Assessing the Distribution and Stability of Groundwater Climatic Refugia: Cliff-Face Seeps in the Pacific Northwest

Abstract

Microrefugia can be critical in mediating biological responses to climate change, but the location and characteristics of these habitats are often poorly understood. Groundwater-dependent ecosystems (GDEs) represent critical microrefugia for species dependent on cool, moist habitats. However, knowledge of the distribution and stability of GDE microrefugia remains limited. This challenge is typified in the Pacific Northwest, where poorly studied cliff-face seeps harbor exceptional biodiversity despite their diminutive size (e.g., ~1–10 m width). To improve knowledge about these microrefugia, we regionally modeled their distribution and stability. We searched for cliff-face seeps across 1608 km of roads, trails, and watercourses in Washington and Idaho, while monitoring water availability plus air and water temperatures at selected sites. We detected 457 seeps through an iterative process of surveying, modeling, ground-truthing, and then remodeling the spatial distribution of seeps using boosted regression trees. Additionally, we used linear and generalized linear models to identify factors linked to seep thermal and hydrologic stability. Seeps were generally most concentrated in steep and low-lying areas (e.g., edges of canyon bottoms), and were also positively associated with glacial drift, basalt or graywacke bedrock types, high average slope within 300 m, and low average vapor pressure deficit. North-facing slopes were the best predictor of stable air and water temperatures and perennial seep discharge; low-lying areas also predicted stable seep water temperatures. These findings improve possibilities to manage seep microrefugia in the Pacific Northwest and safeguard their associated biodiversity under climate change. Lastly, our iterative method adapts techniques commonly used in species distribution modeling to provide an innovative framework for identifying inconspicuous microrefugia.

1. Introduction

Climate change has emerged as a major global threat to biodiversity over the past few decades and necessitates rapid conservation responses to minimize impacts on biota [1,2,3,4,5]. For example, increasing heat waves and drought can result in the loss of key habitats that were once reliably cool and moist, jeopardizing species that rely on those habitats [6,7,8,9]. Although often framed in the context of poleward and upslope range shifts [10], species’ responses to climate change can also heavily depend on access to in situ climatic refugia, which we define herein as pockets of locally favorable conditions enabling species persistence on landscapes that would be (or would soon become) otherwise climatically unsuitable [11,12,13,14]. Here, localized conditions become decoupled from larger scale climatic trends, due to phenomena like groundwater-based thermal buffering or cold air pooling [15]. As putative refugia can vary widely in their refugial capacity (e.g., thermal or hydrologic stability [11,16]), identifying and protecting the most stable climatic refugia—which are most likely to shield sensitive species from climate change impacts (e.g., increasing droughts and warming)—is an increasingly critical yet challenging conservation task [13,17,18]. While links between climatic refugia and environmental factors (e.g., topography, climate, and geology) have sometimes been investigated with respect to specific species [15,19], distributions of discrete refugium types (e.g., seeps, caves, etc.) with community-wide impacts remain poorly studied; particularly across large scales.

Microrefugia—i.e., refugia of diminutive size that often span hundreds of square meters or less—are among the most important to safeguard biodiversity under climate change, despite unique research and conservation challenges. Among the most important microrefugium types for conservation are climatic microrefugia, which are highly localized areas where organisms experience more stable, suitable conditions than on surrounding landscapes, enabling localized population persistence as the surrounding climate becomes unsuitable [13]. For example, non-vagile organisms (e.g., many amphibians, gastropods, and small mammals) can depend heavily on microrefugia, because they are disproportionately sensitive to climate change and may also struggle to colonize newly suitable habitats at a rate sufficient to offset losses [20,21,22], making the protection of climatic microrefugia critical for these species to persist. Notably, microrefugia—once properly identified—may also be particularly economical to protect relative to other biodiverse habitats, as their small size makes them compatible with cost-efficient “micro-reserve” conservation planning frameworks [23].

Currently, a lack of information about the distribution and quality (e.g., stability) of microrefugia is a major hurdle for conservation planning, although efforts to account for these habitats are increasing [24,25]. For example, localized stability at climatic microrefugia is often driven by a combination of discrete habitat features and unusual microclimates, but specific combinations promoting maximal “refugial capacity”—i.e., differential relative abilities of individual refugia to safeguard focal species from threats—have only recently begun to receive more research attention [26]. As a result, species distribution models often ignore the existence and capacity of relevant climatic microrefugia, which can lead to highly inaccurate climate change projections for associated species [27,28]. Unfortunately, the inconspicuous nature (e.g., small size, low discharge, and similar color to the surroundings) of many microrefugia can hinder efforts to identify them remotely (e.g., using satellite data), and manual field-based surveys may also struggle to detect these small features [29]. Therefore, new approaches are needed to streamline the detection of microrefugia and enable assessments of their refugial quality (e.g., temperature stability).

Groundwater-dependent ecosystems (GDEs; e.g., seeps and springs—defined interchangeably herein due to strong overlaps in their biologically relevant characteristics) are an excellent example of microrefugia that can shield sensitive species from rapid climate change impacts [12]. However, GDEs are often inconspicuous, and thus challenges in detecting them can lead to high uncertainty about their distributions and stability. In this context, we define stability as the degree to which air and water temperatures at groundwater outflow locations remain constant across several months, and the consistency with which these habitats supply water to the surface. Challenges in understanding GDE distributions and stability are notably exemplified by cliff-face seeps within the Pacific Northwest (PNW) United States, which are biodiverse yet understudied [30]. These seeps occur where impermeable rock layers force groundwater to the surface, creating a thin film of water that trickles down steep cliff faces [31]. The supply of water to the surface is perennial at some cliff-face seeps, and intermittent at others (pers. obs.). These seeps support a wide variety of moisture-dependent amphibians, mollusks, and hydrophilic plant species endemic to the PNW, many of which have narrow moisture and temperature requirements that may amplify their vulnerability to climate change [21,32,33]. The thermal and hydrologic regimes are of seeps tend to be decoupled—to varying degrees—from short-term weather patterns (e.g., daily or weekly precipitation and temperature), and locations with maximal decoupling (e.g., due to long groundwater residence times) may serve as critical refugial habitats for species dependent on cool and moist conditions, especially during severe heat waves and droughts [11,12]. Protecting microrefugia like PNW cliff-face seeps may therefore offer species a “slow lane” to climate change, allowing them more time to adapt [13]. However, pinpointing the location of these features can be challenging due to their small size and frequently inconspicuous appearance (e.g., <300 m² [30,34]); as much of the region is covered in evergreen forests, verdant, mossy seeps may not visually contrast with their surroundings to the same degree as in more arid regions. In the absence of obvious indicators, ad hoc efforts to locate these habitats have tended to be intuition-based and thus have had limited efficiency (pers. obs.). Lastly, the capacity of GDEs to serve as effective climatic microrefugia (i.e., safeguard species under climate change) likely varies based on thermal and hydrologic stability, which have not been quantified for cliff-face seeps.

