We conducted a Scopus search on 17 February 2021 for field decomposition studies of all times from aridlands published in peer-reviewed journals. We used the search words “arid” OR “dry season” AND “decomposition” and excluded results from unrelated subjects. This search produced a list of 1142 publications. To be included in the database, studies additionally needed to fulfill certain criteria: (a) field studies in which leaf, shoot or root litter of terrestrial plants was used, and (b) minimum of three time points of mass loss data. We did not include wood or dung decomposition studies. We also included publications from our personal libraries. In total, this left us with a list of 184 eligible publications.

We named the database aridec and uploaded it to a repository in GitHub (GitHub, 2022; Sarquis et al., 2022). From each publication selected, we created a database entry consisting of three separate files: a file containing time series of mass loss (timeSeries.csv), a file containing metadata of the study site and the experimental setup (metadata.yaml), and a file with relevant information of the initial characteristics of the litter at the beginning of the experiment (initConditions.csv). We saved each entry in an individual folder named after the last name of the first author and the year of publication. If there was more than one paper per author and year, we added lowercase letters to differentiate them (e.g., Austin2006a and Austin2006b). We included all entries inside the data folder. Other folders in the repository include the Rpkg folder containing an R package for querying and manipulating the database, a test folder with scripts for testing the integrity of the data and the R package, and an additional folder with miscellaneous scripts that demonstrate additional functionality. The overall structure of the database is similar to the structure of SIDb (Schädel et al., 2020), a database of soil incubation time series, and contains the following folder structure: –

aridec –

Rpkg

The metadata.yaml file includes additional information reported in the original paper. It is a yaml type file (“YAML ain't markup language”), which allows us to write lists of items in a hierarchical form and is both machine and human readable. It includes four main sections: entry identification data, the siteInfo, the experimentInfo, and the variables sections. A template for this file, with a full description of how to complete it, is available inside the data folder. The first part includes citationKey, which is a unique identifier for the whole entry in the format LastnameYEAR (lowercase letters must be added when there are two or more entries by the same author and year, i.e., LastnameYEARa and LastnameYEARb). This citationKey name should be the exact same as the folder name. Next is the doi, which stands for the digital object identifier where data is published. entryAuthor and contactName are both the first and last name of the person who wrote the entry file and their supervisor (only if applicable), respectively. If the entry author works independently of a supervisor, both fields should be filled with the same name. contactEmail should be filled with the supervisor's e-mail address. entryCreationDate stands for the date when the file was created, following the format: YYYY-MM-DD. entryNote should include any notes or comments related to this entry, such as missing data or additional data sources used to complete this file. Lastly, study requires a short study description of not more than one sentence.

The second part of the metadata.yaml file is the siteInfo section, which includes environmental information of ecological interest from the study site. First is the site field that requires an identification name for the site (not necessarily the site's real name). If the study includes more than one site, an array format should be used in this field, and the rest of the items in this section should be arrays of equal length. The coordinates field should be completed using decimal units, checking for the negative sign that denotes Southern and Western Hemispheres. If absent from the publication, coordinates can be approximately obtained from Google Earth (Google LLC, 2020). The country field should be completed avoiding full names (e.g., “China” instead of “People's Republic of China” or “USA” instead of “United States of America”). Mean annual temperature (∘C) and mean annual precipitation (mm) should be entered in the fields MAT and MAP, respectively. When climatologic data are absent from the paper, they can be retrieved from other databases like the POWER database (NASA Langley Research Center, LaRC, 2021). The rainySeason field should be filled with either one of five options: whole year, spring, summer, autumn, winter; if precipitation does not follow a unimodal pattern, this item is left blank. Elevation of the study site in m a.s.l. should be entered under the elevation field, which if absent from the publication can be retrieved from other sources such as Google Earth. The type of vegetation cover of the site should be specified in landCover, with possible options: marsh, greenbelt, farmland, mangrove forest, subalpine, shrubland, urban, sandland, forest, steppe, desert, grassland, and savanna. The item vegNote should include a short description of not more than one sentence of the species or functional type composition at the site, if available. The cover item should be completed with percentage values of total plant cover or with cover values for specific plant functional types, as available. Lastly, the soilTaxonomy item must be completed using the taxonomic classification of the soil at the site. If the classification system used in the paper is unknown, it is better to leave this section blank, for exact equivalences between soil classification systems are unlikely.

