader with information about the different sources of water supply over five defined test areas over the South Tyrol (Italy) and Tyrol (Austria) areas in alpine environments. The snow cover fraction (SCF) and Soil Moisture Content (SMC) datasets are derived from machine learning algorithms based on remote sensing data. Both SCF and SMC products are characterized by a spatial resolution of 20 m and are provided for the period from October 2020 to May 2023 (SCF) and from October 2019 to September 2022 (SMC), respectively, covering winter seasons for SCF and spring–summer seasons for SMC. For SCF maps, the validation with very high-resolution images shows high correlation coefficients of around 0.9. The SMC products were originally produced with an algorithm validated at a global scale, but here, to obtain more insights into the specific alpine mountain environment, the values estimated from the maps are compared with ground measurements of automatic stations located at different altitudes and characterized by different aspects in the Val Mazia catchment in South Tyrol (Italy). In this case, an MAE between 0.05 and 0.08 and an unbiased RMSE between 0.05 and 0.09 m³·m⁻³ were achieved. The datasets presented can be used as input for hydrological models and to hydrologically characterize the study alpine area starting from different sources of information.

Dataset: The dataset has been submitted and will be published as a supplement to this paper in the journal Data.

Dataset License: CC-BY-4.0

1. Summary

Understanding the condition of individual components within the hydrological cycle is crucial for effective water resource management and acts as a key tool for adapting to climate change. For example, snow plays a key role in the hydrological cycle and is an important indicator of climate change [1,2]. The large amount of water stored in the snowpack during the winter season is crucial for spring runoff, impacting the downstream agricultural production, which relies on irrigation [3]. Acquiring information on the snowpack, which in turn gives information on the amount of water stored, is therefore crucial to improving knowledge of water availability in mountain basins. In the same way, the soil moisture content (SMC) represents a crucial state variable in the global cycles of water, energy, and carbon, and it therefore plays a key role in the study of Earth’s climate and weather. However, direct measurement of these geophysical parameters by means of field measurements is not always possible, especially in inaccessible areas with complex topography, which is typical of mountain catchments. Moreover, although automatic stations (e.g., meteorological stations) can provide accurate observation data for a long time series [4,5], the number of stations in mountainous regions remains low [6], impeding a spatially distributed monitoring of these variables.

The scarcity of ground measurements also affects the accuracy of model-based estimation of geophysical parameters, as these measurements represent the input variables for the models. Especially in remote areas, where the availability of observed meteorological data on the ground is very limited, potential errors can be introduced on a regional scale [7]. Thus, collecting snow and soil moisture observations and modeling the snowpack and the soil moisture content can be extremely challenging due to its high spatial and temporal variability in mountainous areas.

Beyond the more traditional modeling approaches and in situ measurements, the development of techniques to detect geophysical parameters using remote sensing images has significantly contributed to improving snow and soil moisture mapping and providing spatially continuous measurement over large and remote areas [8,9]. Traditionally, snow cover mapping techniques identify pixels as binary snow cover, i.e., either snow-covered or snow-free. However, the binary snow cover area (hereafter referred to as SCA) classification does not efficiently capture subpixel characteristics. The variability of snow at the sub-grid level requires, instead, the knowledge of the fractional snow cover and its distribution as accurately as possible to integrate this information in numerical models simulating the hydrological or atmospheric surface energy exchange processes [10].

The Fractional Snow Cover (FSC), that is, the snow-covered fraction of the pixel area, naturally provides finer information than the binary SCA, which may be insufficient to characterize the snow distribution in areas where partially snow-covered (mixed) pixels are prevalent [11], especially during the melting season.

Regarding the SMC, the currently available operational products rely on data with a coarse to medium spatial resolution by means of passive or active microwave sensors such as SMOS (passive) [12], SMAP (passive) [13], or ASCAT (active). In the case of radiometers, the typical spatial resolution is about 40 km, while in the case of the ASCAT scatterometer, algorithms have been developed that can obtain the soil water index at 12.5 km [14].

In this context, this study aims to provide the reader with two datasets representing two time series of SCF and SMC over alpine regions located in South Tyrol (Italy) and Tyrol (Austria). The presented datasets were developed in the framework of the project ACR_Water (Assimilating Cosmic-Ray Neutron and Remote Sensing Data for Improved Water Resource Management) and are based on remote sensing data and machine learning techniques. They are part of an approach based on complementary data sources aiming to develop a strength-driven approach to combine remote sensing data with a Cosmic-Ray Neutron Sensing (CRNS) observation network. The provided datasets can be used, for example, in the cryosphere sciences (regarding the SCF data) and for hydrological or climatological purposes (e.g., hydrological models).

