Novel frontier in wildlife monitoring: Identification of small rodent species from fecal pellets using near‐infrared reflectance spectroscopy (NIRS)

Abstract Small rodents are prevalent and functionally important across the world's biomes, making their monitoring salient for ecosystem management, conservation, forestry, and agriculture. There is a growing need for cost‐effective and noninvasive methods for large‐scale, intensive sampling. Fecal pellet counts readily provide relative abundance indices, and given suitable analytical methods, feces could also allow for the determination of multiple ecological and physiological variables, including community composition. In this context, we developed calibration models for rodent taxonomic determination using fecal near‐infrared reflectance spectroscopy (fNIRS). Our results demonstrate fNIRS as an accurate and robust method for predicting genus and species identity of five coexisting subarctic microtine rodent species. We show that sample exposure to weathering increases the method's accuracy, indicating its suitability for samples collected from the field. Diet was not a major determinant of species prediction accuracy in our samples, as diet exhibited large variation and overlap between species. fNIRS could also be applied across regions, as calibration models including samples from two regions provided a good prediction accuracy for both regions. We show fNIRS as a fast and cost‐efficient high‐throughput method for rodent taxonomic determination, with the potential for cross‐regional calibrations and the use on field‐collected samples. Importantly, appeal lies in the versatility of fNIRS. In addition to rodent population censuses, fNIRS can provide information on demography, fecal nutrients, stress hormones, and even disease. Given the development of such calibration models, fNIRS analytics could complement novel genetic methods and greatly support ecosystem‐ and interaction‐based approaches to monitoring.


| INTRODUC TI ON
Small rodents are prevalent in ecosystems across the globe, with many species acting as ecosystem engineers (Dickman, 1999) or keystone species (Ims & Fuglei, 2005); with numerous rare species of high conservation value (Green et al., 2013) and many invasive populations with profound effects on ecosystem functioning (Dickman, 1999).
Monitoring and research of small rodents is thus a globally salient enterprise for ecosystem management, conservation, forestry, and agriculture. Yet, obtaining population-or community-level data on small rodents is often challenging, as these small and cryptic animals are elusive (Green et al., 2013) and cost-effective methods for largescale sampling of their multispecies communities are largely missing (Engeman & Whisson, 2006;Heisler et al., 2016). Estimation of small rodent community composition, population size, and density relies on a variety of trapping, sampling, and indexing efforts, for example, pitfalls, live-, snap-, hair-, or camera traps and systematic incidental observations (Engeman & Whisson, 2006;Fauteux et al., 2018;Soininen, Jensvoll, et al., 2015), or on counts of burrows, runways, winter nests, owl pellet contents, and feces (Green et al., 2013;Heisler et al., 2016). Trapping methods, while providing actual estimates on species identity, abundance, and key demographic parameters, are laborsome (Engeman & Whisson, 2006;Villette et al., 2016) and often inadequate for high-intensity sampling with large spatial or temporal coverage (Heisler et al., 2016;Villette et al., 2016).
Multitude of wildlife censuses rely on feces counts (Campbell et al., 2004;Green et al., 2013;Karels et al., 2004;Kohn & Wayne, 1997), as feces are a noninvasive source of readily available samples. Acquiring feces does not require being in contact with the animal, thus decreasing the risk of infections and animal stress, and feces can provide information from seasons and times when researchers are not present (Kohn & Wayne, 1997;Whisson et al., 2005). Feces are organic material whose structural and chemical properties are determined by the mechanical, biochemical, and microbiological processing of ingested biomass throughout the digestive pathway. Feces-and the chemical and genetic information they store-could therefore allow for the determination of diverse ecological and physiological variables indicative of species interactions (c.f. Ehrich et al., 2019;Vance et al., 2016), if combined with suitable molecular, endocrinological, or spectral methods (Galan et al., 2012;Schwarzenberger, 2007;Zemanova, 2021). Indeed, realizing the potential of fecal data for large-scale ecological monitoring calls for a set of cost-effective high-throughput analytical methods (Vance et al., 2016;Zemanova, 2021). Such methods, if embedded in appropriate monitoring schemes (Yoccoz et al., 2001), could provide access to questions pertinent to ongoing biodiversity shifts and ecosystem changes (Lenoir & Svenning, 2015;Pecl et al., 2017;Wintle et al., 2010). To be feasible and applicable, these methods should be robust and sensitive across space and time, utilize easily obtainable samples, be quick in the preprocessing of the samples, and be cheap and fast (Engeman & Whisson, 2006;Whisson et al., 2005).
Analytical methods that are nondestructive for the fecal samples themselves increase their utility by allowing subsequent application of further analytical methods. For instance, rapid advances in noninvasive and nondestructive genetic methods allow for increasingly cost-effective taxonomic analysis of feces (Zemanova, 2021), given sufficiently preserved genetic material.
Alongside genetic methods, near-infrared reflectance spectroscopy (NIRS) is a promising and highly versatile method for diverse ecological monitoring (Vance et al., 2016). NIRS is widely applied in agriculture and petrochemical industry (Pasquini, 2003), with increasing representation in different fields of ecology (Aw & Ballard, 2019;Murguzur et al., 2019;Vance et al., 2016;Villamuelas et al., 2017). NIRS is a rapid and nondestructive spectral analytic tool for assessing quantitative and qualitative variables based on the physicochemical information stored in the NIR spectra, including, for example, taxonomic or demographic information and fecal chemistry, or even stress or disease (Vance et al., 2016). After initial investment, the method is very affordable. Sample preprocessing is minimal, mainly involving sample material homogenization and drying (Pasquini, 2003). The scanning procedure of a sample takes only seconds and does not require specialized personnel or laboratories. Once a calibration model for the variable of interest (e.g., species identity) is validated, scanning and analysis of any number of further samples comes with no extra cost of, for example, reagents or genetic primers. After scanning, samples can be analyzed with other methods, and the existing spectra can be later analyzed for new variables using respective calibrations. Fecal NIRS (fNIRS) has been used to predict species identity, demographic parameters (Aw & Ballard, 2019;Tolleson et al., 2005;Wiedower et al., 2012) and diet quality (Foley et al., 1998;Villamuelas et al., 2017) of several wild animals from large mammals to insects (Vance et al., 2016). However, the use of fNIRS in wildlife research and monitoring remains rather unknown to the ecologist community, in comparison with the widespread use of remote sensing (Kerr & Ostrovsky, 2003;Pettorelli et al., 2014) and high-throughput genetic barcoding (Yoccoz, 2012).
As with any analytical method, there are potential caveats to using fNIRS for small rodent monitoring. First, as feces are collected from the field, they are exposed to ambient weather. This can compromise prediction accuracy because leaching and microbiological processes can change pellet chemical composition (Jenks et al., 1990;Kamler et al., 2003), similar to issues linked with degrading DNA. Second, variation between animal individuals may affect their fecal composition and NIR spectra and hence decrease or confound prediction accuracy of species. In particular, diet quality affects fNIR spectra (Stuth et al., 2003;Villamuelas et al., 2017)

