No place to hide: Rare plant detection through remote sensing

Detection of rare species is limited by their intrinsic nature and by the constraints associated with traditional field surveys. Remote sensing (RS) provides a powerful alternative to traditional detection methods through the increasing availability of RS products. Here, we assess the capacity of RS at high and medium resolution to detect rare plants with direct and indirect approaches, and how the performance of RS can be influenced by the characteristics of species.

Remote sensing (hereafter "RS"; see Box 1) has become an important tool for the scientific community in addressing these field survey issues, offering an inexpensive method to assess biodiversity characteristics over large areas at regular intervals (Corbane et al., 2015;Kerr & Ostrovsky, 2003). RS allows both (i) detection of individual biological entities, species assemblages or ecological communities (direct approach) and (ii) acquisition of biodiversity-related information from environmental proxies (indirect approach; Turner et al., 2003).
RS has become a widely applied tool in plant studies (e.g. Kopeć et al., 2020;Wan et al., 2020) with the increasing availability of RS products that capture a wide variety of environmental features (Corbane et al., 2015;Kerr & Ostrovsky, 2003;Turner et al., 2003). Studies using RS to focus specifically on rare plants remain uncommon, although their number has been growing in recent years (e.g. Arenas-Castro et al., 2019;Gonçalves et al., 2016;Zhu et al., 2016). As an emerging research field, the potential benefits of RS for the detection of rare plants remain unclear. In this context, the objectives of this synthesis were i) to evaluate the capacity for RS to detect and predict the occurrence of rare plants, and ii) to assess how the main characteristics of rare plants influence the performance of RS. Our concept of "rarity" is based on Rabinowitz's rare species classification (Rabinowitz, 1981), which discerns seven rarity types based on three dichotomous criteria: geographic distribution range (large versus. restricted), habitat specificity (wide versus. narrow), and local population size (large versus. small). Since these criteria are characterized by a continuous transition among the different rarity categories (absence of defined thresholds) and make abstraction of causes of rarity, it is a flexible

BOX 1 Remote sensing (RS)-related terms and abbreviations
Active sensor: Emits radiation and measures the energy returned after being reflected.
Hyperspectral sensor: Discriminates many narrow spectral bands across the electromagnetic spectrum.
Multispectral sensor: Discriminates a few relatively broad spectral bands across the electromagnetic spectrum.
Multi-temporal imagery: Multiple images of the same location acquired on different dates.
Passive sensor: Measures energy emitted or reflected by the earth's surface without emitting radiation.
Remote sensing: Methods of detecting the electromagnetic radiation coming from the earth's surface via aircraft or satellite sensors (Campbell & Wynne, 2011;Turner et al., 2003).

Spatial resolution:
Basic unit of captured information that corresponds to pixel or grain size and determines the minimum spatial scale at which variation can be observed. Categories of spatial resolution in this paper follow Corbane et al. (2015): very high resolution <3 m; high resolution 3-29 m; medium resolution 30-300 m; and low resolution >300 m.

Spectral resolution: Width (and thus number) of bands into
which the electromagnetic spectrum is divided. concept for the continuous and complex nature of rarity. While we will discuss the capacity for RS to feed into and improve SDMs of rare plants, a comparative evaluation of the performance of different modelling techniques is beyond the scope of this synthesis and has been addressed in the literature (e.g. Elith & Burgman, 2002;Williams et al., 2009;Wiser et al., 1998). We will discuss the suitability of predictive performance measures for rare plant modelling studies, as well as the potential influence of rarity types on RS effectiveness and the detection of target species in the field (hereafter "practical utility").