The third part of the metadata.yaml file is the experimentInfo section, which includes information regarding the experimental design of the study. incDesc stands for incubation description and must include a short list of treatments and sampling points in time. The number of replicates should be specified, paying attention to occasional pseudo-replication in decomposition studies. The experiment duration in days should be completed with the maximum time length that samples stayed in the field. The month in which the experiment started should be specified under startingMonth. The name of the litter used for the experiment should be specified under litter, and it should match the name used in the initConditions.csv file (see below). Under the litterbag field, many sub-fields for different characteristics of interest should be completed, such as mesh material, mesh size (one side of a square in mm), dimensions (in cm), mesh transmittance (as a percentage of full sunlight), and litterbag position (full list of options available in the template file). A general rule for the experimentInfo section is that when there is more than one option for a field, they should be considered as different levels of a treatment. In this case, that field should be left blank in this section, and a new field should be created in the variables section by replacing the experimentalTreatment placeholder in each variable (see below).

The last section of the metadata.yaml file is the variables section, which serves as a link between columns in the timeSeries.csv file and metadata. Thus, this section should have as many variables as columns in the timeSeries.csv file. The first variable (V1) must always be called “Time” and only time units should be modified accordingly. The rest of the variables (V2 to Vn) must be adequately edited to represent treatment application as described in the original publication. Variable names should match column names in the timeSeries.csv file. Litter mass loss units should be expressed either in (dry) mass remaining, organic matter remaining, or C remaining. In our database, organic matter remaining is a synonym of ash-free dry mass remaining. This is because the ash-free dry mass correction assumes that ash is inorganic matter, and thus ash-free mass is equivalent to organic matter for the purpose of this database (Harmon et al., 1999). Under varDesc (as in variable description) one should write a brief sentence indicating specific treatment levels applied to this variable. The site field should be completed using the same site name entered in the siteInfo section. The experimentalTreatment item is a place holder for treatments with multiple levels. It should be replaced by any of the listed variables in experimentInfo and completed with an appropriate treatment level. In compTreat complementary treatments not included in the rest of the metadata items should be indicated using key words (e.g., grazed, ungrazed, water addition, control, etc.). Finally, transmittance and wavelength threshold (nm) data for radiation filters should be indicated under filter. This sub-section should be completed only for photodegradation studies.

The last file in the data folder is the initConditions.csv file, which contains details on the plant litter substrate used for each experiment. The first row contains column names. The first column name is species and is the only mandatory item; nonetheless we strongly recommend completing all items, if possible. We suggest checking for the correctness of scientific names in the Global Biodiversity Information Facility database (GBIF.org, 2022). Names in the species column should be used to complete the litter item in the metadata.yaml file. Four options are valid for the type column: deciduous or evergreen (for woody plants) and forb or graminoid (for herbaceous plants). For the N-fixer item, we recommend consulting the NodDB database (Tedersoo et al., 2018). Units for the sample amount column are in g, for the nutrients and fibers in percentage, and for SLA (specific leaf area) in mm2 mg-1. When litter quality traits are not provided in the original paper, they can be obtained upon request from the TRY database (Kattge et al., 2020). We created a template for the initConditions.csv and a README.md file with further instructions in the data folder. Special attention should be paid to the material section of the README.md file, for litter substrates are highly variable among studies, and this is key for database consistency. In Fig. 1, we present a flowchart with the full process of entry submitting for potential contributors.