T A X O N O M Y C L A S S I F I C A T I O N
Biodiversity ecology, Biogeography, Community ecology, Conservation ecology, Invasion ecology, Movement ecology, Population ecology, Zoology and may compromise prediction accuracy whenever considerable dietary overlap between species occurs. Especially for coexisting and competing generalist rodent species, the effect of diet is likely to be complex, as diets overlap and are dependent on, for example, season and forage item availability (Soininen et al., 2013). In addition, sex and reproductive status may affect fNIR spectra (Tolleson et al., 2005), increasing variation in the spectral data. Third, if factors driving the species-signal in fNIR spectra differ between populations, fNIRS calibrations that cover only spatially limited target populations can yield unreliable results on samples from a different population; an issue avoided by genetic methods of well-preserved samples. However, recent studies presented NIRS calibrations applicable across species and geographic regions (Murguzur et al., 2019;Villamuelas et al., 2017), revealing the potential for cross-regional or global NIRS and fNIRS calibrations.
Here, we develop fNIRS calibration models for rodent taxonomic determination from single pellets (>0.025 g of dry weight). Our main hypothesis is that fecal properties differ between taxa, allowing for classification of individuals to genus and species based on their fNIR spectra (H1 and H2). However, this separation capability might erode with the effect of exposure (H1 vs. H2) due to leaching, irradiation, and decomposition. Prediction accuracy of individual samples may also be linked with diet composition (H3), which we tested by comparing fNIRS data to DNA metabarcoding data of the same pellets.
Additionally, spectra may display regional differences between populations (H4). For the latter, we contrast two subhypotheses: H4.1 and H4.2 that calibration models based on samples from one region may perform poorly with samples from other region and H4.3 that a calibration model including all regions successfully classifies independent test individuals from all regions. Finally, we demonstrate how our calibration model and the modeling framework would be used in practice to predict species identity of new fecal samples.