In this study, we developed new methods to identify and characterize small and inconspicuous climatic microrefugia across a broad region (Figure 1), using PNW cliff-face seeps as a focal example that might shelter species from future climate change impacts. Given the ubiquity of at least mild thermal buffering conferred by groundwater discharge, we considered all cliff-face seeps (Figure 2) to be climatic microrefugia to at least some degree, but with varying levels of refugial capacity (i.e., prospects for sheltering climate-sensitive species) indicated by factors like seep thermal and hydrologic stability (e.g., air and water temperature variation at seeps, and their consistency of groundwater discharge). We sought to address these gaps at cliff-face seeps in our study, specifically within humid portions of the PNW, where seep-associated biodiversity is high [21,30,32]. Our primary objective was to identify factors associated with (1) seep occurrences, and (2) seep stability—the latter based on air and water thermal variation, alongside the consistency of surface flow. We considered these seep stability metrics to also be indicators of general refugial capacity, as areas with consistent surface moisture and high thermal buffering are more likely to shield specialist (e.g., stenothermic) species from negative climate change impacts [19]. In addition, we tested an iterative “survey-model-predict” approach wherein repeated rounds of field-based model validation, model updating, and prediction generation were carried out with the goal of maximizing cumulative seep detections over time. To our knowledge, our study is the first to apply this iterative approach to study inconspicuous microrefugia, despite the prominence of similar approaches used in species distribution modeling. More broadly, our study provides novel insights about how to identify rare and inconspicuous microrefugia, as well as estimates of seep occurrences and biologically relevant characteristics (e.g., temperature stability) that are publicly available as GIS layers and can aid in future studies of seep species. Considering the above outputs holistically, we also qualitatively assessed whether synergies or tradeoffs existed between different factors impacting seep refugial capacity, outlining consequences for future research and management.

2. Methods

2.1. Overview

We used a multi-pronged approach to locate, monitor, and then model the distribution and multi-metric (thermal and hydrologic) stability of seeps (Figure 1). To locate seeps, we surveyed 1608 total kilometers of roads, trails, and watercourses in the Olympic Peninsula, Washington Cascades, and Idaho Rocky Mountains, using biologically oriented criteria (defined below) to identify these habitats. These surveys were carried out in three rounds as part of an iterative survey, modeling, and prediction generation approach, designed to improve seep occurrence predictions using repeated validation-based methods that also iteratively increased sample size (i.e., total seeps detected). We used Boosted Regression Trees (BRTs [35]) to model the distribution of seeps, relative to physiographic and other spatial predictor variables (described below). To monitor thermal and hydrologic stability, we deployed aquatic and air-based temperature loggers at 43 seeps, representing a subset of 113 seeps where we also visually monitored the presence or absence of surface water (i.e., hydrologic stability). We calculated the relative thermal stability (for air and water) of each seep by comparing average temperature ranges among loggers, then used linear and generalized linear models (LMs and GLMs) to model three metrics of seep stability (air and water temperature stability and seep permanence) as a function of physio-climatic predictor variables. Relative stability metrics were calculated by comparing loggers of the same type (air or water) to others at different seeps, rather than deploying any loggers in non-seep habitats. After generating and mapping model predictions for each metric, we integrated all four metrics into a map of “total seep refugial potential”, which represented an equal combination of seep stability metrics, weighted by the likelihood of seep occurrence. Prediction raster layers and corresponding R code are freely available on ScienceBase, with an emphasis on regions within the physio-climatic envelope of our study areas.

2.2. Study Areas

Our cliff-face seep detection and monitoring efforts included three regions within the PNW, including the Washington Cascades, Olympic Peninula, and Idaho Northern Rockies. We chose these three study regions because they (1) have high cliff-face seep biodiversity of amphibians and gastropods [21,36], (2) are representative of a diversity of climatic conditions and geology across moist portions of the PNW [37], and (3) are known for their refugia that have shielded seep-associated species from historical climate change [38]. Amphibian and gastropod genera contributing to seep-associated biodiversity in these regions include—but are likely not limited to—Allogona, Ariolimax, Carychium, Dicamptodon, Euconulus, Haplotrema, Hemphillia, Paralaoma, Plethodon, Pristiloma, Prophysaon, Rana, Rhyacotriton, Vertigo, Vitrina, and Zonitoides in both the Washington Cascades and Olympic Peninsula, and Allogona, Ambystoma (pers. obs.), Anguispira, Cryptomastix, Dicamptodon, Euconulus, Haplotrema, Hemphillia, Kootenaia, Magnipelta, Oreohelix, Paralaoma, Plethodon, Polygyrella, Pristiloma, Prophysaon, Securicauda, Udosarx, Vertigo, Vitrina, Zacoleus, and Zonitoides in the Idaho Northern Rockies [21,36]. While many of the above are moisture-dependent terrestrial taxa that tend to occupy seep margins, it is plausible that fully aquatic biodiversity also depends considerably on these habitats within the PNW while being less well-documented. The three regions feature heterogenous moisture regimes but differ in terms of past and ongoing volcanic activity (which often dictates surface geology and can impact groundwater flow [39]), with the Washington Cascades being the only region wherein present-day volcanoes exist. Thermally buffered refugia in or near each of the three study regions are likely to have facilitated the persistence of seep-associated species like Plethodon idahoensis and P. vandykei during the Last Glacial Maximum, and all three regions continue to feature patches of thermally buffered rainforests today [37,40].

2.3. Seep Discovery

We conducted transect-based surveys (described below) to identify cliff-face seeps, which we defined—based on biologically oriented criteria—as groundwater-driven wet cliff faces (>30% maximum slope; threshold chosen subjectively based on amphibian and gastropod observations) containing predominantly madicolous (hygropetric) habitats (i.e., thin films of water). This definition aimed to focus on habitat features that are most important for seep-associated organisms (e.g., madicolous habitats), while deprioritizing other potentially less relevant hydrologic features (e.g., precise groundwater discharge locations); see Section S1 for additional information.

