In 2021, the European Commission published the renewed Soil Strategy (European Commission, 2021), in which the vision for healthy soil ecosystems by 2030 was presented. Soil health, defined as “the continued capacity of soils to provide ecological functions for all forms of life, in line with the Sustainable Development Goals and the Green Deal” (European Commission et al., 2020), is still an evolving term and concept. It received heightened attention since the implementation of the EU Mission: A Soil Deal for Europe (European Commission et al., 2020) and particularly the discussed Soil Monitoring Law (European Commission, 2023). Narrowing down the definition of soil health and, most importantly, quantifying it has therefore become a key issue for soil scientists. Characterizing soil health in the form of multiple indicators and creating a “soil health index” (Lehmann et al., 2020) suitable to measure and monitor soil health in the EU (European Commission, 2023; Matson et al., 2024) are pivotal in achieving healthy soils until 2030. Efforts to define such an index are multi-fold (Matson et al., 2024; Wade et al., 2022; Rinot et al., 2019), but generally they agree on including chemical, physical and biological indicators that go beyond crop production to encompass soil ecosystem services (Shen and Teng, 2023). Properties such as soil organic carbon (SOC), soil nutrients, pH or cation-exchange capacity (CEC) are frequently included in soil health assessments (Lehmann et al., 2020). These properties are routinely analyzed in laboratories with traditional methods but are often time-consuming and require resources such as expensive equipment and chemicals. The increasing requirements for soil health assessments and monitoring posed by the Soil Monitoring Law are creating a demand for less cost-intensive alternative methods (Safanelli et al., 2025).

The application of soil visible (vis) and near infrared (NIR) spectroscopy to predict soil properties, particularly chemical ones, using statistical and machine learning methods has increased rapidly in recent decades (Viscarra Rossel et al., 2011; Gholizadeh et al., 2013; Stenberg et al., 2010). Spectroscopy was shown to generate fairly to very accurate estimates of e.g., total carbon (Ma et al., 2023), SOC (Guerrero et al., 2016) and its fractions (Jaconi et al., 2019b), as well as soil texture, specifically clay (Jaconi et al., 2019a), and carbonates (Tavakoli et al., 2023). This approach also bears great potential for soil fertility assessments of total nitrogen (Park et al., 2024) along with total and critical available phosphorus (Recena et al., 2019). Compared to traditional laboratory analyses, spectroscopy has many advantages because it is fast, simple, cost-effective, reproducible, repeatable, non-destructive and environmentally friendly (Viscarra Rossel et al., 2006; Nocita et al., 2015; Soriano-Disla et al., 2013). To improve predictions and fill data gaps, large reference training data sets, so-called soil spectral libraries (SSL), are being built and often made freely available. These extend from local, regional to national and even global scale (FAO, 2022), thereby helping to describe soils and their health whilst improving soil data availability (Cornu et al., 2023). Nevertheless, large areas without available data in the global coverage of NIR SSL remain, calling for active participation to fill these gaps (Viscarra Rossel et al., 2016; Safanelli et al., 2025). This is necessary because the predictive capacity of models relies on the number and diversity of soil samples and conditions represented in the spectra. Concurrently, new methods for coordinating existing SSL are being developed, enabling interoperability between labs, data harmonization, engagement of communities and model development (Safanelli et al., 2025; Peng et al., 2025).

The application of NIR spectroscopy for Austrian soils is limited so far. Beyond sample analysis within the LUCAS inventory (Fernandez Ugalde et al., 2022), Ludwig et al. (2023) tested the suitability of vis-NIR and MIR (mid infrared) spectroscopy for forest soils whilst comparing different modelling approaches. Moreover, the suitability of vis-NIR for measuring soil carbon contents and the effects of agricultural management methods was analyzed by Bieber (2023) on a regional scale. To date, no national SSL for Austria is openly available. By using available legacy soil samples, including their results from chemical and physical analysis, we wanted to fill this gap and analyzed them via NIR spectroscopy. Therefore, the objectives of this study were (i) to provide a first dataset on Austrian NIR soil spectra and reference laboratory analysis for several soil health properties, covering all Austrian environmental zones and (ii) to apply partial least square regression (PLSR) for model calibration and validation and test the model's applicability for national soil health assessments.