We generated a global aridity index (GAI) map with the study sites from the database. We retrieved GAI data from the Consortium for Spatial Information global climate datasets (CGIAR-CSI; Trabucco and Zomer, 2019). This index is calculated after dividing the mean annual precipitation by the mean annual reference evapotranspiration. The raster dataset that we used is based on WorldClim2 database (Fick and Hijmans, 2017). We chose this dataset because it encompasses a relatively long period of time (from 1970 to 2000) and it has a high spatial resolution (∼1 km at the Equator). We then classified each study site by its GAI value as hyper-arid (0–0.05), arid (0.05–0.2), semi-arid (0.2–0.5), dry sub-humid (0.5–0.65), and humid (>0.65; United Nations Environment Programme, 1997). Complementarily, to explore how representative our sites in aridec are of the whole climatic range of aridlands, we first made a random point sampling of 6793 pixel units separated at least 1 km away from each other within the range of aridlands (GAI: 0–0.65). For each sample, we averaged mean monthly temperatures from WorldClim2 to obtain mean annual temperature values. We did the same for our aridec coordinates and plotted both sets of data together to evaluate how well aridlands are represented in our database. We used the QGIS software to process data and created a map (QGIS Development Team, 2021).

A central application of this database is the development of models of litter decomposition for aridlands. In an attempt to explore what the most appropriate model structures to integrate with data from the database are, we selected different structures of decomposition models based on recent theory of models of organic matter decomposition (Sierra and Müller, 2015). These model structures are already implemented in the SoilR package (Sierra et al., 2012), and we provide here an interface between our database and this R package. SoilR is a modeling framework that contains a wide set of functions and tools to model soil organic matter decomposition within the R computing platform (R Core Team, 2020).

Organic matter decomposition in SoilR is represented by systems of linear differential equations that generalize most compartment-based models. A simple general structure to represent litter decay with no inputs follows Eq. (1): 1dC(t)dt=AC(t),Ct=[Cpool1,…,Cpoolm]T,A=-k1⋯a1i⋮⋱⋮aj1⋯-km, where C(t) is a m×1 vector with m pools of litter mass observed at time t, and A is a square m×m matrix that contains decomposition rates (km) for each pool and transfer rates (aij) between them. These different pools may correspond to different ways in which the quality of the litter is expressed in different studies. For example, they may correspond to different compounds obtained from a specific extraction method (e.g., water soluble sugars or acid detergent lignin), or they can be defined by general decay classes such as fast and slow decay compounds. These pools have different decomposition rates, pool 1 being the fastest decomposing pool and pool m being the slowest. The linear dynamical system represented by Eq. (1) has many different solutions, but we are only interested in the solution that satisfies 2C(t=0)=C0=[total C0⋅p1,…, total C0⋅pm]T, where C0 is an m×1 vector with the value of initial litter mass content in the different compartments m. Total C0 is set to be 100 % in SoilR for our database, and the resulting parameters pm are the initial proportions of litter in m pools. Using this framework, we chose to fit a total of five different models with an increasing number of parameters (Table 1).

For this set of models, we performed an identifiability analysis following the procedure described by Soetaert and Petzoldt (2010). Non-identifiability is a common issue with inverse-modeling approaches. It is a type of model over-parameterization that makes precisely determining parameter values virtually impossible; thus parameters are “non-identifiable”. When parameters are functionally related, changes in one parameter can be compensated by changes in others. This produces different parameter sets that have similar probability distributions, thus the inability to determine a single parameter set for the model (Sierra et al., 2015). Analyzing for parameter identifiability in models fitted with aridec data allowed us to assess which model structures are the most appropriate to use in this context.