To locate cliff-face seeps, we established transects along 1608 km of roads, trails, and watercourses in the Olympic Peninsula, Washington Cascades, and Idaho Rocky Mountains during September 2020–September 2022. We concentrated our seep transects within the known ranges and elevational limits of seep-associated amphibians (e.g., Plethodon idahoensis, P. vandykei, and Rhyacotriton olympicus) and various lesser-studied gastropods [21], comprising a focal guild of potentially at-risk species. Most of these species (with some exceptions like R. olympicus) are entirely non-aquatic yet rely on moisture at the margins of seeps for survival throughout life [31,36]. In accordance with the distributional limits of these species, transects were generally limited to humid climates (i.e., >80 cm mean annual precipitation), below ~2000 m in elevation. We defined the start and end points of seep transects based on the termini of accessible road, trail, and canyon segments that were either intuitively likely to contain seeps (e.g., based on steep slopes and nearby headwater streams) or predicted to contain seeps based on preliminary modeling outputs (see below). Transects did not have a defined width; rather, we scanned for seeps visually in all applicable directions (e.g., along both banks of the watercourse, for canyon transects). As nearly all seeps were detected within <10 m of transect locations, the size of the field of view on any given transect did not generally introduce strong detection biases. Transects were conducted primarily during the dry season (July–October) to facilitate accurate approximations of the distance to dry-season groundwater emergence points, but we also conducted opportunistic transects during fieldwork for other projects, in either late spring (for low elevations) or early summer (for higher elevations). We defined dry-season groundwater emergence points as locations where water would come to the surface (based on clues like vegetation and rock staining; Section S1), assuming the smallest non-zero surface flow possible (i.e., a small trickle) possible. We included intermittent seeps in our definition as hydrological stability was modeled separately. However, we excluded locations where groundwater discharge appeared highly ephemeral (e.g., lasting for only hours, days, or weeks, after heavy rainfall or snowmelt) based on a lack of physical evidence for sustained surface flow (e.g., a lack of rock discoloration or sharp changes in mossy/algal surface cover); this step served to minimize biases in seep detections when conducting opportunistic spring surveys. If multiple groundwater discharge points emerged near one another, then we considered them separate seeps if they were not simultaneously visible when standing ~10 m back from the cliff base (judged by the same observer each time, for consistency); otherwise, we considered them one seep. We defined edges of seeps based on the outermost edges wherein signs of water discharge (e.g., based on discolored rocks or reduced moss cover) were evident.

2.4. Seep Distribution Model

We used BRTs and the “gbm” package in R (Version 4.3.2) [41] to identify relationships between the presence or absence of seeps and climate or landscape variables. Briefly, BRTs use iterative tree-based models to explain variation in the response variable, wherein each new set of models is scored based on how well it explains residuals from the previous set of models, while applying pre-specified levels of regularization, tree complexity (interaction depth), and proportions of training/testing data [41]. We used BRTs because of their high predictive accuracy, ability to accommodate missing data (e.g., for climate variables that are available in some regions but not others), and ability to fit irregular, threshold-based relationships between predictor variables and response data [35]. We fit BRTs with a binary response variable, coding all detected cliff-face seeps as 1 s and 10,000 random absence points (i.e., a random selection of points along transect lines; sample size chosen based on general statistical accuracy tending to plateau above this level) as 0 s. Absence points were randomly selected from portions of transects that did not contain seeps, so that they matched the same pattern of sampling bias as seep occurrence points; this avoided underestimating seep occurrences within unsurveyed areas [42]. To limit spatial autocorrelation, we required that random absence points be located >100 m from seeps and each other when creating these points in ArcGIS Pro (Version 3.5; ESRI Inc., Redlands, CA, USA). We constructed BRTs using a bag fraction of 0.5 to accurately estimate out-of-bag uncertainty, as this yielded the highest performance in initial exploratory analyses, wherein we tested various common values [35]. We used a tree complexity of three to strike a balance between acknowledging the existence of complex (3-way) interactions dictating seep locations while also acknowledging that increasingly complex interactions fit to our geographically restricted data might not be as broadly applicable across the entire PNW. Based on cross-validated AUC scores and model runtimes from initial BRTs, we used settings of maximum trees = 20,000 and learning rate = 0.0005 to maximize model consistency while allowing for reasonable runtimes.

Predictor variables for seep occurrences encompassed several potentially relevant factors, including 50 initial variables related to topography, groundwater storage, climate, vegetation, geology, and location (Table S1). A large suite of initial (but not final; see below) predictor variables was appropriate given high uncertainty about where seeps would occur, as well as our goal to accurately predict seep locations rather than testing specific hypotheses. Topographic predictor variables were included to account for the inherent steepness of cliff-face seeps, and because topography often influences groundwater flow dynamics [43]. We included predictors related to groundwater storage (e.g., base flow index) and climate variables (e.g., precipitation) because these variables sometimes influence how much groundwater can be theoretically supplied to the surface [44,45]. In addition, this supply of groundwater can be altered by plants consuming groundwater or creating shaded microclimates linked to low evaporation rates [46,47], making vegetative variables (e.g., volume of live trees) similarly relevant for modeling seep occurrences. Dominant rock type was included as it can also influence seep locations by impacting where groundwater flows, such as by dictating the size and concentration of impermeable rocks that force groundwater upwards [39,48]. This variable was compiled by combining various state and provincial geology databases [49,50,51], after standardizing their naming conventions. Notably, one of our predictor variables (“riparian climate corridor index” hereafter) was not a direct measurement of environmental conditions, but rather an index of potential refugial capacity for riparian animal movement corridors [52], based on a combination of predictor variable categories above. We chose to include this variable because it also integrates information about landscapes that might synergistically help shape groundwater dynamics (e.g., solar radiation exposure, stream width, and canopy cover). Lastly, we included region (Northern Rockies, Cascades, or Olympic Peninsula) as a predictor variable to account for otherwise unmodeled variation (e.g., driven by variables that lacked comprehensive data for region-wide modeling) in seep distribution patterns across different portions of the PNW. For example, these regions vary considerably in factors like volcanism, plate tectonics, and fire history [37,53]. To account for the small size of cliff-face seeps, we used the finest computationally viable (e.g., not exceeding storage/memory limits) resolution available for each predictor variable, yielding scales ranging from ~30 m (for topographic variables) to ~1 km (for climate variables). Given the general sparsity of seeps, we took an iterative approach to locating, modeling, and predicting seeps (referred to as “survey-model-predict” hereafter), similar to that used in previous studies to optimize distributional models for rare species [54,55,56,57]. First, we surveyed roads, trails, and watercourses that either appeared intuitively likely to contain seeps (i.e., in areas with steep slopes, adjacent to high-gradient headwater streams) or were located near known seep locations based on amphibian locality data. To minimize biases imposed by surveying only “intuitive” seep locations, we also surveyed less intuitive areas that were located in between hypothesized locations of seeps. After detecting >200 cliff face seeps, we used these to construct an initial seep distribution model (described previously), then used this model to predict unvisited locations (amongst a computer-generated set, placed every 100 m along unsurveyed potential transects) that were most likely to contain seeps. Next, we carried out a second iteration of transect surveys, designed to intersect as many points within the top 0.1% of predicted seep occurrence probabilities as possible (corresponding to an experience-based a priori estimate of 0.1% seep prevalence on the landscape). This approach served to validate or invalidate positive model predictions based on whether seeps were observed in the predicted locations. We also conducted opportunistic transects intersecting negative model predictions (i.e., unsuitable for seeps), but these were not a priority due to an expected false negative rate of <1% for even a poorly structured model, as cliff face seeps likely cover <1% of most landscapes. Given such sparsity, even completely random predictions would be expected to correctly predict >99% of seep absence locations, thus BRTs had inherently little opportunity to improve on this metric. After detecting >100 additional seeps during the second survey iteration, we rebuilt BRTs using all cumulative seep data and the same methods as previously and generated predictions as previously. We repeated this survey–model–predict approach once more to locate additional seeps and generate final BRTs and predictions.