The selection of soil samples for the Austrian NIR soil spectral library was based on a wide distribution of soils with different properties to strengthen the dataset usability and predictive power of applied models. The dataset includes legacy soil samples (analyzed and stored in the AGES archive) from several long-term field experiments, different project campaigns and from farmers' land. Moreover, samples from the so-called AGES “soil box” (AGES soil box, 2025), which are sent in by private persons, were included. The soil box allows individuals to have their soils analyzed for common properties such as pH, SOC and phosphorus. These samples are special because they stem from a wide range of land uses including grassland, arable land, forests, orchards, hedges or lawn, but also compost or garden soils. This particularly promotes the diversity of included soils and enhances the geographical distribution and coverage throughout Austria. All included samples were collected between 2016 and 2023 and soil sampling depth ranged between 0 cm down to a maximum of 110 cm (mean: 24.5 cm). This yielded a total of 2129 samples which were considered for the soil spectral library. Most samples represent the Pannonian zone (n=1059), followed by the Continental (n=778), Alpine South (n=286) and Mediterranean Mountains (n=5) zones (Fig. 1). For one sample, the location and environmental zone are unknown (Sample_number 743), hence it is not included in Fig. 1.

Information about the soil samples included in the dataset was extracted from the AGES internal database “LISA”. Missing data regarding zip codes and land-use type was gathered by contacting responsible persons within and outside of AGES. MS Excel© (version 2410) and the statistical programming language R within the RStudio interface (version 2024.04.0) were used for data processing. The dataset contains 1846 columns, of which columns 1 to 11 provide information on the soil sample, columns 12 to 25 contain chemical and physical reference analysis results, and columns 26 to 1846 contain the measured absorbance from 680 to 2500 nm for every nm. Column “Sample_number” assigns each sample a number, starting from one, and column “Sampling_year” shows the year of soil sampling. In column “Experiment_number” the samples were grouped into experiments which indicate that samples were collected on the same site. When several samples were sent in together but originated from different sites (e.g., many sites sampled within one research project), this was indicated by an underscore (e.g., 9_1, 9_2). Column “Sample_source” provides information on the source of the soil sample, e.g. whether it was taken from a long-term experiment (LTE), part of a research project sampling campaign, or submitted by a private person or advisor for analysis. The columns “Zip_code” and “Municipality_code” indicate the location of the sampling site, which in some cases needed to be approximated due to limited location information. For the majority of samples, coordinates were not available, as they were either sent by private individuals and farming advisers, or were taken by other research facilities than AGES. Therefore, only zip and municipality codes are provided in our dataset. The environmental zone (Metzger et al., 2005; Metzger, 2018) of the sampling location is shown in column “Environmental_zone”. Sampling depths (sampling_depth_from and sampling_depth_to) are reported in columns 8 and 9, respectively, and land use types in column 10. Most samples were taken from arable land (n=1485), followed by vegetable plots (n=202) and lawn (n=174). Other land use types (orchards, forests, grassland, hedges, ornamental plants and vineyards) were sampled less than 100 times. For 73 samples the land use type was unknown. Lastly, column 11 indicates whether samples were sent in for analysis with a “soil box”. As these samples are taken and submitted by private individuals, the land use type defined by those individuals might not always be accurate. Finally, such samples may contain artificial substrate, high amounts of compost or fertilizers. Regarding the chemical and physical reference analysis and NIR spectroscopy, the methodologies and results are described in detail in the following sections. The dataset is available at https://zenodo.org/ (last access: 19 December 2025) in the form of an excel file, which is accompanied by document metadata and a legend (Fohrafellner et al., 2025).