This identifiability analysis is based on the calculation of the collinearity index (Brun et al., 2001). This index is a measure of the degree to which changes in one parameter are compensated by changes in other parameters for a certain model structure and dataset. We used the modCost function from the FME R package to first adjust a model cost function (Soetaert and Petzoldt, 2010). This function estimates weighted residuals of the model output versus the observed data and calculates sums of squared residuals, according to the formula: 3resk,l=Modk,l-Obsk,lerrork,l⋅nl, where Modk,l and Obsk,l are the modeled and observed values for any data point, k, of a variable l, respectively. Errork,l is a weighing factor that makes the term non-dimensional.

The model cost function, together with a set of pre-set initial parameter values, is then used as an input to calculate a matrix of sensitivity functions using the sensFun function from FME. This function estimates the sensitivity of the model output to the parameter values using the expression: 4Sij=∂ri∂Θj⋅wΘjwri, where Si,j represents each entry of the matrix, ri are model residuals calculated from the cost function, Θj is a model parameter, wriis the scaling of ri, and wΘj is the scaling of parameter Θj (Soetaert and Petzoldt, 2010).

The final step in this analysis is calculating the collinearity index γ. The collin function from FME uses the sensitivity matrix as an input to calculate γ for every combination of parameters; γ is defined as 5γ=1min⁡EVS^TS^. Here S^ij is calculated as: 6S^ij=SijΣjSij2, where S^ij contains the columns of the sensitivity matrix that correspond to the parameters included in the set, and EV estimates the eigenvalues. The collinearity index equals 1 if the columns are orthogonal, and the set is identifiable. The index equals infinity if columns in the sensitivity matrix are linearly dependent (Soetaert and Petzoldt, 2010). The interpretation of the collinearity index is thus a change in the residuals caused by a change in one of the parameters can be compensated by a proportional change 1/γ in another parameter. For practical purposes, if γ>20, the parameter combination is considered non-identifiable (Sierra et al., 2015).

For the identifiability analysis, we first selected a representative group of 30 entries from the database ranging from three to 19 time points (Table 2). The number of data points in time limits the number of parameters that can be fitted because it affects the number of degrees of freedom. Thus, models with more parameters require longer datasets. This meant that, a priori, not all entries could be used to fit all model structures. On the other hand, it is possible to test identifiability for restricted model versions, that is, models with some of their parameters fixed to a known value. This implies that there are fewer parameters to be determined, and thus it allows us to use shorter time series. The details of all the models tested are reported in Table 3. From this first analysis, we noticed that two pool parallel and series structure models with a restricted initial proportion of litter in pool 1 (p1) were the two models more likely to meet identifiability with our data. Because of this, we tested collinearity for all the 184 entries in the database, but only for these two models and for the respective models with the full set of parameters, for comparison. The R code for this analysis can be found in the collinearity.R script inside the scripts folder of aridec.

Our collinearity analysis (see below) showed that although most entries could be used to fit two-pool parallel and series models with a fixed p1 parameter (i.e., the initial proportion of litter in pool 1), this was dependent on each dataset. Restricting the p1 parameter is a sensible way of achieving identifiability because it is common to find information on litter lignin content in decomposition publications, and this can be used as an initial proportion value for the slow-decomposing litter pool (i.e., p2). Since 7p1+p2=1, it is possible to estimate the p1 as the complementary value of p2.

To give a practical example of what can be done with this database, we chose to fit these models using one of the entries where both the models were identifiable, and the initial proportion of litter lignin was available. Together with these models, we fit a simple one-pool model for comparison. We used variable 2 (V2) from the Day2018 entry, which corresponds to Simmondsia chinensis (Link) C. K. Schneid. leaf litter decomposed under full sunlight treatment in the field (Day et al., 2018). The initial proportion of lignin (p2) was 0.09. We used the Bias Corrected Akaike Information Criteria (AICc) to assess the model fit (Shumway and Stoffer, 2017).