2. Study Area Anddata Description

2.1. Study Area

The study area covers the surroundings of the five CRNS used in the project, which are located in South Tyrol (Italy) and Tyrol (Austria): Corvara, Weisssee, Leutasch, Dresdner Hütte, and Obergurgl. The CRNS probes, which are geographically distributed around the study area, are located in regions with different topographical, geological, and meteorological characteristics. The probes are situated in the Eastern Alps within the catchments of the Inn, Adige, and Leutascher Ache Rivers. The CRNS locations are shown in Figure 1, where the red lines delimit the study area and separate the North (and East) Tyrol (Austria) from the Southern part (Italy).

2.2. Data Description

The SCF and SMC maps are provided as a percentage (representing the percentage of snow and the soil moisture content per pixel, respectively) in NetCDF format. The time series covers the period from October 2019 to September 2022 for the SMC product and from October 2020 to May 2023 for the SCF product. All the maps have a spatial resolution of 20 m.

The characteristics of the datasets are reported in

Table 1. Characteristics of the datasets.

Characteristic	Snow Cover Fraction Dataset	Soil Moisture Content Dataset
Name	Snow Cover Fraction	Soil Moisture Content
Data type	Integer (0–100 = % pixels snow coverage; 255 = No data)	Integer (1–100 = % soil moisture content; 0 = No data)
Data format	NetCDF	NetCDF
Projection	WGS84/UTM zone 32N	WGS84/UTM zone 32N
Spatial coverage	10 × 10 km surrounding the CRNSs	10 × 10 km surrounding the CRNSs
Spatial resolution	20 m	20 m
Temporal coverage	2020–2023	2019–2022
Temporal resolution	Approximately every 5 days (depending on the cloud coverage)	On average approx. a map every 2–3 days. After 23rd of December 2021, only half of the maps are available due to the failure of S1–B satellite. Each satellite pass only partially covers the study area.

. The maps were generated for the areas surrounding the CRNSs, whose characteristics are listed in

Table 2. Test site coordinates and altitude.

Test Site	Latitude	Longitude	Altitude [m]
Corvara	46.544	11.9	1924
Leutasch	47.376	11.162	1111
Weisssee	46.873	10.714	2464
Dresdner Hütte	46.997	11.14	2293
Obergurgl	46.849	11.031	2644

. From now on, we will refer to these areas as “test sites”.

3. Methods

3.1. Snow Cover Fraction (SCF) Algorithm

The input data used for snow cover retrieval consist of Sentinel-2 (S2) images. The constellation is composed of two twin satellites, namely, S2-A and S2-B, which together provide a revisit frequency of 5 days in the area of interest. The study area is covered by five tiles: 32TNT, 32TPT, 32TPS, 32TQT, and 32TQS. The S2 Level 1C spectral bands are downloaded from CREODIAS (https://creodias.eu/ (accessed on 2 April 2024)) and appropriately preprocessed in preparation for the core snow cover fraction (SCF) algorithm. Three main steps are applied: (i) Conversion from digital number (DN) to Top of the Atmosphere (ToA) reflectance values using the quantification value provided in the Sentinel-2 metadata. (ii) Resampling of all bands to a spatial resolution of 20 m using bilinear interpolation and cropping to the area of interest. Here, the raw S2 grid and coordinate reference systems are preserved to minimize changes as much as possible. (iii) Cloud masking using the S2 cloudless algorithm, which is available at https://github.com/sentinel-hub/sentinel2-cloud-detector (accessed on 2 April 2024) [15].

The SCF algorithm retrieval is based on a Support Vector Machine (SVM) classification that is manually trained using an Active Learning (AL) procedure [16].