| MATERIAL S AND ME THODS
We outline the workflow of building NIRS calibrations and testing our hypothesis in Figure 1. We detail the following steps: rodent fecal sample extraction, sample processing and experimental treatment, NIRS scanning and spectra pretreatment, calibration modeling including model building, validation and testing, diet molecular analysis including DNA metabarcoding, bioinformatics and modeling potential confounding of diet, as well as regression modeling of NIRS calibration model results. Further details on molecular methods and calibration modeling are provided in the Supplementary material S1

| Rodent data and fecal samples
We collected fecal samples from animals trapped as part of a longterm rodent population monitoring time series. The set of rodent individuals included in the calibration modeling incorporated substantial variation in environmental conditions of their trapping site, season, and year, as well as variation in individual physiological condition related to reproductive status and age ( Table S1a) Table S1a).
Another set of fecal samples (n = 85) came from animals trapped in three river catchments in East Finnmark-Ifjordfjellet (70° N, 27° E), F I G U R E 1 Illustration of the analytical workflow. Detailed description of calibration models is presented in Table 1.  Table S1a).
All rodents were frozen after trapping, and information of trapping date and sampling quadrant was recorded. We also recorded individual rodent body mass, sex, species identity, and female reproductive status (visible pregnancy). Dataset included individuals of varying age based on the wide body mass distribution (L. lemmus Feces were carefully sampled from the intestine without damaging or breaking the intestinal tissue, avoiding contamination with blood or other compounds that could affect the NIR spectra. From each individual from WF, we collected consecutive pellets in two Eppendorf tubes, to form two parallel samples, while of EF samples no parallel samples were taken. All pellets were dried at 40°C for minimum of 48 h or until dry and stored in Eppendorf tubes in room temperature. One parallel WF sample and all EF samples were NIRscanned without exposure treatment (intestinal samples), while the other WF parallel sample was scanned after the exposure treatment (exposed samples).

| Exposure treatment
We subjected one set of parallel WF samples to ambient weather conditions (hereafter "exposure treatment") in Tromsø, Northern Norway, during autumn 2014 to incorporate the effects of leaching and bleaching due to UV radiation on NIR spectra in the calibration model. Only four species were included in the weathering treatment, as Myodes rutilus pellets were not suited for weathering due to a frequently liquid consistency and small size of the pellets. We placed the pellets on wooden frames of 50x50 cm, with a 1 x 1 mm nylon mesh in the bottom. Nylon mesh was used as a bottom material because it does not release chemicals to the feces and allows for water drainage and evaporation. Each frame had pellets of only one species to avoid contamination. Each frame had ca. 150 pellets.
For L. lemmus and Myodes rufocanus pellets, we used two frames per species, whereas Microtus oeconomus and M. agrestis had one frame each. Frames were placed on heath vegetation in September, and weather conditions during the treatment were variable, including rain, sunshine, frost, and temperatures both above and below zero. During 6 weeks, we weekly took ca. 25 pellets for scanning from each frame.

| NIRS scanning of the fecal pellets
We scanned individual intestinal and exposed samples without further sample prepreparation. Due to the small size, we used a custombuilt, Ø4mm sample holder and pressed the pellet flat with an Ø4mm metal rod in order to prevent light from penetrating through the sample. A few samples were too small to cover the whole area, but they did not appear as outliers during the modeling process NIR spectra were collected using a FieldSpec 3 (ASD Inc., Boulder, Colorado, USA): each spectrum was recorded as reflectance using monochromatic radiation at 1.4-nm intervals in the TA B L E 1 Summary of calibration model methodology. Column « Hypothesis » refers to the tested question. 350-1050 nm range and 2 nm intervals in the 1050-2500 nm range, and interpolated to 1 nm resolution. We scanned each sample three times by rotating the sample holder between scans in order to account for angular effects on light scattering. We used the mean of the three spectra in the further analysis. We used the R package prospectr (Stevens & Ramirez-Lopez, 2015) for spectra preprocessing, that is, applying splice correction and normalizing the spectra.
The spectral regions in the 350-399 nm and 2451-2500 nm ranges were removed from the dataset due to instrumental noise