The 184 studies in the database included data for 212 unique study sites around the world. Twenty-four of these sites were repeated in two or more studies. According to the GAI, ∼3.6 % of sites were classified as hyper-arid (0–0.05), ∼35.9 % as arid (0.05–0.2), ∼43 % as semi-arid (0.2–0.5), ∼6 % as dry sub-humid (0.5–0.65), and ∼12 % as humid (>0.65). We recognize that humid sites do not classify as aridlands, but we included them nonetheless because these sites had marked dry seasons according to the original publications. A total of 33 countries was represented in our database. The top-five countries with the largest number of study sites were China (58 sites), USA (49 sites), Argentina (32 sites), Israel (22 sites), and Brazil (12 sites). Most sites in the database correspond to arid regions where the mean annual temperatures are above zero degrees Celsius, with a very low representativity of colder regions (Fig. 3)

Out of the 184 database entries, we retrieved 1752 series of litter mass loss over time. The oldest publication in the database is from 1975, and the newest is from 2021. Moreover, there has been a considerable growth in the number of publications per year (Fig. 4a). The study duration in the database ranged from 18 d to 10 years, with a median of 365 and a mean of 430 d (Fig. 4b). The number of sample harvests from the field went from 2 to 23, with a mean of ∼6 and a median of 5 (Fig. 4c). The sampling frequency ranged from 0.08 to 11.1 samples per month, with a median of 0.4 and a mean of 0.8 samples per month (Fig. 3d). Elevation at the study sites varied from -375 to 4000 m a.s.l., with median and mean values of 557 and ∼811 m a.s.l., respectively (Fig. 4e). The mean annual temperatures ranged from -0.45 to 29.5 ∘C at the study sites with a mean value of 14.9 ∘C and median of 15.6 ∘C (Fig. 4f). The mean annual precipitation in aridec ranged from 2 to 1700 mm, with median and mean values of ∼375 and ∼494 mm, respectively (Fig. 4g). Out of all sites, 23 % were reported by the authors to be deserts, 17 % forests, 16 % agroecosystems, 12 % grasslands, 10 % shrublands, 10 % steppes, 8 % savannas, 2 % coastal ecosystems, and 2 % urban sites (Fig. 4h).

Figure 5 shows results from the first identifiability analysis carried out on a subset of 30 entries. Here, we can compare how entries with an equal number of time points behave under each model structure. For the two-pool parallel model structure, four parameter combinations were compared: a full three-parameter model and three alternative models with one restricted parameter each. There were 30 points for the full model with 62.1 % of values below the collinearity index γ=20 threshold, i.e., almost 40 % of the models were not identifiable. For the models restricted to two parameters, 95.6 % out of the 90 models compared were identifiable. When specifically looking at the model with a restricted p1 value, 100 % of the models were identifiable. This was expected because usually the fewer the parameters to be estimated, the lower the collinearity in models.

The two-pool series model structure analysis included a full four-parameter model and ten alternative models with one or two restricted parameters each. From a total of 25 full parameter models, 48 % were identifiable according to their γ values. Out of 150 models with two restricted parameters, 94.7 % of models were identifiable, while for models with only one restricted parameter, 68.3 % of models were identifiable. When specifically checking for the collinearity index in models with a restricted p1 parameter, 84.6 % of them were identifiable. The non-identifiable values in this last case corresponded to four entries with four time points each. That means that 100 % of models with >5 time points were identifiable for the restricted p1 model version. This highlights the importance of having longer time series available for modeling.

The case of the two-pool model structure with feedback included a full parameter model and 24 other model variants with one, two, and three restricted parameters each. The analysis of 15 models with all five parameters showed 100 % of non-identifiable results. The results for the restricted model version with four parameters showed 100 % of non-identifiable models out of 75 data points, while only 4.7 % of the 150 data points with three parameters gave γ values lower than 20. The analysis of the restricted model version with two estimated parameters generated 57.3 % of identifiable results out of 150 models.