All spectral bands are utilized as features, along with informative indices such as the Normalized Difference Vegetation Index (NDVI), which is calculated as the normalized difference between the near-infrared (NIR) and red bands, as well as the Normalized Difference Snow Index (NDSI), which is calculated as difference between the green and the shortwave infrared (SWIR). The data are normalized with a standard scaler. The SVM model utilizes a radial basis function kernel. To determine the model parameters, we employed a grid search strategy to identify the regularization parameter C and the kernel coefficient gamma. The grid is initialized with a user-defined range. The model selection process begins with a coarse grid and, based on the obtained results, is refined around the values of C and gamma that performed the best. The best values are selected by evaluating the mean and standard deviation of the accuracy calculated in a cross-validation strategy with k-fold validation (k = 5).

The SVM model is trained exclusively with pure class pixels, i.e., 100% snow-covered pixels (class “snow”) and 0% snow-covered pixels (class “snow-free”). These pixels are collected under different illumination conditions, i.e., diffuse light, direct light, and shadow. This is accomplished through visual inspection of the spectral signatures of the collected training samples, as well as considering the characteristics of adjacent pixels, such as vegetation in the surrounding area. For example, these pixels are collected distant from vegetated areas because of the difficult interpretation of those areas. In this phase, it is helpful to visualize an RGB false-color composition together with representative spectral indices such as the NDSI and NDVI. Furthermore, it is also useful to include DEM-derived features for excluding steep areas in training selection and ensure that training samples are equally spread over different elevations and aspects.

The selected training data should represent distinct classes of “snow” and “snow-free” accurately, as they play a crucial role in defining the hyperplane, which is the decision function that separates the two classes. In detail, the training aspects that define the hyperplane are the so-called support vectors. Once the hyperplane is defined, we can establish a correlation between the SCF and the distance to the hyperplane.

The AL procedure helps accelerate the learning process of the classification by involving the user in collecting training samples iteratively. The training selection was performed ad hoc for each scene (image) and iteratively by the user, visually assessing the results for each scene. Consequently, different trainings might be selected for different scenes, as well as different model parameters. The model parameters are the Lagrangian multipliers of the optimization function and gamma and C [17,18]. This approach ensures a scene-specific classification tailored to the final purpose of achieving the highest possible accuracy for the intended product assimilation.

In previous studies [19,20], a similar approach was applied. However, they were focused on simple binary classification, while here, we estimated SCF (snow cover fraction). Additionally, an improved version of this algorithm is presented by [21], which is based on the automatic selection of training data. This represents a big advantage when considering large areas and long time series of data, while it is preferable to collect ad hoc training to achieve better performance.

3.2. Soil Moisture Content (SMC) Algorithm

The algorithm used to estimate the SMC was developed by Greifeneder et al., and it is based on the Machine Learning (ML) approach described in [22]; the source code is available online at https://zenodo.org/records/4552813 (accessed on 23 March 2024), and the related documentation is available at https://pysmm.readthedocs.io/en/latest/ (accessed on 23 March 2024). The algorithm uses Gradient Boosted Regression Trees (GBRTs), which is a family of tree-based methods. This characteristic entails that it is compatible with different data and scales. Moreover, it has a relatively low computational cost associated with algorithm training and target prediction. Finally, the algorithm exploits the server-side processing capabilities of Google Earth Engine (GEE), eliminating the requirements to download or preprocess the input datasets.

The primary input data for the Soil Moisture Content Estimation algorithm consist of Sentinel-1 (S1) images. This means that soil moisture maps were generated at each Sentinel-1 satellite pass over the study area. For each satellite pass, the dual-polarization (VV + VH) Ground Range Detected products were used as SAR input features for soil moisture estimation. Sentinel-1 satellites carry a synthetic aperture radar (SAR) sensor onboard that can provide images of the Earth’s surface with better resolution than other microwave sensors (radiometers or scatterometers) independently of weather conditions and sunlight, unlike optical sensors. The joint use of the two satellites S1-A and S1-B made it possible to obtain a repeat cycle of 6 days over the area of interest with SAR Interferometric Wide (IW) swath mode. However, Copernicus Sentinel-1B encountered an anomaly related to the power supply of the instrument’s electronics on 23 December 2021. Since then, the satellite has no longer been able to provide radar data. As a result of this failure, the only available satellite is S1-A, and consequently, the repeat cycle has increased to 12 days.

Despite that repeat cycle, many orbits partially cover the area of interest of Tyrol and South Tyrol in ascending or descending mode, increasing the temporal resolution of the SMC time series exploiting the six relative orbits: 66, 168, 95, 15, 117, and 44 (Figure 2). The coverage for each test site is shown in

Table 3. In the second and third columns, the relative orbits of Sentinel-1 in ascending and descending mode, respectively, are listed, covering the test sites.