| DNA-metabarcoding of rodent feces
After scanning, we analyzed a subset of intestinal WF samples (n = 385) for diet species composition with DNA metabarcoding.
All wet laboratory work was performed by Center of Evolutionary Applications, University of Turku, Finland. We used a DNA metabarcoding approach similar to Soininen, Gauthier, et al. (2015). We here give a summary of the methods, and a detailed description is included in supplementary materials S1 (p. 2: Detailed DNA-metabarcoring of rodent feces). DNA was first extracted and thereafter amplified using universal plant primers "g-h" and "c-h" that target chloroplast trnL intron (Taberlet et al., 1991(Taberlet et al., , 2007. Final DNA libraries were constructed in the second PCR (see, e.g., Vesterinen et al., 2018). Resulting DNA libraries were purified and size-selected using an SPRI bead clean-up (see Vesterinen et al., 2016), and concentrations were measured by Qubit (Invitrogen; www.invit rogen.com) and finally pooled in equimolar quantities. Sequencing was done on 316 v2 and 318 v2 chips with Ion PGM (Life Technologies, manual cat nr 00009816, Rev C.0).
The four sequencing runs yielded altogether 12,575,869 qualitycontrolled reads that could be assigned to original samples. The reads were uploaded to CSC servers (IT Center for Science, www. csc.fi) for trimming and further analysis. Bioinformatic steps were carried out following Schmidt et al. (2018) with several modifications (p. 3-4: Bioinformatics). To summarize, reads were merged, filtered for quality, cut for primers, collapsed into unique haplotypes, and finally clustered into zero-radius OTUs using softwares usearch (Edgar, 2010) and cutadapt (Martin, 2011). We were able to map 7,726,911 reads (~93% of the trimmed reads) to our original samples. The trnl OTUs were initially identified to taxa using the usearch "sintax" classifier with a database consisting of artic plant trnl sequences (Willerslev et al., 2014) using 70% probability threshold for taxonomic assignation. After this, read counts for each diet taxon in each sample were transformed to relative read abundances (RRA), as read abundances may actually be less misleading than presence/ absence conversions (Deagle et al., 2018).

| Calibration modeling of genus and species identity
After obtaining NIR spectra and diet data on individual samples, we built calibration models for genus and species identification and further used regression, multivariate, and latent variable modeling in order to address our four hypotheses and test of concept ( Figure 1, Table 1). To build calibration models, we applied multivariate adaptive regression splines (MARS; Friedman, 1991) both directly and in a flexible discriminant analysis framework (FDA; Hastie et al., 1994; see Table 1). For this, we used R packages earth (Milborrow, 2014) and mda (Hastie et al., 2015). We chose this approach as it deals well with high-dimensional input data and allows for additive effects and interactions between variables (Friedman & Roosen, 1995) and assumes nonlinear responses or decision boundaries (Hastie et al., 1994). A more detailed description of the modeling and a comparison with other common chemometric approaches is provided in Supplementary material S1 (p. 4, Calibration modeling of genus and species identity). We used R version 3. reported as model misclassification rate and species prediction accuracy; Pasquini, 2003). The purpose of applying rdMCCV was to estimate how model performance and misclassification rates vary across sets of calibration data (Filzmoser et al., 2009) and to estimate error rates for individual samples (cf. Liu et al., 2008). Furthermore, we repeated the rdMCCV procedure with 95%/5%, 90%/10%, and 80%/20% data splits between calibration and test sets. This increased the number of total rdMCCV iterations per model to 600 (Table 1). To ensure the allocation of targeted variability between calibration and test sets, the data were split (depending on the model) as fixed percentage of each species; species and exposure treatment duration; or species, exposure treatment duration, and region ( with "0" denoting a misclassification, and "1" a correct classification.
We then averaged classification result across all iterated model runs.
Prior to analysis, we transformed the IPA values as to avoid zero-one inflated data (Smithson & Verkuilen, 2006). As predictor variable, we used exposure treatment time (number of weeks from 0 to 6 as a numerical fixed factor). Due to the beta-distribution (0 < y < 1) of IPA values, we fitted beta-regressions with a logit-link using package rstanarm and with the function's default weakly informative priors (Muth et al., 2018). We fitted the model using Markov and visually with trace plots, and we inspected and confirmed good model fits by visual posterior predictive checking (Muth et al., 2018) with package bayesplot (Gabry & Mahr, 2019).