When testing for the three-pool parallel model structure, we used one full model with five parameters and 24 model variations comprising from two to four parameters each. None of the full model data points showed collinearity index values lower than 20. Out of the restricted four-parameter models only 34.7 % could be identifiable. When we specifically looked at the models where either p1 or p2 were fixed, none of them were identifiable. Restricted models with three parameters produced a 68 % of identifiable results. Further, restricted models that only had k values were not identifiable. Finally, 89.3 % of models restricted to two parameters were identifiable according to our analysis.

The last model structure analyzed was the three-pool series structure, which produced comparisons for models with all seven parameters, plus 118 other model variants with different restricted parameters. None of the models with all seven parameters were identifiable. Only 5.7 % of the models with six parameters were identifiable, none of which corresponded to the models where either p1 or p2 were fixed. Models restricted to five parameters produced 10.5 % of identifiable results. Specifically looking at models with both fixed p1 and p2, 100 % of those were not identifiable. Restricting models to four parameters generated 28 % of results with γ>20. Models with three estimated parameters produced 54.9 % of identifiable results. Lastly, 84.8 % of models restricted to only two parameters were identifiable.

Because two-pool parallel and series models with a fixed p1 parameter showed the highest percentage of identifiable cases in this first analysis, we did a second test with the whole database for these models and for their respective full-parameter versions for comparison (Fig. 6). For the two-pool parallel model with the full set of parameters, 58.7 % of entries were identifiable, whereas restricting the p1 parameter yielded 99.5 % of identifiable entries. For the more complex series models, the percentages of identifiable entries were much lower, with only 20.1 % for the restricted version and no identifiable cases for the version with a full set of parameters. Clearly, restricting the number of parameters to be estimated decreases collinearity, but the results are highly variable and dependent on each particular dataset.

The previous analysis of the collinearity index (Fig. 6) shows that from the proposed model structures to be fitted, our data can, in most cases, be fitted to two pool parallel and series structure models with a restricted initial proportion of litter in pool 1, aside from single-pool models. Figure 7 shows the results from the simulation of the dynamics of organic matter loss from leaf litter fitted from the Day2018 entry. The one-pool model showed how a single reservoir of organic matter from leaf litter decomposed at a k rate of 0.0142 per month (Fig. 7a). At the end of the almost 3-year period, the remaining percentage of total organic matter was 61.02 %. The two-pool parallel model showed a fast decomposing organic matter pool with a k1 of 0.0158 per month (Fig. 7b). This pool went from representing 91 % of total organic matter to 52.4 % after almost 3 years. On the other hand, the slow decomposing pool, which we defined as the initial lignin content of litter, had a value of k2 of 1.5×10-16 per month. We defined this pool as 9 %, and it remained unchanged after 2 years, which was expected from a k2 of nearly zero. Lastly, the two-pool series model showed a fast decomposing pool with a k1 of 0.11 per month (Fig. 7c). This pool went from 91 % of total organic matter to 1.7 % at the end of the experiment. In this case, the slow decomposing pool had a k2 of 0.02 per month. This model also had a transfer coefficient from the fast decomposing pool to the slow pool of 1 (i.e., 100 % of organic matter in the fast pool that decomposed in a month transformed into more recalcitrant forms, adding to the slow decomposing pool). Then, the slow decomposing pool went from 9 % at the start of the simulation to 54.08 % after almost 3 years. Judging by their AICc values, the three models were similarly supported by the data.

The aridec database is a comprehensive database with a wide range of information on decomposition studies from aridlands worldwide, which includes litter mass loss data, litter traits, and experimental design information. Our exhaustive bibliographic search gave us close to 200 papers that fulfilled our criteria. Notably, we did not limit our work to studies published in English; aridec also includes papers in Spanish, Portuguese, and Mandarin. This widens the scope of our work to achieve a more inclusive database.