Test Site	Ascending Relative Orbit	Descending Relative Orbit
Leutasch	117	168, 095
Dresdner Hütte	117	168, 095
Weisssee	015, 117	168
Obergurgl	117	168, 095
Corvara	044, 117	168, 095

.

Sentinel-1 data and all the other input datasets are available on GEE. They are used as input features and/or for masking pixels where the algorithm cannot estimate the SMC (i.e., pixels with snow cover presence, with high vegetation coverage, or in layover shadow).

The other input datasets collected on GEE include Landsat-8 (shortwave reflectance and thermal radiance) [23], MODIS MOD13Q1 Enhanced Vegetation Index (EVI) [24], soil temperature and snow-water-equivalent from Global Land Data Assimilation System (GLDAS) [25], Copernicus Global Land Cover Layer [26], and soil information from OpenLandMap (OLM) [27] and SRTM Aster DEM [28].

The model applied was trained and extensively validated on a global scale in [22]. For this reason, 461 automatic stations of the International Soil Moisture Network (https://ismn.geo.tuwien.ac.at (accessed on 30 October 2024)) were exploited by selecting those where soil moisture content was measured at a depth of no more than 5 cm, and input features of the algorithm were available on the same date. All data collected from ISMN were used to create the training and test datasets through a randomly partitioning procedure in a proportion of 80% and 20%, respectively. Using an iterative approach, the training dataset was divided into N groups: the first, N-1, was validated on the last. Finally, the test score was evaluated on the independent test dataset.

The original algorithm was implemented to achieve a spatial resolution of 50 m. For this purpose, bilinear interpolation was applied to all input features before applying the GBRT model. In this work, the same procedure was used to obtain a spatial resolution of 20 m (Figure 3). In this way, the spatial resolution is consistent with that of the SCF product.

The landcover classes included in the estimation from the Copernicus Global Land Layer are the following: bare/sparse vegetation, cropland, herbaceous vegetation, open forest, and shrubs. Moreover, masking based on several thresholds is implemented so that pixels characterized by layover, shadow, or foreshortening phenomena for S1 images; radiometric saturation, terrain occlusion, or clouds for L8; EVI values greater than 0.5 for MODIS; and below a threshold temperature of 275 K and with SWE values greater than 0 kg/m² for GLDAS data are filtered out, as defined in [22]. The goal of this masking is to not assign values at points characterized by geometric SAR distortion, dense vegetation conditions, presence of snow, and frozen soil where the satellite data cannot estimate soil moisture. For all these cases and in the pixels not covered by the Sentinel-1 satellite pass under consideration, the algorithm assigns a null value to the pixel.

The SMC dataset made available covers 1 October 2019 to 30 September 2022. This way, three spring–summer seasons are available, like what was delivered for SCF. The algorithm also estimates the SMC in the winter season, although, in this case, considering that CRNS stations are located at high altitudes, the thresholds set by GLDAS severely limit the number of points at which an estimate is available and, sometimes, no estimate is available. In the latter case, the maps were discarded.

The dataset covers an area of 500 × 500 pixels centered on the 5 CRNS stations to be consistent spatially with the SCF dataset coming from S2 data.

3.3. SCF Validation

A very-high-resolution dataset was used to evaluate the quality of the developed product, as the lack of ground data prevents a systematic validation considering the diverse topographical and morphological conditions of the predominantly mountainous test area. Indeed, the use of very high spatial resolution images allows us to really verify the fraction of snow coverage at S2 subpixel level. This information is not available, for example, from automatic stations giving only the punctual and binary information relative to the presence or not of snow.

For this purpose, we obtained six WorldView images at 1 m spatial resolution, which we first prepared through a process of orthorectification and co-registration to make them comparable with S2-derived maps. The collected WorldView images used for verifying the reliability of the developed product do not always correspond to the areas covered by the maps presented here (see Livigno and Sonthofen in

Table 4. Comparison of SCF maps with VHR images: metrics and results.