BOX 1 Practical application of the calibration model
Step 1: Building the calibration model.
We used 95% of exposed WF samples (n = 798, modeling dataset) to fit an FDA/MARS model with a two-way interaction. In total, 66 wavelengths were selected into the model (Figure 9 lower panel, Supplementary material S1, Figure S3).
We used the calibration model to predict the remaining 5% of WF {n = 35), and all EF samples to species.  Figure 3 lower right panel).
The misclassification rate of EF samples was 0.875, with prediction accuracies of 13.2% for M. rufocanus and 11.2% for M. oeconomus.
Step 3: Visualizing the calibration model and sample misclassification.
First, we plotted the canonical discriminant variables (CA1-3) of the modeling dataset ( Figure 7) and the WF and EF test sets (Figure 8). Second, to illustrate areas of high and low sample misclassification rates in the discriminant space, we applied the rdMCCV procedure on the modeling dataset and retrieved sample IPA values to color the plotted samples (Figures 7 and 8). Misclassification rate of the modeling data based on the rdMCCV calibration models was 0.056 ± 0.028. Across the 600 rdMCCV models, in total 1961 variables had importance greater than zero in one or more models (Figure 9 upper panel, Figure S3).

| H1: Prediction accuracy of genus and species identity
Calibration models for genus identity resulted in excellent prediction accuracy and a low misclassification rate of 0.041 ± 0.019 (mean ± sd, Figure 2, upper right panel). Across all model iterations, prediction accuracy for Myodes was 96.2% ± 3.13%, for Microtus 95.1% ± 3.62%

F I G U R E 2
Predictive ability of genus-specific fNIRS calibrations for small rodents from Finnmark, Norway. Left panels: prediction accuracy of each genus (mean ± sd) separately. Right panels: density plots of model misclassification rates. Upper row shows results from calibration models used for H1 (i.e., including all data from West Finnmark), lower row for H2 (i.e., including only samples used to test exposure effect). Results are divided by data splits (fractions) with 200 iterations each, where 5%, 10%, and 20% of data were randomly assigned as validation data (using rdMCCV framework). and for Lemmus 96.7 ± 3.47% (Figure 2, upper left panel). For species, the prediction accuracy was more variable and with a misclassification rate of 0.129 ± 0.032 (Figure 3,  Microtus agrestis (75.1% ± 9.85%; Figure 3 upper left panel).

| H2: Effect of feces exposure
Calibration models based on only exposed samples performed better than calibration models including intestinal samples, especially in predicting species identity (Figure 3). Effect of feces exposure to ambient weather for 1-6 weeks was visible on fNIR spectra, with a clear decline in reflectance at 700-1400 nm and increase in reflectance between ca. 1500 and 2500 nm ( Figure S1). The rate of change in fNIR spectra seemed to decline with exposure duration for all species; however, the Microtus oeconomus and M. agrestis spectra changed strongly again after week 5 (Supplementary material S1, Figure S1). Model misclassification rate for genus identity was 0.031 ± 0.02 (Figure 2,  for Microtus agrestis (Figure 3, lower row).  Figure 4, Table S3).

| H3: Does diet confound prediction accuracy?
The fNIR spectra were only weakly correlated to diet, as indicated by both Mantel test (r = .153, p < .001) and Procrustes test (

| H4: Regional vs. cross-regional calibrations
Calibration models built with data from one region predicted samples from the same region well, but samples from the other region poorly (Figure 6; cf. Box 1 and Figure 8). WF calibration (H4.1) predicted WF samples with a misclassification rate of 0.041 ± 0.027, while misclassification rate for EF samples was as high as 0.349 ± 0.071 (Supplementary material S1, Figure S2).  Figure 6 upper row).

Mean prediction accuracy of
Similar region-specific pattern emerged with calibration models built on EF data (H4.2). Misclassification rate was lower for EF (0.171 ± 0.140) than for WF samples (0.265 ± 0.064, Figure S2). from EF test data reached decent prediction accuracies ( Figure 6 middle row).