From a geographic perspective, study sites included in aridec cover most of the main aridlands of the world (Fig. 2). As expected, countries like China, USA, and Argentina had the largest number of studies, which might be related to the extension of aridlands in these countries, since China has 6.07×106 km2 of drylands (Huang et al., 2019), and around 40 % and 69 % of USA and Argentina territories are considered as drylands, respectively (Verbist et al., 2010; White and Nackoney, 2003). In contrast, some of the biggest deserts in the world, such as the Sahara, the Kalahari, the Australian Outback, and the Arabian desert are underrepresented, if not absent, in our database. Future efforts should focus on this information void, and the aridec database will be available to include these coming studies in our framework.

Study sites in the database represent a big part of the climate range where aridlands occur, from hyper arid deserts to dry sub-humid ecosystems (Fig. 2). Some of the study sites (∼12 %) were classified as humid according to the GAI. We chose to include them because those studies reported marked dry seasons at the study sites and focused on seasonal patterns of litter decomposition. The range of climatic variables such as mean annual temperature and precipitation, physical variables like altitude, along with land cover types are also very well represented in this database (Fig. 4). This wide representativity of climates in aridec is a crucial asset if the database is to be used to answer global-scale questions. Nonetheless, dry sub-humid lands but mainly hyper-arid deserts are the least represented in the database, suggesting more studies should be developed in these areas. Moreover, there is a void in the colder end of the climatic range of aridlands (Fig. 3). This might be related to the lack of sites in aridec of ecosystems like tundra where the GAI is mostly low, but a bibliographical search like ours could not retrieve the studies in those areas. There is clearly much room to expand our database and increase its potential applications.

One of the advantages of the aridec database is that its files are compatible with R and, specifically, with the SoilR library (Sierra et al., 2012). In addition to the fact that this database is an open-source and open-code project, there is huge potential for broadening the extent of the information in aridec and for developing code to work with it. Nonetheless, we should mention that the formats included in the database are not R-exclusive and can be used with most commonly available software. YAML files can be read and edited from any text processor, and CSV files can also be opened with any spreadsheet software. Although statistical analysis on R scripts cannot be used elsewhere, raw data itself can be freely processed with any statistical software.

The aridec database can be used on its own, but we recommend complementing our information with other publicly available databases to expand the application possibilities. Metadata are not always fully reported in publications, so it is possible to fill these gaps with climate and altitude data from databases like NASA Prediction of Worldwide Energy Resources (POWER; NASA Langley Research Center (LaRC)). Leaf litter traits are a big part of the database and are sometimes poorly reported in publications. A general source of litter traits data can be obtained upon request from the TRY database (Kattge et al., 2020). Another more specific source of information is NODdb, where nodulating-N-fixing plant genera are detailed (Tedersoo et al., 2018).

Based on our collinearity analyses (Figs. 5 and 6), we suggest that before fitting any models to our data, it is crucial to check the identifiability of the parameters with each database entry. We assume that all entries can be fitted to a one-pool model, since there is only one parameter to estimate. As for the more complex models, the situation is highly dependent on which data entry is being used. For instance, although most entries could be used to fit two-pool models with parallel and series structures, there are some exceptions. Another example can be seen in Fig. 5b, where one dataset of more than eight points was not identifiable for a model with all four parameters. Besides collinearity, the number of degrees of freedom will restrict which models can be fitted to the data so that these two aspects should be considered together. As models get more complex, we had to progressively exclude entries with less time points because they did not have enough degrees of freedom. As a concluding remark, the results in Figs. 5 and 6 are not to be interpreted as how those models perform in general but how they perform with the specific data in aridec. That is why we provide an R script in the database to test collinearity for individual datasets.

Moreover, in most cases, fitting these models is only possible by restricting the parameters estimated to only decomposition constants and transfer coefficients. We suggest restricting models to these parameters because it is more likely to find data on initial proportion of lignin or cellulose to be used as proxies for parameters p1 or p2. Not only might they be more difficult to find in the literature but estimating values for decomposition constants and transfer coefficients might altogether be a better use of this database. Further, we recognize that for some specific combinations of parameters and datasets, the more complex models might be identifiable (data points below the log10(γ=20); Fig. 5).