	Date S2	Date WV	Bias	RMSE	Unbiased Rmse	Correlation
Corvara	13 April 2022	12 April 2022	−0.62	17.44	17.43	0.90
Dresdner	27 November 2020	27 November 2020	−2.65	17.51	17.31	0.91
Leutasch	27 November 2020	27 November 2020	3.45	22.38	22.11	0.86
Livigno	23 September 2021	21 September 2021	−16.33	34.19	30.05	0.51
Sonthofen	12 March 2022	14 March 2022	0.45	21.11	21.1	0.87
Weisssee	13 January 2022	13 January 2022	1.17	15.46	15.42	0.90

). Indeed, to obtain a dataset that covers as many regions with different characteristics as possible in terms of different topographies as well as different periods of the year or snow coverage, in some cases, we also exploited images close to the areas of the maps but not perfectly coincident with them.

Furthermore, a cross-comparison with the Copernicus Fractional Snow Cover (FSC) product (https://sdi.eea.europa.eu/catalogue/copernicus/api/records/3e2b4b7b-a460-41dd-a373-962d032795f3?language=all (accessed on 3 April 2024)) was performed to evaluate the developed product in complex situations as, for example, in case of shadow. This comparison, not really a validation, is useful to verify the goodness and the improvement of our product with respect to a reference product such as Copernicus snow product.

3.3.1. Very-High-Resolution Data Preparation

Six very-high-resolution (VHR) WorldView (WV) images were acquired following the criteria of selecting cloud-free scenes, acquisitions possibly coincident or very near to an S2 acquisition, and snow cover conditions that might be interesting to test as variable SCFs over the scene (mixed pixels due to vegetation or melting process). The WV images used were both WV2 and WV3.

First, orthorectification is needed. VHR images often contain geometric distortions due to sensor orientation, topographic relief, and Earth’s curvature. The orthorectification procedure aims to correct these distortions and transform the image into a georeferenced and geometrically accurate representation. These data are provided with an orthorectification kit consisting of the Rational Polynomial Coefficients (RPCs). These are the coefficients of a fractional polynomial function that link the image coordinates with the object space.

After being orthorectified, the image is co-registered and aligned to the S2 grid and reference system by keeping a final resolution of 1 m.

Once the WV and S2 grid are made comparable, the VW needs to be classified. Analogously to the classification technique applied for the S2 images, we also used, for this case, an SVM model trained separately for each VW image by manually collecting the training samples. The procedure is very similar to what is described in the previous section, with the main difference being that our target is a binary classification.

Finally, we aggregated the VW images to 20 m spatial resolution, thus obtaining the SCF to be compared with the S2 images.

3.3.2. Cross-Comparison with WorldView

The comparison with WV data was performed by analyzing different test sites through different metrics (bias, RMSE, unbiased RMSE, and correlation coefficient) to understand the goodness of the SCF estimation in different areas and variables conditions, such as different seasons, topography, and snow coverage conditions. Figure 4 shows an example of comparison between the developed product and the reference WV image classified: on the left side, the S2 map is shown, while on the right side, WV map is shown.

The following table (Table 4) reports the metrics calculated for the different areas, together with the dates of WV and S2 images, respectively, where the comparison with very-high-resolution images was performed.

In Table 4, we can see that, on average, the correlation is, for all test sites, around 0.9, except for Livigno, where it is 0.5. The reason behind this drop in performance could be related to the different acquisition dates in the two images, S2 and WV. Indeed, as you can see in Figure 5, before the WV acquisition, there was a snowfall event, which, considering the time of year (September 2021), was only a transient snowfall, lasting only a few days. In fact, it is no longer visible in the image captured by S2 just two days later. On the left side of Figure 5, the RGB false colors (FC) of SCF maps by S2 (up) and WV (down) are shown; on the right side, the classification results used for the comparison and the metric estimation are shown.

3.3.3. Cross-Comparison with Copernicus

To evaluate the snow cover fraction dataset, a cross-comparison with a standard snow product, such as the Copernicus high-resolution FSC product, was performed [29] (Dataset link: https://sdi.eea.europa.eu/catalogue/copernicus/api/records/3e2b4b7b-a460-41dd-a373-962d032795f3?language=all (accessed on 3 April 2024)). This comparison allows us to evaluate the performance of the developed product with respect to the reference product, especially in complex situations such as, for example, in case of shaded areas.

We performed the comparison for the five test sites in the period ranging from 1 October 2020 to 10 May 2023.

Table 5. Comparison between SCF maps derived from S2 and Copernicus product: number of the scenes compared, period, and extent.