| ME THODS
An extensive literature review was conducted to synthesize the use of RS to detect or predict rare plant distributions. Although the term "remote sensing" is defined precisely in the literature, the concept of "remote sensing variables or predictors" remains ambiguous.
Henceforth, RS predictors refer to i) continuous spectral information obtained from aircraft or satellite sensors, either as raw spectral bands or as indices; ii) landcover products developed from the classification of spectral information; and iii) digital elevation models ("DEMs") developed from satellite or airborne sensor information as well as derived topographic indices. In cases where the information on the origin or generation process of the predictors were not provided nor accessible, they were considered as non-RS predictors, except for DEM-derived topographic indices. DEMs are commonly generated through RS techniques and their direct survey is rare; therefore, when the DEM source is other than RS, it is usually stated in the literature (e.g. Padalia et al., 2010;Sperduto & Congalton, 1996).
A literature review of peer-reviewed articles was carried out using the search engine Scopus by combining terms related to plant with keywords related to RS and rarity or species at risk for the period 1990-2020. Studies targeting species at risk were included since they can be considered rare according to Rabinowitz's rarity classification (Rabinowitz, 1981). Specifically, the search was carried out using the following combination of keywords: (plant OR tree OR bryophyte OR moss) AND (rare OR endemic OR "at risk" OR endangered OR threatened OR red-list) AND ("remote sensing" OR "remotely sensed" OR sensor OR satellite OR drone OR "unmanned aerial vehicle" OR spectral OR lidar OR radar OR airborne OR aircraft). A total of 1,112 articles matched our search criteria. These articles were reviewed individually to identify and keep only those relevant for our topic, that is, articles using RS data for the purpose of detecting or modelling and predicting the presence of rare plant species. We excluded (i) articles that model richness distribution patterns of rare plants as they do not evaluate the capacity for RS to predict individual species, nor the influence of species' characteristics on RS performance, and (ii) those studies performed at low spatial resolutions (>300 m) since SDMs at these pixel sizes hinder the identification of local environmental factors driving species occurrence patterns (Engler et al., 2004), can result in large areas predicted as suitable habitat of limited practical use to detect rare species through field surveys (Guisan et al., 2006), and normally provide less accurate predictions for sessile species (Guisan & Thuiller, 2005). A total of 43 articles were selected for the development of this study.
These articles were first classified by RS approach (direct or indirect) and then by spatial resolution used (Table 1).

| REMOTE S EN S ING D IREC T APPROACH-DE TEC TION OF R ARE PL ANTS
The direct detection of rare plants and their traits through RS requires previous knowledge of a species' ecology and distribution, as well as the use of high spatial resolution imagery. Despite these constrains, 19 articles following this RS approach were found (Table 1; Figure 1 The utility of high and very high spatial resolution satellite sensors for direct detection of rare plants has also been highlighted. Omer et al. (2015) detected with high accuracy 5 of the 6 target endangered tree species in Dukuduku forest in South Africa using WorldView-2 satellite spectral imagery at 2 m spatial resolution.
Similarly, the use of 5 m resolution SPOT imagery allowed to map the endangered and endemic alpine tree Larix chinensis Beissn.
(Shaanxi larch) on Mount Taibai in China . Other studies have also tested the combined use of RS data from passive and active sensors for direct detection purposes. This combination provides a powerful approach, since active sensors, which allow the assessment of rare plant structural properties,    The RS direct approach not only offers the possibility to detect and map rare plants but also to assess their status (e.g. water stress, health; Chávez et al., 2013Chávez et al., , 2016Murfitt et al., 2016). This ability may allow the implementation of monitoring systems for these species, which can provide valuable additional information for management and conservation purposes. The studies reviewed here well exemplify the potential of a direct RS approach not only to detect rare plants, but also to monitor them in space and time (Landenberger et al., 2003;McGraw et al., 1998), or even to discover new populations (Fletcher & Erskine, 2012).

| REMOTE S EN S ING IND IREC T APPROACH-PREDIC TION OF R ARE PL ANT DIS TRIBUTIONS
The indirect RS approach allows the prediction of rare plants under environmental conditions where their direct detection is not possible (Levin et al., 2007). Most of the studies included in this section were performed in the Northern Hemisphere, while only three were conducted in the Southern Hemisphere ( Figure 1).