After constructing initial BRTs, we used recursive feature elimination to improve model predictions [35], which involved removing predictor variables until every remaining predictor yielded a >5% contribution to the model and was non-redundant with other predictor variables. We addressed pairwise redundancy among predictor variables (i.e., |r| > 0.7) for this same purpose, removing the variable with the lower contribution to the preliminary BRT model. This parsimonious structure also aimed to improve the applicability of model predictions for areas beyond those used in model fitting, acknowledging that tradeoffs between model complexity and generalizability can occur [58]—particularly when BRT predictor variables are redundant [35].

Models from our first two iterations contained ~800 × 800 m BioClim variables plus the other variables described above, whereas our final modeling iteration considered two different formulations to BRTs; one in which BioClim predictor variables were experimentally downscaled to ~180 × 180 m (see Section S2 for details), and another in which they (as previously) were not. We evaluated downscaled climate data as a potential modeling necessity due to the potential for microclimates to vary over fine spatial scales in steep regions in ways that could potentially impact seeps [14,27]. To do so, we constructed BRTs in parallel for these two resolutions of climate data while otherwise following the same modeling approach as above, then used cross-validated AUC scores and Pearson’s correlation coefficient (for model predictions) to compare the two models and determine whether downscaling was necessary (see Section S2 for details).

2.5. Monitoring Thermal and Hydrologic Stability

We used multiple methods to monitor the thermal and hydrological stability of a subset of the seeps detected during transects. To assess hydrologic stability (i.e., seep permanence or intermittency), we visited 113 seeps 2–3 times each during 2021–2022 (including at least one visit during September-October, when seeps are driest) and noted the presence or absence of flowing water during each visit. In addition, we monitored thermal stability at a subset of these seeps (n = 43; manually identified to maximize variation in predictor variables) by deploying HOBO MX2201 temperature loggers (Onset Inc., Bourne, MA, USA) in the water and air at each seep in Fall 2021 and Spring 2022 (loggers installed in 2021 remained in place but were inaccessible during winter). We attached water-based temperature loggers to small (~12 cm) garden stakes using zip-ties, then installed them at the bottom of the deepest part of each seep with a metal mallet. To monitor air temperature, we deployed loggers ~0.6–0.7 m above the ground at the most heavily shaded portion of the seep by attaching them near the top of ~91 × 2 cm metal stakes using zip-ties. We deployed temperature loggers at seeps located in Olympic National Park, Mount Rainier National Park, and the North Fork Clearwater Drainage in Idaho to infer seep thermal stability, whereas seeps monitored for permanence or intermittency included these same areas as well as the North Fork Coeur d’Alene Drainage in Idaho (Figure 3).

To quantify biologically important thermal stability at different seeps, we cleaned seep temperature measurements, then calculated “relative thermal variation” for each seep (with separate calculations for air and water; see Section S3 for details). We calculated “relative thermal variation” as the difference between the observed thermal range (90th–10th percentile of daily mean temperatures) at a focal logger and expected thermal range (90th–10th percentile of daily mean temperatures across all loggers; see Section S3 for details). We used only the middle 80% of daily temperature data—and not brief extremes beyond this range—because most seep-associated species can temporarily burrow to avoid briefly unsuitable or volatile surface temperatures. Precise cutoff values chosen to account for these situations (e.g., 10th and 90th percentiles within this study) were inherently somewhat arbitrary, due to a sparsity of relevant physiological data for seep-associated species.

2.6. Seep Thermal and Hydrologic Stability Models

We used LMs and GLMs [59] to model seep thermal and hydrologic stability due to their performance with modest sample sizes [60], and because we had a sample size smaller than what tends to be required to construct robust BRTs [61]. To limit overfitting, we tested only five predictor variables that we expected would be among the most informative for predicting seep thermal and hydrologic stability: slope, elevation, aspect, vertical distance to the highest point within 1 km (calculated by using the “focal statistics” tool in ArcGIS Pro (Version 3.5) to find the highest elevation within 1 km of each cell of a DEM raster, then subtracting the local elevation of the cell), and the maximum riparian climate corridor index within 90 m (represents a one-cell radius given the layer’s native 90 m resolution; values coded as zero for all seeps outside of riparian corridors [52]). We expected these variables to be associated with seep stability, due to their strong impacts on surface temperature [62]; all other BRT predictor variables were excluded from stability models due to sample size-based limitations. Due to modeling limitations imposed by sample size (n = 113, 30, and 23, for hydrologic, air temperature, and water temperature stability, respectively), we constructed a set of 15 LMs corresponding to single-variable models (n = 5) plus all additive combinations of two predictor variables (n = 10) for relative seep air and (separately) water temperature variation. While this additive design was necessary given small sample sizes, we also acknowledge that a lack of interaction terms in these models may have limited their predictive power. For seep hydrologic stability (i.e., permanence or intermittency), we constructed 25 binomial GLMs corresponding to single-variable models (n = 5), plus all additive and interactive combinations of two predictor variables (n = 20). We also constructed a null model for comparison in all three modeling sets and ranked models based on second-order Akaike’s Information Criterion (AICc [63]). We judged variation in AICc values on a continuum rather than defining singular thresholds for important or unimportant models. As we were primarily interested in generating accurate predictions rather than uncovering mechanistic links between individual predictor variables and the response variable, we used AICc-weighted model averaging to generate predictions for seep thermal and hydrologic stability [64,65]. We excluded models that performed worse (i.e., higher AICc score) than the null model (which explained no variation in the response data, by definition) from these calculations, due to their uninformative nature.