Test Site	Processed S2	Available Copernicus	Matching Scenes	Period		Number of Pixels (Centered on the CRNS)	Extent
Corvara	105	212	101	1 October 2023	10 May 2023	500 × 500	717,320.0; 5,153,620.0; 727,320.0; 5,163,620.0
Dresdner	158	295	142	1 October 2023	10 May 2023	500 × 500	657,660.0; 5,202,080.0; 667,660.0; 5,212,080.0
Leutasch	163	295	162	1 October 2023	10 May 2023	500 × 500	658,180.0; 5,244,180.0; 668,180.0; 5,254,180.0
Obergurgl	68	327	66	1 October 2023	10 May 2023	500 × 500	649,810.0; 5,185,346.0; 659,810.0; 5,195,346.0
Weisssee	184	329	177	1 October 2023	10 May 2023	500 × 500	625,640.0; 5,187,460.0; 635,640.0; 5,197,460.0

lists the number of analyzed scenes for each test site, the period, and the extent to which the comparison took place.

Starting with the available scenes, we computed bias, unbiased RMSE, RMSE, and cross-correlation to compare the two products. The results are shown in

Table 6. Results of the intercomparison between SCF product developed and the standard product derived from Copernicus.

Test Site	Bias	RMSE	Unbiased RMSE	Correlation
Corvara	−0.06	16.8	16.05	0.73
Dresdner	5.29	21.89	20.27	0.69
Leutasch	0.76	17.73	18.54	0.61
Obergurgl	0.01	23.74	22.87	0.66
Weisssee	3.74	20.69	19.48	0.72

.

The results of Table 6 show good agreement between two products, and the differences highlighted by the metrics could be related to the different behavior in specific situations: in particular, we found significant improvements in case of snow detection in the shaded areas. This is shown in Figure 6, where two scenes from proposed method and Copernicus product over Weissee area are compared with reference to a false-color (FC) RGB image highlighting shaded areas (the left one). As the reader can see, the large “black” (shaded) areas are often classified by Copernicus as snow-free areas, whereas the product presented manages to capture even these situations by correctly classifying snow pixels. The presence of snow in these situations was verified by a careful visual inspection and through a slope exposure evaluation. The ability in shaded pixel snow detection of the developed product leads, therefore, to a general improvement with respect to the reference product, Copernicus.

3.4. SMC Validation

In [22], the SMC algorithm was already validated at a global scale. However, in this work, a few alpine automatic monitoring stations measuring soil water content (SWC) in Val Mazia (South Tyrol—Italy) were considered (Figure 6) for comparisons with SMC maps estimated by the ML algorithm [30]. These stations record meteorological and biophysical variables of the long-term socio-ecological research LT(S)ER site Matschertal/Val di Mazia (https://browser.lter.eurac.edu/ (accessed on 30 October 2024)). The data can be downloaded from this website after registration. This approach was considered because most of the ISMN network used to train the SMC algorithm is not located in alpine or, in general, in mountainous areas like those considered in the project. Consequently, this comparison can help to better understand the behavior of the algorithm in this specific environment. Among many parameters, like air temperature, relative humidity, precipitation, and wind speed, these climate stations record the SWC at different depths from 2 to 50 cm, averaging 15 samples taken every minute, aggregating them into a single value. Considering the penetration characteristic of the SAR signal, only the measurements at 2 and 5 cm were considered. For comparison, the satellite acquisition time was considered, and the ground measurements from one hour before to one hour after that time were further averaged and utilized.

The stations were equipped with Campbell Scientific TDR (Time Domain Reflectometry) sensors able to measure soil temperature and SWC (range: 0 to 52%, accuracy: ± 3% for Electrical Conductivity ≤ 10 dS/m). They were calibrated using a standard procedure for a generic soil type.

The monitoring stations are located in the western part of South Tyrol (Val Mazia catchment, an inner-alpine dry valley in the Italian Alps), and their locations are indicated in Figure 7 and shown in Figure 8. The four stations were selected among a network of 24 climate stations considering their peculiar characteristics, especially in terms of elevation, exposure, slope, and landcover type, as observed in

Table 7. Climatic stations described in terms of GPS coordinates, elevation, land use, and soil characteristics.