The Kennard–Stone algorithm (Kennard and Stone, 1969) of the “prospectr” package (version 0.2.7) (Stevens and Ramirez-Lopez, 2024) in R Studio (version 2024.04.0) was used for calibration and validation set sampling, as it allows to create sets that have a flat distribution over the spectral space (Stevens and Ramirez-Lopez, 2025). The dataset was split into 80 % calibration and 20 % validation samples. To avoid pseudo-independence of spectra, grouped samples from the same sites were either assigned to the calibration or validation set via the algorithm. Simultaneously, a uniform distribution of samples was ensured by applying the Mahalanobis distance, which first performs a PCA on the input data and computes the Euclidean distance (Stevens and Ramirez-Lopez, 2024). As these settings generated a very small validation set for clay in suspension (n=7), a separate split using only samples with data of this soil property was conducted. Here, the Euclidean distance, which selects points which are the farthest apart from its closest neighbor (Stevens and Ramirez-Lopez, 2024), was used because the Mahalanobis distance requires more samples than individual spectra for point distance computation. The UCal™ Chemometric Software (Unity Scientific, Brookfield, CT, USA) was then used to apply the partial least square regression (PLSR) for model calibration and validation. A maximum of 15 factors were set, and outlier detection and removal were performed by calculating the Mahalanobis distance, which was set to a maximum global distance of 13 and 3 for the calibration and validation sets, respectively. The spectra were trimmed to the range of 780–2400 nm and sensor switches between 1330–1350 nm were removed. After model calibration, predictions were generated based on the validation set for each soil property. Models were tested for all samples or for subsets, i.e., SOC values below 7 %, as described in Sect. 4.1. This was done as the modelling results revealed that using subsets with SOC<7 % generated better estimates of the properties SOC, SOC/clay ratio and TC, as it is the upper limit for most agricultural mineral soils in Austria. Statistical measures used to determine the performance of the model were the standard error of prediction (SEP) and the coefficient of determination (R2) from linear regression as well as the ratio of performance to inter-quartile distance (RPIQ) (Bellon-Maurel et al., 2010).

NIR estimates of the 14 soil health indicators (Table 2 and Fig. 4) revealed that PLSR explained total nitrogen contents (TN) best (R2=0.9, SEP=0.05), followed by CaCO3 (R2=0.89, SEP=3.95), soil organic carbon (SOC) (R2=0.82, SEP=0.45) and clay (R2=0.74, SEP=4.35). Nonetheless, the standard error of prediction was large for these properties. This is a common issue in larger-scale spectral libraries because they can comprise a wide range of soils with high mineralogical variability (Nocita et al., 2015; Stenberg et al., 2010), as it is the case for Austria. Regarding SOC, the regression shows that up to concentrations of 4 %, the model predicted SOC well, but with increasing concentrations it underestimated NIR predictions due to the limited number of samples. Such limited numbers of samples for “extreme” values impacted the predictive quality of several other soil parameters, particularly phosphorus or sand content. Although SOC and clay results were sufficient, the SOC-to-clay ratio was not well estimated by the model (R2=0.49, SEP=0.08); only narrow ratios showed good prediction by NIR spectroscopy. For total carbon (TC), only a small sample size for validation (n=30) was available, which nevertheless revealed potential for accurate prediction by NIR (R2=0.68, SEP=1.02). Labile C had few low and high reference values, which impacted the prediction on NIR negatively (R2=0.54, SEP=83.79). When comparing the results for soil pH, pH in acetate achieved better estimates than pH in CaCl2 (R2=0.67, SEP=0.2 and R2=0.61, SEP=0.61, respectively). Phosphorus values above 200 mgkg-1 were underestimated by NIR, yielding overall poor model results (R2=0.2, SEP=82.84). Clay analyzed in suspension had a moderate coefficient of determination (R2=0.58, SEP=2.71). Other soil properties, i.e., cation exchange capacity (CEC), sand and silt, achieved R2 below 0.5.