2.7. Generating Output Maps

After computing AICc-weighted model averages, we generated output maps by predicting seep occurrence likelihood and our three seep stability metrics across space using the ‘raster’ package in R (Version 4.3.2) [66]. In addition, we generated a raster map depicting an integrative metric of “total seep refugial capacity”, using equal (z-scored then summed directly, reflecting a lack of comparative knowledge about their relative biological importance) contributions of three seep stability metrics, weighted by predicted seep occurrence probability (see Section S4 for details). These raster layers included predictions for every 30 m pixel within the Pacific Northwest, thus they did not require interpolation. Lastly, we clipped each output raster to portions of the PNW that fell near or within the physio-climatic envelope of our study regions (defined in Section S5). Briefly, we considered predictions for areas within the physio-climatic envelope of our study locations to be of high or moderate confidence and included these in all output maps, whereas less similar regions (e.g., deserts) were presumed to contain less accurate predictions (“low” or “very low” confidence) and thus included only in a supplementary version of these maps (see Section S5 for details).

3. Results

We detected 457 cliff-face seeps across 1608 km of transects, including 99 seeps across 332 km (0.30 seeps/km) in the Olympic Peninsula, 264 seeps across 743 km (0.36 seeps/km) in the Washington Cascades, and 94 seeps across 533 km (0.17 seeps/km) in northern Idaho (Figure 3). We detected roughly half of all seeps within the first survey iteration, one quarter during the second, and one quarter during the third. Our third iteration of surveys detected seeps at a ~5–10× higher (depending on location) rate per unit of transect distance compared to the first iteration, indicating increasingly accurate predictions. Based on an average wetted width of 2.9 m calculated for cliff-face seeps as part of a separate ongoing study (Button et al., in prep.), these habitats were rare, occupying only ~0.09% of the total transect length. Qualitatively, cliff-face seeps exhibited high spatial clustering, wherein relatively short transect segments with abundant seeps were often punctuated by much longer segments completely lacking in these habitats.

Our final BRT model predicted cliff-face seep occurrences with relatively high accuracy (cross-validated AUC = 0.891) and included nine predictor variables that were sufficiently influential to be retained (i.e., >5% relative influence;

Table 1. Relative importance of variables included in the final seep distribution model.

Variable	Relative Influence (%)
Slope (°)	16.6
Vertical distance to highest point within 1 km (m)	11.7
Dominant rock type	11.5
Average slope within 300 m (°)	10.0
Average vapor pressure deficit (1981–2010; kPA)	9.9
Precipitation of warmest quarter (1991–2020; mm)	9.9
Elevation (m)	8.7
Average potential evapotranspiration (1981–2010; mm/day)	8.6
Landsat Enhanced Vegetation Index (September 2022)	7.2
Maximum Riparian Climate Corridor Index within 90 m	6.1

; Figure S1). Slope was the most informative predictor of cliff-face seep occurrences (16.6% relative influence), while vertical distance to highest point within 1 km, dominant rock type, average slope within 300 m, average vapor pressure deficit, and precipitation of the warmest quarter each exerted similar secondary levels of influence (9.9–11.7%). In general, cliff-face seeps were most likely to occur within sharply sloping areas located relatively far below adjacent high-points (e.g., along steep-sided canyon bottoms). In addition, seeps tended to be most common in areas with a low mean vapor pressure deficit (based on CHELSA-BIOCLIM+ climate data [67]) and basalt, glacial drift, or graywacke as the dominant rock type. The results did not suggest any clear interactions between bedrock type and other predictor variables. During the final round of ground-truthing the seep predictions, false positives occurred >50% of the time at the 30 m scale of the map pixels, but in such cases, we nevertheless tended to find seeps within a few hundred meters (or less) of the targeted pixel. Predicted seep occurrence probabilities were similar for models constructed with ~800 m and downscaled ~180 m BioClim variables (r = 0.97 after logit-transforming) and yielded CV AUC scores within 0.01 units of each other. Downscaling climate data was considered unnecessary given that it did not substantially alter our results at model-fitting points and would have been arduous to conduct on a region-wide basis. Thus, our final BRTs used the pre-existing ~800 m resolution of available BioClim data.

The strongest predictors of seep stability depended largely on the specific output metric (air temperature stability, water temperature stability, and seep permanence probability). However, aspect was an important predictor for all three outputs, and more northerly aspects tended to feature more thermally stable and permanent seeps (

Table 2. AICc results for seep air temperature models. K = number of model parameters; CCI_Max = maximum riparian climate corridor index [52] within 90 m; VDist1km = vertical distance to highest point within 1 km. Aspect values were transformed such that they represented the angular difference from a north-facing slope.

Predictor Variables	K	AICc	ΔAICc	AICc Weight	Cumulative Weight	Log Likelihood
Aspect + Slope	4	131.22	0.00	0.58	0.58	−60.81
Aspect + Elevation	4	132.90	1.68	0.25	0.83	−61.65
Slope + Elevation	4	134.91	3.69	0.09	0.92	−62.65
Elevation	3	135.60	4.38	0.06	0.89	−64.34
Aspect + VDist1km	4	137.70	6.48	0.02	0.94	−64.05
Elevation + CCIMax	4	137.77	6.55	0.02	0.96	−64.09
Aspect	3	138.00	6.78	0.02	0.95	−65.54
VDist1km + Elevation	4	138.26	7.04	0.02	0.98	−64.33
Slope	3	138.28	7.06	0.02	0.98	−65.68
Slope + CCIMax	4	140.13	8.91	0.01	0.99	−65.26
Aspect + CCIMax	4	140.23	9.01	0.01	0.99	−65.32
Slope + VDist1km	4	140.69	9.47	0.01	1.00	−65.54
None (null model)	2	141.93	10.71	0.00	1.00	−68.74
VDist1km	3	142.52	11.30	0.00	1.00	−67.80
CCIMax	3	144.41	13.19	0.00	1.00	−68.74
VDist1km + CCIMax	4	144.84	13.62	0.00	1.00	−67.62

,

Table 3. AICc results for seep water temperature models. K = number of model parameters; CCI_Max = maximum riparian climate corridor index [52] within 90 m; VDist1km = vertical distance to highest point within 1 km. Aspect values were transformed such that they represented the angular difference from a north-facing slope.