Station	Latitude [DD]	Longitude [DD]	Elevation A.S.L. [M]	Aspect [°]	Slope [°]	Land Use	Soil Type	Texture
B1	46.661183	10.590244	980	230	12	Irrigated Meadow	Brown Earth	Sandy Loam
P2	46.684305	10.585125	1540	230	27	Pasture	Brown Earth/Ranker	Loam
B3	46.691694	10.591936	1920	220	11	Pasture	Ranker	Loam
M2	46.711303	10.691794	2210	250	9	Pasture	/	/

. By referring to the period May–October of the three observation years, 2020–2022, a set of metrics was considered based on what is described in [31,32,33]. Mean Absolute Error, bias, and Unbiased Root Mean Square Error are reported in

Table 8. Performance of the comparison between SMC in m³ m⁻³ estimated by the algorithm and measured by the monitoring stations. The suffix _2 and _5 indicate the depth of the station in centimeters.

Station	MAE_2	MAE_5	URMSE_2	URMSE_5	BIAS_2	BIAS_5
B1	0.055	0.047	0.054	0.048	−0.043	−0.035
P2	0.086	0.083	0.091	0.086	0.046	0.050
B3	0.059	0.107	0.071	0.104	0.018	−0.076
M2	0.055	0.050	0.068	0.061	0.006	−0.006

. The selection of the study period considers that these months have the lowest probability of snow cover, making it easier to obtain a consistent and continuous series of SMC maps, which is helpful for time series analysis.

Figure 9, Figure 10, Figure 11 and Figure 12 show examples of the trends of the SMC estimated by the algorithm compared to that extracted from the automatic meteo stations at 2 and 5 cm. The graphs for different ground stations and years are related to May–October, when the stations are usually not affected by snow, to highlight the temporal trends. The measurements and estimations are acquired on the same day, as described above. In 2022, the number of SMC estimates from remote sensing data was reduced as only the S1-A satellite was available then.

The three curves show similarities in the trends, although the estimated values present a more compressed dynamic. A possible reason for this can be related to the temporal filter applied to SAR data to reduce the speckle of the SAR images, which is fundamental at this spatial resolution. The measurements at 2 and 5 cm are very similar except for B3 and M2 stations, which are a meadow and a pasture station, respectively, but are characterized by a flat terrain. In the B3 case, where the dynamics are broader, the estimated value is closer to the ground measurement at 2 cm, while in M2 case, the estimation over the entire analysis period is between the two ground estimations.

Comparison with CRNS and Point-Scale SMC Data

A further comparison with CRNS illustrates a potential use case for the data at the site of Leutasch. The Leutasch site is part of the COSMOS Europe network [34]. The CRNS-based SMC values were obtained using the calibration procedure described in [35]. The data need to be corrected for changes in incoming cosmic radiation, atmospheric pressure, and air humidity. The residual signal is highly sensitive to changes in hydrogen content in soil, vegetation, and snow [36]. It should also be noted that the ML product and CRNS estimates are related to different spatial resolutions, depths, and calibrations. For example, the soil moisture CRNS signal is sensitive to a soil depth of up to 80 cm in dry conditions [37], where satellite observations can only detect the moisture in the first centimeters of the soil. Furthermore, ML algorithm has a spatial resolution of 20 m, and CRNS gives a value corresponding to a large area of several hectares. Therefore, due to different integration volumes and spatial coverages, the combination of remote sensing and in situ-based SMC data can give further insights into the hydrologically relevant dynamics. This is particularly true if combining spatially representative CRNS data with regionally available remote sensing products.

However, because of the larger measurement volume, in particular, with regard to the depth of the signal, the CRNS data show different absolute values with respect to ML data. To better understand the behavior of two datasets, we thus compare the mean-centered time series (Figure 13 and Figure 14). Deviations in temporal dynamics are plausible by considering the differences in the measurement volume and depth. This demonstrates that the two datasets can be used as complementary sources of knowledge by providing different information about the same variable (SMC).

4. User Notes

4.1. Data Access

The two datasets presented in this work are freely accessible as a supplement to this paper in the journal Data and are organized as follows: for each CRNS station, an NCDF file is available for both the SCF and SMC products, providing the entire time series for the study period.