We also calculated ratio of performance to inter-quartile distance (RPIQ) for all 14 soil health indicators. The proposed threshold for satisfactory model performance of a RPIQ of >1.89 (Ludwig et al., 2019) was used. Based on this metric, CaCO3 (RPIQ=5.06) performed best, followed by TC (RPIQ=3.33), clay (RPIQ=2.52), clay in suspension (RPIQ=2.21), pH in CaCl2 (RPIQ=2.20), SOC (RPIQ=2.09) and TN (RPIQ=1.95). The other properties achieved RPIQs below 1.89. As previously described by Ludwig et al. (2019), it should be acknowledged that the threshold of 1.89 is predicated on the work by Chang et al. (2001), whose values were not based on theory or experiment, and that model performance must be evaluated according to the specific context. When assessing the suitability of NIR as a substitute for routine laboratory measurements, particular attention should be given to the SEP, since even high R2 or RPIQ values cannot compensate for a large SEP that indicates limited precision. Accordingly, we concentrated our assessment on R2 and SEP as key indicators of model performance.

The results of data validation show that applying partial least square regression to the Austrian NIR soil spectral library yields estimates of TN, CaCO3, SOC and clay contents with moderate, and other properties with low accuracy. Although several properties were predicted with an R2>0.7, standard errors of prediction were generally high. Considering the applicability for routine soil health analyses and monitoring, the predictive quality is currently insufficient compared to standard laboratory analyses. Nevertheless, in cases where funding is limited or rough estimates are sufficient, the tested models could be of benefit. Other user-cases might include the validation of citizen science data or the estimation of soil texture, as laboratory analysis is a tedious process. Increasing the sample number, particularly in the Alpine South, and including samples with “extreme” values, could improve the predictions and reduce standard errors. Grouping by environmental zones or according to the parent material are further options to assess for model improvement. Testing the performance of other models, particularly more advanced machine learning algorithms such as random forest, cubist, artificial neural networks or support vector regression (Viscarra Rossel et al., 2016; Minasny et al., 2024; Minasny and McBratney, 2008), could further enhance the useability of the Austrian NIR soil spectral library. Lastly, we encourage merging with other SSLs, for example by using transfer functions (e.g., calibration transfer, Feudale et al., 2002, harmonization functions, Francos et al., 2023) or spiking for adaptation of models to the characteristics of the target sites (Guerrero et al., 2016) and using them for developing localized calibration sets for specified contexts and pedologic domains (Viscarra Rossel et al., 2022). A weakness of our dataset is the missing coordinates, which, due to inclusion of soil samples taken by private individuals, farming advisors, and external research organizations, were not available.

The dataset described in this manuscript can be accessed at 10.5281/zenodo.15772618 (Fohrafellner et al., 2025).

We present a first Austrian near-infrared (NIR) soil spectral library with over 2100 legacy samples and test the performance of the partial least square regression (PLSR) to predict 14 chemical and physical soil health indicators. Several properties, i.e., total nitrogen, soil organic carbon, CaCO3 and clay were estimated with moderate accuracy, whereas the potential for predicting other indicators varied. We conclude that applying the PLSR to the Austrian NIR soil spectral library is currently not suitable for precisely estimating soil properties compared to routine laboratory analysis, but that the library presents a valuable foundation for future soil health assessments. Enlarging the spectral library would no doubt improve the predictions, specifically by including extreme values outside of common ranges and by including data from the Alpine South. Moreover, testing novel models using machine learning algorithms to train and validate the spectral library, for example, could improve predictions of soil health indicators. We therefore encourage the use of this open dataset and of merging the spectra with other existing or forthcoming libraries. This effort is an important step forward in supporting the expansion of soil spectral libraries globally, facilitating the amplification of NIR analysis as a fast and simple method to assess, monitor and map soil health.