Predictor Variables	K	AICc	ΔAICc	AICc Weight	Cumulative Weight	Log Likelihood
Aspect + VDist1km	4	116.93	0.00	0.27	0.27	−53.35
Aspect	3	117.73	0.79	0.11	0.28	−55.23
VDist1km	3	117.96	1.03	0.10	0.38	−55.35
None (null model)	2	118.11	1.18	0.15	0.41	−56.76
Elevation	3	118.35	1.42	0.08	0.55	−55.54
VDist1km + CCIMax	4	118.66	1.73	0.11	0.52	−54.22
Aspect + Slope	4	118.97	2.03	0.10	0.62	−54.37
Aspect + Elevation	4	119.11	2.17	0.09	0.71	−54.44
Slope	3	119.27	2.34	0.05	0.79	−56.00
Elevation + CCIMax	4	119.41	2.48	0.08	0.79	−54.59
CCIMax	3	120.11	3.18	0.03	0.87	−56.42
Aspect + CCIMax	4	120.49	3.55	0.04	0.83	−55.13
VDist1km + Elevation	4	120.51	3.58	0.04	0.88	−55.14
Slope + VDist1km	4	120.54	3.61	0.04	0.92	−55.16
Slope + Elevation	4	120.68	3.75	0.04	0.96	−55.23
Slope + CCIMax	4	120.72	3.79	0.04	1.00	−55.25

and

Table 4. AICc results for seep permanence models. K = number of model parameters; CCI_Max = maximum riparian climate corridor index [52] within 90 m; VDist1km = vertical distance to highest point within 1 km. Aspect values were transformed such that they represented the angular difference from a north-facing slope.

Predictor Variables	K	AICc	ΔAICc	AICc Weight	Cumulative Weight	Log Likelihood
Aspect × Slope	4	132.87	0.00	0.20	0.20	−62.25
Aspect + Slope	3	133.36	0.49	0.16	0.35	−63.57
Slope × Elevation	4	133.54	0.66	0.14	0.50	−62.58
Slope	2	133.81	0.94	0.11	0.55	−64.85
Slope + VDist1km	3	133.98	1.11	0.11	0.61	−63.88
Slope + CCIMax	3	134.11	1.24	0.11	0.72	−63.94
Slope × CCIMax	4	134.75	1.88	0.08	0.79	−63.19
Slope + Elevation	3	135.17	2.29	0.06	0.86	−64.47
Slope × VDist1km	4	135.65	2.78	0.05	0.91	−63.64
Aspect × CCIMax	4	135.97	3.10	0.04	0.95	−63.80
Aspect + CCIMax	3	136.95	4.08	0.03	0.98	−65.36
Aspect × VDist1km	4	139.48	6.61	0.01	0.98	−65.56
Aspect + VDist1km	3	140.43	7.55	0.00	0.99	−67.10
CCIMax	2	140.67	7.79	0.00	0.98	−68.28
Aspect	2	140.74	7.86	0.00	0.99	−68.31
Aspect + Elevation	3	141.76	8.89	0.00	0.99	−67.77
Elevation + CCIMax	3	141.82	8.95	0.00	0.99	−67.80
None (null model)	1	141.90	9.03	0.00	0.99	−69.93
VDist1km	2	142.02	9.15	0.00	0.99	−68.96
Elevation	2	142.18	9.31	0.00	0.99	−69.04
VDist1km + CCIMax	3	142.23	9.36	0.00	1.00	−68.01
VDist1km + Elevation	3	143.12	10.25	0.00	1.00	−68.45
Elevation × CCIMax	4	143.76	10.89	0.00	1.00	−67.70
VDist1km × CCIMax	4	143.87	11.00	0.00	1.00	−67.75
Aspect × Elevation	4	143.88	11.01	0.00	1.00	−67.76
VDist1km × Elevation	4	145.22	12.35	0.00	1.00	−68.43

, Figure 4, Figure 5 and Figure 6), and the influence of this variable was quantitatively clear in top-ranking models despite small sample sizes (95% CIs: β = 0.44–1.80 for the air temperature model, −0.01–2.18 for the water temperature model, and β = −1.75–(−0.11) for the seep permanence model). In addition, seeps with relatively gentle 30 m-scale slope values (e.g., <20°) had overall relatively stable air temperatures (95% CI: β = 0.55–2.34 [corresponding to thermal variation—i.e., the opposite of stability]) and high permanence (95% CI: β = −1.44–(−0.33); see also Figure 6), meaning that localized seep-bearing cliffs in these pixels were relatively small and were surrounded by flatter areas within the same 30 m pixel. Seeps in low-lying areas had the most stable water temperatures (95% CI: β = −2.79–0.06 [corresponding to thermal variation]); see also Figure 6). Model performance relative to the corresponding null model was better for the top-scoring seep air temperature stability model (ΔAICc = 10.7 for the null model; Table 2) and seep permanence model (ΔAICc = 9.2; Table 3) than for the top-scoring water temperature stability model (ΔAICc = 1.2; Table 4).

Our integrative estimate of total seep refugial capacity, as defined based on seep occurrences and stability combined (see methods), was primarily dictated by seep occurrence probability (Figure 4 and Figure S2). However, parameters associated with thermal and hydrological stability contributed substantially to calculations of overall seep refugial capacity for areas where seeps were deemed likely to occur (e.g., medium elevation portions of the Olympic Peninsula); this was clearest when visualizing results for these areas at a relatively fine spatial scale, wherein divergent relative pixel scores for the different seep stability metrics reflect the sensitivity of total seep refugial capacity to additional factors beyond those linked to seep occurrences (Figure 5). Numerous portions of the Washington and Oregon Cascades, Olympic Peninsula, Northern Rockies, and Washington and Oregon Coast Ranges (Figure 4 and Figure S3) possessed sufficiently similar traits to sampling locations for predictions to be classified as “high” or “medium” certainty (see Section S5 for definitions and methods), whereas other areas featured “low” or “very low” predictive confidence (Figure S3) and were therefore only included in supplementary maps.

4. Discussion

4.1. Summary of Key Findings

Our study highlights the utility of our iterative survey–model–predict approach to identify inconspicuous microrefugia and improves knowledge of the distribution and stability of Pacific Northwest (PNW) cliff-face seeps. We found that these groundwater-dependent ecosystems (GDEs) were closely linked to steep, low-lying areas (e.g., steep-sided canyon bottoms) and certain rock types (e.g., basalt, glacial drift, and graywacke) and were most stable on north-facing slopes. Encouragingly, we detected seeps at a ~5–10× higher rate per unit of distance traveled during our final round of surveys than when searching for seeps initially, highlighting the importance of our iterative approach to maximize sample size for these inconspicuous features. While similar iterative modeling frameworks have proven valuable in species distribution modeling [57], our study is among the first to adapt this framework to microrefugia, conceptualizing it in terms of discrete habitat features in place of more continuous populations. Moreover, our seep air temperature and permanence models explained considerable variation in seep refugial properties, with a northerly aspect being the strongest predictor of seep stability overall. Seep water temperature stability models performed worse than air temperature and permanence models, however; therefore, a version of our holistic model outputs (i.e., “total refugial capacity” estimates) that excludes water temperature as an input has been included alongside others on ScienceBase. Our study’s publicly available output maps are an important first step towards understanding how seep microrefugia and their associated species may be impacted by future climate change.