Again, interconnection between datasets like aridec and others like TRY (Kattge et al., 2020) is a key workaround to the collinearity problem by providing data for parameter restriction (Sierra et al., 2015). We recognize that limitations in data available from field studies ultimately restrict our capability to fit more complex models (Brun et al., 2001). This limitation can have further implications if we consider the proportion of identifiable datasets per ecosystem type or level of aridity. For example, semi-arid and dry sub-humid ecosystems show the lowest proportion of identifiable datasets for a two-pool parallel model with all parameters (data not shown). This would lead to an under-representation of some aridlands because of a lack of suitable data available. Moving forward, new decomposition studies should consider making more measurements and including data on litter initial chemical quality, as well as expanding studies to less represented climates and ecosystems. This will allow for the detection and modeling of finer scaled dynamics of organic matter (see Appendix A).

The applied example in Fig. 7 serves as a glimpse of the potential of this database. Choosing which model to fit with decomposition data is not a minor task. For instance, none of the three models were supported as the best model from their AICc, which suggests that they must all be considered when making conclusions from this simulation (Anderson and Burnham, 2004). Whether a time series has a better fit with one model or another depends mostly on mass loss dynamics that were captured by the experimental design. While longer datasets can generally be fitted with more complex models, if the overall mass loss dynamics fit better for a one-pool model, higher model complexity will not be chosen using AIC. It is also to be considered that ecological theory may come into play here instead of applying purely statistical reasoning. As was pointed out recently, litter decomposition is not as much a process of “what is lost” but more of “what is left” (Prescott and Vesterdal, 2021). A balance between statistical fit and theoretical support should be found when choosing which model is best for each study case.

The aridec database is available for open access and download at http://github.com/AgustinSarquis/aridec (last access: 31 May 2022; Sarquis et al., 2022). Our hope is that newer studies on dryland litter decomposition will be added to the database by new collaborators (see Appendix A). It is important to follow our user guidelines to ensure consistency, all of which are available in the database itself. File templates for uploading new entries to the database are given, and further details can be found in them. Additionally, users will find in the database a README file, scripts to test file consistency and many examples on how to apply functions and to fit models using R code.

To our knowledge, the aridec database is unique. Other databases that focus on land C studies include the Soil Incubation Database (SIDb; Schädel et al., 2020), a peatland productivity and decomposition parameter database compiled by Natural Resources Canada (Bona et al., 2018), and the Chilean Soil Organic Carbon Database (CHLSOC; Pfeiffer et al., 2020). Although they all intend to assess questions related to C budgets in terrestrial ecosystems to some extent, not all of them present decomposition data (i.e., Pfeiffer et al., 2020). Moreover, only SIDb (Schädel et al., 2020) and aridec contain time series of organic matter loss. This is a unique asset that allows for future studies to make new assessments of decomposition without having to worry about inconsistencies in the calculation of k parameters. Finally, none of these other databases are centered around plant litter decomposition in aridlands like aridec.

The extent of the information included in aridec in addition to its open-science approach makes it a great platform for future collaborative efforts in the field of aridland biogeochemistry. In this sense, the main purpose of this database is to further our understanding of C dynamics at the earth system level. Complete datasets like aridec are necessary to test which model structures and parameters best explain decomposition processes and to help develop more realistic representations of the global C cycle in drylands (Luo et al., 2016). Further, additional parameters could be used to test the importance of mechanisms that are relevant in aridlands but are under-represented in the literature. Studies on processes like photodegradation (Adair et al., 2017) could be expanded to a wider geographical range and to soil processes thanks to the representation of sites in aridec using the SoilR framework (Sierra et al., 2012). Another potential application of our database is to combine ecological data with climatic data in earth system models, which is a promising framework to assess future global change stresses and their effects on the biosphere (Bonan and Doney, 2018).