Calibration model including both WF and EF samples (H4.3)
predicted test samples nearly as well or better than either of the regional models alone ( Figure 6). Misclassification rate for WF samples was 0.041 ± 0.027, and for EF samples, 0.104 ± 0.108 ( Figure S2). Mean prediction accuracy for M. rufocanus was 96.7% ± 3.25% for WF and 85.3% ± 18.5% for EF samples (Figure 6 lower row

| DISCUSS ION
Our results demonstrate fNIRS as an accurate method for predicting vole and lemming genus and species identity based on single fecal pellets (H1 and H2). While the calibration models predicted genus identity extremely well (at >95% accuracy), prediction accuracy of species identity varied more but was still good-to-excellent at 85%-98%. Exposure of samples to ambient weather resulted in the best calibration model performance, contradicting the hypothesized negative effect of collecting samples from the field (H1 vs. H2). Surprisingly, we did not find support for diet being a major determinant of species prediction accuracy (H3). Calibration

F I G U R E 5
Comparing variation in diet composition with individual sample misclassification rate. Each individual (sample) is plotted as a separate dot, and all five species are plotted in their separate subplots. Spreads of species-specific diets are shown by separate ellipses. Point color darkness indicates lower IPA value, that is, higher sample misclassification rate based on rdMCCV calibration models on plotted samples. Point locations are latent variables of relative read abundances. Latent variable modeling is equivalent to unconstrained ordination, that is, proximity of samples in two-dimensional space indicates diet similarity, and distance indicates dissimilarity. Our results lend strong support for using fNIRS on samples exposed to ambient conditions, that is, samples collected from the field, and for further development of fNIRS for large-scale rodent population monitoring, including cross-regional calibrations. For future applications, the biggest appeal lies in the versatility of fNIRS.
In addition to small rodent censuses or community sampling, fNIRS calibrations can provide information on demography, fecal nutrients, stress hormones, and even disease, variables linked with fitness and not all captured by genetic methods-all from the same NIR spectra.
This makes fNIRS a compelling addition to ecosystem monitoring and sampling toolkits. If diet quality and composition are only loosely associated, the link between fNIR spectra, species identity, and generalist diet composition may not be easily detected. Even so, it is likely that group-specific variation in diet composition affecting any diet quality variable would still affect misclassification rates (e.g., Tolleson et al., 2005). Further research is needed to determine whether and which diet composition or quality variables co-vary with species identity to affect fNIRS rodent species classification.

| Diet and misclassification
Furthermore, the large sample size from WF may have incorporated sufficiently high levels of intraspecific dietary variation, to render the effect of diet on misclassification rate small, compared with other fecal constituents. With lower sample sizes, systematic variation in diet, for example, between years could confound species identification as found by Tolleson et al. (2005). Similarly, systematic variation in species diet between regions could explain high rates of misclassifications between WF and EF. Species-specific diets be-

| Juxtaposing misclassification with phylogenetic history
In general, classification model performance emerges likely as a complex product of individual traits linked with phylogenetic history. Evolutionary constraints may manifest in fNIR spectra through morphological, physiological, or microbiological differences in the digestive system, along with ecological niche factors (Blomberg et al., 2003). Dental and gut morphologies are examples of phylogenetically variable traits that may affect fNIRS species identification through differences in fecal particle size (Clauss et al., 2015;Sheine & Kay, 1977) and fecal nitrogen levels or fiber digestibility (Clauss et al., 2015;Lovegrove, 2010), respectively (Foley et al., 1998;Tolleson et al., 2005 and references therein; Steyaert et al., 2012).
Indeed, divergence of Lemmus, Myodes, and Microtus at tribal level (Buzan et al., 2008) is congruent with good performance of calibration models at genus level. Similarly, Myodes rufocanus and M. rutilus which were well separated in classification models show marked phylogenetic differentiation within the genus (Buzan et al., 2008;Cook et al., 2004;Kohli et al., 2014). While the two Myodes species' ecology can be similar in, for example, subarctic birch forests (cf. Ehrich et al., 2009), clear behavioral niche differences may exist (Nations & Olson, 2015) and for instance their habitat use at JRA differs markedly. By contrast, higher misclassification rates and similarities in Microtus fNIR spectra link with the relatively recent and rapid radiation history of the genus (Barbosa et al., 2018). While M. oeconomus is ancestral to M. agrestis, there is ongoing intraspecific divergence within both species (Barbosa et al., 2018), as well as strong interspecific competition (e.g., Hoset & Steen, 2007). Divergence patterns in dental structure converge roughly with these phylogenetic patterns, with Microtus oeconomus and M. agrestis displaying the smallest differences (Herrmann, 2002).