4.2. Example Usage

Although the products show some limitations (as presented in Section 4.3, Cautionary Notes), the results show an improvement with respect to other products currently available for this type of environment. Regarding the SMC product, in fact, radiometer-based soil moisture estimations have a very coarse resolution compared with the heterogeneity of this type of environment (tens of kilometers), considering that, for example, the farthest stations considered in the comparison are less than 10 km apart. Even compared to higher resolution products, such as the 1 km resolution Copernicus product, this approach is more informative because the alpine areas in Copernicus are almost completely masked (https://land.copernicus.eu/en/technical-library/algorithm-theoretical-basis-document-surface-soil-moisture-version-1/@@download/file (accessed on 30 October 2024). Regarding the SCF product instead of the “standard” product (typically used as a reference product), Copernicus, proved to be less efficient in snow detection in the shaded areas than our product by showing an improvement in our approach.

For these reasons, the developed products can be exploited in different applications, for example, as inputs for hydrological models: complementary information sources can indeed be exploited to improve model output and reduce uncertainty by overcoming limitations of individual data related, for example, to an inadequate scale representation or to the uncertainties of the single products, as reported in the following section. If relying on one data source only, errors and biases propagate into hydrological modeling [38]. The use of high-resolution products developed in this work could refine assimilation models by better characterizing local spatial and temporal patterns. Another example of usage is the analysis of the temporal behavior of the soil moisture and the snow cover fraction over a specific point of interest included in the area. This analysis can be useful to study trends and anomalies—for example, extreme drought characteristics in terms of onset and duration.

4.3. Cautionary Notes

A few cautionary notes are necessary before using the datasets presented in this paper. The uncertainties related to the proposed products are due to several factors, as explained below, and must be kept in mind when using the data.

Concerning snow maps, because of the lack and nature of station data, no comparison has been performed with this type of measurement. The stations provide, when available, punctual measurements of snow height at the station location, and this kind of data is not easily comparable with a spatialized value derived from the SCF map where a percentage of snow cover for each pixel (20 by 20 m) is provided. However, an inter-comparison with very-high-resolution products (WorldView) and with what is considered a reference product (Copernicus) was performed, and the results are encouraging by suggesting a valid and reliable product.

Regarding the SMC product, it should be noted that the model used was obtained starting with the ISMN stations, which are mostly located in the USA, and, in any case, none of them are in the Alps or in European mountain areas. The consequence is that considering URMSE, the error obtained in the tested alpine sites is higher with respect to 0.04 m³·m⁻³ obtained in the validation executed at a global scale by [22]. Moreover, although the automatic stations in Val Mazia (used for the validation in this paper) cover different kinds of mountainous conditions in the Alps, their locations do not correspond to those of CRNS probes, where the SMC time series are provided. Nevertheless, the tests carried out show that the data follow a realistic temporal trend. Another limitation is that the algorithm can exclude many points from the SMC estimation because of the masking procedure described in Section 3.2. For example, regarding forested areas, the only type considered is that one classified as “open forest” by the Copernicus Global Land Cover Layer (CGLS-LC100), which has a spatial resolution coarser than the spatial resolution of the SMC maps.

Finally, both satellite products provided were obtained by exploiting the use of optical sensors (S2 for SCF and L8 for SMC), which are reliable instruments for observing geophysical parameters from space. In particular, the distinctly high reflectance of snow in the visible spectrum and its highly absorptive nature in the short-wave infrared wavelengths of the electromagnetic spectrum are captured by multispectral sensors [39]. Nevertheless, optical-based satellite sensors are impacted by atmospheric conditions (e.g., cloud presence, [40]) and by land cover conditions (e.g., forest presence, [41,42]). Cloud coverage works well for parameter estimation with optical sensors if, on one side, the optical signal cannot penetrate the cloud coverage by preventing parameter detection in that portion of the image; on the other hand, in the case of snow detection, cloud cover can create errors in snow-detection algorithms because of the similar reflectance of clouds and snow at wavelengths utilized for snow identification. Thus, misclassification errors can occur between snow and clouds during cloudy days.

The other challenge for geophysical parameter retrieval from optical sensors is forested areas because the sensor only detects the viewable soil fraction of a satellite pixel (snow or soil moisture in this case). In the case of snow, optical sensors only view snow that is visible beneath leafless deciduous stands (such as aspen), in clearings and forest gaps between coniferous trees (i.e., pine, spruce, and fir) and through thin foliage or needles, or snow that has been intercepted by the forest canopy. Since the study area is in a mountainous environment, this kind of error must be considered by the user who wants to exploit these products in forested areas.