4.2. Seep Distribution Model

Findings of our seep distribution model (Figure 4) were generally consistent with prior studies on aquifers and groundwater discharge patterns. For example, multiple types of seeps (e.g., gushets and hanging gardens; both referred to by hydrologists as “springs”) are tied to “near-vertical” instances of “cliff seepage conditions” [68,69], which supported the most seeps in our study. In addition, our finding of high seep occurrence probabilities in relatively low-lying areas (i.e., pixels with a large vertical distance to the highest point within 1 km) is consistent with their larger expected aquifers and upslope groundwater recharge zones [43]. Similarly, seeps were relatively common in areas with certain volcanic rocks (e.g., basalt) or glacial drift, consistent with previous evidence that these rock types efficiently form aquifers [70,71]. However, it is also possible that the relationship between rock type and seep occurrences largely reflects topographic influences of rock type (e.g., on slope) that can have subsequent impacts on seep formation, rather than more direct causal pathways. Climate can also impact aquifer characteristics [43], consistent with our finding of relatively high seep occurrence probabilities in areas with high summer precipitation (which tends to occur only in temperate rainforests, which indeed have high seep densities) and low vapor pressure deficit (i.e., the disparity between the theoretical maximum and actual amount of moisture held in the air, given the temperature). However, these findings may not directly correspond to precise mechanisms for seep formation, as BRT analyses sacrifice mechanistic interpretations in exchange for maximizing predictive accuracy [35]. We favored this approach because our goal was to identify geographic targets for future seep research and conservation, rather than precise seep-generating mechanisms. However, our region-wide seep predictions may allow future studies to more easily identify such mechanisms, by making it easier to locate potential study sites. Lastly, as downscaling climate data had little impact on BRT results, it is likely that including fine-scale topographic variables in the models—and allowing them to interact with 1 km climate variables—was sufficient to capture sources of finer-scale climatic variation inasmuch as these impact cliff-face seeps.

4.3. Seep Stability Models

Results of LMs and GLMs suggested high stability on north-facing slopes, as well as modest potential tradeoffs between factors maximizing seep stability versus maximizing the likelihood of seep occurrence. North-facing slopes were the most reliable predictor of seep stability overall and were linked to both thermal and hydrologic stability, consistent with the existence of cool microclimates on north-facing slopes, which results from their limited sunlight exposure [72]. Given these characteristics, and also because moisture-dependent and stenothermic species often prefer sheltered north-facing slopes [73], such areas have high conservation value for species associated with cliff-face seeps. In addition, seeps were—assuming their occurrence in an area—more likely to be permanent and thermally stable (for air temperature) when surrounded by relatively gently sloping terrain on a 30 m scale (Figure 6). This finding was surprising, because steep slopes (i.e., >30% maximum incline) were a prerequisite to identify cliff-face seeps at a more localized scale. However, a modest number of seeps in our study emerged from small, isolated drop-offs within overall more gently sloping areas, and our findings suggest that these were more stable, on average, than more numerous seeps emerging from larger cliff bands. One plausible explanation for this pattern is that small, isolated drop-offs within certain otherwise-flat areas (e.g., floodplains) may expose relatively large (and thus more thermally buffered) groundwater reservoirs to the surface [43]. Alternatively, cold air pooling might also make relatively flat valley bottoms have more stable air (but not necessarily water) temperatures [74], consistent with our results. Given the association between 30 m-scale gentle slopes and air temperature stability, it is conceivable that such slopes are also linked to water temperature stability but that we failed to capture this due to a smaller sample size for water temperatures (n = 23) than air temperatures (n = 30). In contrast to the above linkages, seeps emerging from larger cliff bands (e.g., headwall seeps) are more likely to be fed by perched aquifers, which tend to be smaller and less stable [75]. Therefore, while seeps are most likely to occur near cliff bands (Table 1), modest tradeoffs may exist with seep stability. Nonetheless, a larger number of stable seeps are expected at large cliff bands than in flatter areas overall, because the much greater number of seeps at these cliff bands compensates for overall their lower average (i.e., not accounting for seep-to-seep variation) stability. Therefore, we recommend prioritizing steep, low-lying, north-facing slopes (including large cliff bands) and potential upslope recharge zones [76] to best safeguard stable seep microrefugia overall. Importantly, these suggestions apply only to cliff-face seeps and are likely not generalizable to certain other GDEs (e.g., helocrene and mound form springs) that are less associated with steep slopes [68,69]. However, our model may serve as a partial proxy for other cliff-associated moist habitats (waterfall splash zones, moist talus slopes, and steep ephemeral rivulets) in future studies, as these frequently occurred near cliff-face seeps (pers. obs.).

4.4. Spatial Scale-Based Considerations

Our seep prediction output maps have applications across multiple spatial scales and for both research and management. For example, as the 30 m scale of these raster layers is approximately consistent with fine-scale habitat use by non-vagile seep-associated species (in addition to being the finest resolution that many computers can store, given large spatial extents), such outputs may be useful for assessing species’ distributions. Encouragingly, these layers have already improved SDMs for two seep-associated amphibians (Plethodon idahoensis and P. vandykei; Button et al., in prep.), highlighting their high potential biological relevance. However, additional research is needed to identify potential climate change impacts on cliff-face seeps and how these may in turn affect species, as our seep-related models were strictly correlational and based on contemporary measurements. As such, contemporarily seep-related predictor variables in our models could become decoupled from the ultimate drivers of seep distribution and stability under future climate change and thus become less informative of seep locations and stability. In addition, false positive predictions were generally common (albeit often linked to alternative biologically relevant moist habitats) when modeling seeps at a 30 m scale, based on model validation-focused seep surveys (pers. obs.). However, false positive seep predictions (at a 30 m scale) were often located within a few hundred meters of a seep. Thus, we recommend rescaling the resolution of prediction rasters to ~500 m for many applications; this resolution is likely more accurate and thus potentially more useful for larger-scale conservation planning. Overall, our output maps represent a major improvement for efficiently detecting cliff-face seeps, as these habitats covered <0.1% of our total transect length, making them labor-intensive to manually locate.