F I G U R E 8
Canonical discriminant plot of the calibration model, with predicted locations of WF and EF test samples plotted on top of the modeling dataset points (cf. Figure 7). Point shape indicates species identity. WF and EF samples are plotted with large points with black borders and with fill color indicating correct or incorrect calibration model prediction. Modeling data are plotted in all panels as small transparent gray points as in Figure 7.

| Identification of species and pellet exposure
Surprisingly, misclassification rates of our samples decreased after exposure to ambient weather. This was mainly due to higher prediction accuracy within the genus Microtus, when we excluded intestinal samples from modeling. It appears that species-specific signals in fNIR spectra of the closely related M. oeconomus and M. agrestis became more apparent with exposure. While we found no evidence of reduced prediction accuracy with increasing exposure time, rates of visible changes in spectra varied between weeks and species. Yet, it is likely that exposure times longer than 6 weeks will eventually lead to loss of species-specific signals as the fecal pellets decompose. Further research is needed to determine the maximum or optimal timeframe for successful classification of field-sampled rodent feces. How this compares with degradation of genetic material might indicate the suitability of genetic vs. spectral analysis for old fecal samples.
Comparing models that include both intestinal and exposed samples to models with exposed samples only allows for speculation as to which fecal constituents may have contributed to species identification, based on constituents' known susceptibility to exposure.
Specifically, species-specific signals in fNIR spectra seem not to associate with volatile or rapidly leaching substances. For instance, stability of constituents linked with dietary quality in fecal samples varies strongly under exposure (Jenks et al., 1990;Kamler et al., 2003;Leite & Stuth, 1994;Steyaert et al., 2012). Gut microbiota and the fecal metabolome likely displays phylogenetic and species-specific variation (Anders et al., 2021;Zierer et al., 2018) detectable by fNIRS (Santos et al., 2014;Saric et al., 2008). Fecal steroid metabolite concentrations of various large mammal species decline during days or remain stable for only up to a week (Abáigar et al., 2010;Mesa-Cruz et al., 2014;Parnell et al., 2015). Thus, steroid metabolites may be too short-lived to have accounted for species identification in our study. By contrast, differences in microbial flora could translate to some exposure-resistant differences in rodent fNIR spectra. For instance, diaminopimelic acid (DAPA), a marker of gut microbe-derived N in feces (Karr-Lilienthal et al., 2004), has good NIRS calibrations (Atanassova et al., 1998) and has been found to retain stable concentrations in exposed deer feces (Kamler et al., 2003). In summary, rodent species identification is likely to rest on a multitude of fecal constituents, many of which link with phylogenetic distance (Ley et al., 2008;Zierer et al., 2018) and have varying resistance to exposure and decay. We note that quoted studies on exposure effects on fecal dietary quality indicators and on fecal metabolome cover only up to a few weeks or days, respectively, providing only a preliminary basis for interpretation of constituents affecting fNIRS calibrations.

| Toward application of fNIRS calibrations for rodent monitoring
Our results corroborate previous findings of breaking past limitations of closed sample populations (Murguzur et al., 2019), as we show that increasing the spatial extent of calibration data will produce robust models across sample populations. However, we also caution for high or increased misclassification rates of samples In conclusion, fNIRS can facilitate rodent population censuses with larger sample sizes, combining large spatial extent with small grain if combined with pellet-count-based abundance indices (Engeman & Whisson, 2006;Jareño et al., 2014;Karels et al., 2004). The wide array of data (e.g., diet, disease, and stress) discernible through fNIRS suggests that developing monitoring schemes based on pellet counts and fNIRS could meet the need for ecosystem-and interaction-based approaches to monitoring (Ehrich et al., 2019) and for this purpose complement genetic methods (Zemanova, 2021). Further steps in this direction involve increasing the spatial scope of calibrations, extending the calibration exposure times to meet the needs of field sampling intervals, as well as continuous development of the calibration model algorithm and outlier detection. Key advantage of fNIRS is the ability to use existing spectral data to develop calibrations for any number of qualitative and quantitative variables (Foley et al., 1998;Vance et al., 2016), opening possibilities to link process and pattern across different levels of organization. For small rodents, fNIRS could provide invaluable data for nutritional ecology, stress and disease, underlying population dynamics and biotic interactions.

CO N FLI C T O F I NTER E S T S TATEM ENT
The authors declare no competing interests.