4.5. Interpretation of Output Maps

Given visually apparent patterns in mapped outputs and limitations of our air temperature stability models, we suggest multiple key considerations to improve the future use and interpretation of seep-related predictions. For example, seep thermal and hydrologic stability were intentionally non-influential on predicted overall refugial capacity in areas where seep occurrences were unlikely, which corresponded to the vast majority of all map pixels. Nevertheless, these stability metrics exerted a substantial influence on predicted overall refugial capacity in more localized pockets wherein seeps were deemed more likely to occur (Figure 5). Thus, while predictions of total seep refugial capacity appear nearly identical to a “seep occurrence probability” map when viewed on a region-wide scale (Figure 4), the two should not be used synonymously. In addition, our estimates of seep water temperature stability should be interpreted with caution due to poor-fitting models for this parameter (Table 3). As such, we have included a supplemental version of our total seep refugial capacity layer on ScienceBase that excludes water temperature stability from the corresponding calculation. However, we recommend that future studies—if using this incomplete, supplemental version of the layer to model seep-associated species’ distributions—incorporate other forms of water temperature data simultaneously (e.g., NorWeST temperature data for adjacent streams [77]).

4.6. Key Caveats and Limitations

As associations between seeps and climate or landscape variables may shift continuously across space and between ecoregions (e.g., inland deserts versus coastal rainforests), decisions about how to quantify and define boundaries of acceptable confidence levels, when creating seep prediction maps, were inherently subjective. Our qualitative approach to classifying uncertainty for seep predictions was relatively conservative, removing all map pixels that were geographic or physio-climatic outliers relative to training data. This approach minimized the risk of low-quality predictions in the outputs by eliminating qualitatively risky areas for such predictions to occur. However, although highly speculative, our model predictions—if treated cautiously as hypotheses—may still be informative of seeps in certain areas that fell outside of acceptable physio-climatic limits when defining our model projection area. Therefore, seep-related prediction layers are also available for the entire PNW region as a Supplementary figure herein (Figure S2), and as additional downloadable raster files on ScienceBase. However, we recommend that predictions from areas labeled as “low” or “very low” confidence (Figure S3) be handled with extreme caution, particularly within arid climates that are highly dissimilar to our study areas. For example, our model predicts abundant seeps within certain steep portions of eastern Oregon deserts, which is probably unlikely given this region’s low precipitation [78]. Despite this limitation, our model projection area (i.e., pixels with predictions of high or medium confidence; Figure S3) nonetheless covers large areas within the ranges of many seep-dwelling species [21,36].

Importantly, our metric for “total seep refugial potential” was based on abiotic factors only and does not take the habitat preferences of seep-dwelling organisms into account. Additional research is needed to understand patterns of seep biodiversity and determine where biodiversity and refugial characteristics overlap; these seeps will have the highest conservation value [12]. Ongoing research suggests that amphibian and gastropod biodiversity is highest at low elevation seeps (e.g., <600 m; S. Button unpublished data), whereas the highest predictions of seep refugia in this study were often for moderate elevations (e.g., 900–1700 m), presumably corresponding to “headwalls” below much higher peaks. The total refugial capacity of these moderate elevation seeps may be reduced, in practice, if they can support fewer species than elsewhere. Nevertheless, as the climate continues to rapidly warm around them, these moderate elevation seeps might eventually become viable climatic refugia for some historically lower-elevation species, if they can migrate upslope.

4.7. Management Implications

Our study identified factors associated with the distribution and stability of cliff-face seeps, with a focus on prediction-oriented models and the creation of broadly applicable output rasters, which are freely available on ScienceBase (doi: 10.5066/P1989RPN). An iterative survey–model–predict approach was useful for studying these sparse habitat features, which tended to occur most frequently in areas like steep-sided canyon-bottoms and be most stable on north-facing slopes. Our findings provide an innovative framework for studying rare and inconspicuous microrefugia, and our findings have broad applications both within and beyond our study system. Within moist portions of the PNW, publicly available GIS layers from our analyses may be useful to help formulate biological inventories and geospatial models (e.g., SDMs) for local seep-dependent species (e.g., dozens of amphibians and gastropods), many of which may be sensitive to climate change [21,36]. These outputs can also be further operationalized to protect stable cliff-face seeps, by incorporating areas likely to contain them into “micro-reserves”, habitat corridors, or monitoring efforts (e.g., for seep amphibians and gastropods). While many of these species also use other moist habitats (e.g., headwater riparian areas) currently, seep populations may become increasingly relictual under future climate change as less refugial habitats are destabilized. Thus, seep microrefugia may ultimately function as the most (or only) viable source populations for repatriating certain specialist (e.g., stenothermic) species to their broader historical ranges. In our study region, areas important for seep refugia (e.g., north-facing slopes and bottoms of steep-sided canyons) also sometimes featured land uses that could hamper their conservation, such as land-clearing for road construction or ski resorts (pers. obs.). In such cases, impacts on seep species could be partially mitigated by avoiding the use of overhanging culverts [79], constructing wildlife corridors (e.g., underpasses [80]), and avoiding ground disturbance or canopy removal at seep locations [81]. In addition, as seeps and other GDEs currently face other threats such as anthropogenic depletion, depressurization, and contamination of their contributing aquifers [82,83,84,85], assessments of water quality and flow metrics at seeps detected in this study may aid in developing future management priorities. However, given sample size limitations and somewhat climatic diversity among our study sites, we recommend that our findings for water temperature stability, and for areas of “low” or “very low” confidence levels (Section S5; Figure S3), be viewed as tentative hypotheses.

4.8. Conclusions

Our iterative survey–model–predict approach may be broadly useful for characterizing many types of microrefugia, and our study represents a novel use of this iterative approach to successfully detect rare habitat features, demonstrating its expanded utility beyond past uses to optimize distributional models for rare species [54,55,56,57]. Consistent with previous evidence for rare species, our findings demonstrate that an iterative survey–model–predict approach may be especially applicable for habitats that are rare, inconspicuous, and can be intuitively conceptualized as discrete features (e.g., cave entrances, waterfall splash zones, desert oases, phytotelmata, deep sea thermal vents, hot springs, small rock outcroppings, and isolated small ponds, bogs, and fens). However, while optimal iterative frameworks for identifying such habitats may overlap with those used in past species distribution modeling exercises (e.g., references [54,55,56,57]), we caution against exact replication considering the differential properties of organisms and microrefugia. Most notably, while proximal species occurrence records are often “thinned” in species distribution modeling to avoid overrepresentation of any single population [86], spatially adjacent microrefugia are, in contrast, likely to have additive individual-level value for conservation (e.g., contribute additively to species’ overall regional carrying capacities), making point-thinning undesirable within iterative frameworks for microrefugia. Applying the above principles, a standardized approach that considers different types of microrefugia simultaneously could help streamline future conservation planning by determining whether any physio-climatic variables (e.g., steep slopes) can predict multiple microrefugium types at once. This approach may also assist in identifying viable study sites for monitoring biodiversity and habitat characteristics at diverse microrefugia over time, making it possible to assess how microrefugia impact key ecological factors like metacommunity dynamics, species–area relationships, co-existence mechanisms, dispersal, gene flow, phylogeographic patterns, and population, community, and evolutionary trajectories.