RS has been used to spatially characterize different biophysical
conditions at multiple spatial, spectral and temporal resolutions related to topography, vegetation, structure, climate, soil, geology, moisture, bathymetry and water transparency, as well as to anthropogenic and natural disturbances (Table 1). RS information has been acquired primarily from passive satellite sensors, although active satellite sensors, and airborne sensors both active and passive, have also been used.
Only three studies have used very high-resolution RS to model the distribution of rare plants, being limited exclusively to LiDARderived topographic predictors, NDVI and hyperspectral data.  Table 1 (shrubby reed-mustard) was successfully mapped (AUC = 0.85) by combining a wide variety of RS predictors, including topographic variables, spectral bands (and ratios), as well as vegetation, wetness and soil indices (Baker et al., 2016).
Rare plant studies using medium resolution RS are more common and have used a much wider diversity of RS predictors than those developed at high and very high resolutions (Table 1).
The variety of RS predictors that can be successfully integrated into rare plant models at this resolution was exemplified by Zimmermann et al. (2007). The authors modelled 19 tree species distributions ranging from rare to common and found that models combining RS and non-RS predictors consistently provided better performance for all species, and more so for rare species.
Medium resolution RS-only SDMs are also very useful for predicting rare plant occurrences. Suitable habitats for the endemic tree Adinandra griffithii Dyer were accurately predicted (AUC = 0.99) by using EVI time series (Adhikari et al., 2018). Similarly, robust predictions were achieved for the narrow-range endemic spe- Robust models were also developed for the rare sandstone shrubs Melaleuca triumphalis Craven and Stenostegia congesta A.R. Bean using a RS radiometric map representing thorium, uranium and potassium in combination with two and one topographic predictors, respectively (Crase et al., 2006). Likewise, the usefulness of the indirect RS approach to characterize aquatic habitats of rare plants has been demonstrated by several studies (Borfecchia et al., 2019;Traganos & Reinartz, 2018;Zucchetta et al., 2016). While all these studies focused on the same species, the endemic plant P. oceanica, they exemplify the variety of RS predictors that can be employed for predictive mapping purposes in aquatic environments (Table 1).
Overall, RS has provided valuable information on rare plant niches with good predictability at high and medium resolution.
These results highlight the potential of RS to not only characterize the habitats of rare plants but also to monitor them spatially and temporally (Bartel & Sexton, 2009;Neumann et al., 2015).
Several authors have also demonstrated the practical utility of predictive models built partially (Buechling & Tobalske, 2011;Sperduto & Congalton, 1996;Williams et al., 2009) or completely (e.g. Hernández-Lambraño et al., 2020Lauver & Whistler, 1993;de Queiroz et al., 2012) with RS predictors at those resolutions by discovering previously unknown populations of rare plants (Table 1). While SDM-based predictions are valuable tools to guide the search for rare plants in the field, the integration of abundance estimates derived from species abundance models ("SAMs") could further facilitate their detection. Likewise, when probability of occurrence estimated from SDMs and predicted abundance are uncorrelated and determined by different sets of predictors, SAMs can provide valuable additional information on habitat quality or ecological species preferences (Duff et al., 2012). Since RS also has the ability to feed SAMs (e.g. Arenas-Castro et al., 2019;Duff et al., 2012;Guarino et al., 2012), abundance estimates can also be obtained at high or medium resolutions. Therefore, the combination of both RS-based SDM and SAM model types may represent a new and strong practical approach for detection of rare plants, by guiding field search efforts towards predicted habitats where higher plant abundance makes them more detectable.

| Considerations of predictive performance measures for rare plants
Currently, there is still no consensus on which are the most suitable metrics to evaluate the predictive performance of SDMs, which has the use of multiple metrics as the best solution (Amini Tehrani et al., 2020;Breiner et al., 2015). However, since each accuracy metric provides a type of information, the choice should ideally be based on their intended use (Fielding & Bell, 1997) rather than on the arbitrary selection of different metrics. As rare plants typically show low prevalence (i.e. high absences/presences ratio), overall predictive performance metrics (e.g. accuracy or AUC) can lead to overly optimistic results about model accuracy (Buechling & Tobalske, 2011;Lobo et al., 2008). Furthermore, those metrics are not sensitive to overprediction, which can be common for rare plant presence. Based on these drawbacks, we propose the use of two complementary metrics to evaluate in isolation the ability of the model to predict presences of rare plants, namely sensitivity and precision. Sensitivity is the proportion of true positives correctly predicted, while precision is the proportion of positive predictions corresponding to true positives (Fawcett, 2004). The use of both metrics provides information on the proportion of actual presences correctly predicted and possible instances of overestimation. Therefore, sensitivity and precision are ultimately metrics indicative of the practical utility of models to find new localities of rare plants.

| REMOTE S EN S ING BA S ED ON THE CHAR AC TERIS TIC S OF R ARE PL ANTS
Rare plants, like all plant species, have distinct features that allow their differentiation and identification, but is it possible to capture some of the distinguishing features of rare plants through RS? Can these features influence the performance of RS to detect or predict rare plants? In this section, rare plant features related to morphology, phenology, physiology, and ecological niche are discussed. Since rare and common plants are not categorized as such based on the plant features presented, we are aware that some aspects of our discussion may also apply to common plants.

| Morphology
Morphological features of rare plants can provide decisive information for their direct detection. Two conditions must be met to ensure the success of this approach (also applicable to common plants): i) very high spatial resolution is required to capture morphological characters considered important, and ii) the date on which RS imagery is taken must correspond to a time when the target plant exhibits distinctive morphological characters that allow its discrimination. The direct detection of the rare shrub Boronia deanei based on its pink flowers and growth form exemplifies these criteria (Fletcher & Erskine, 2012).

| Phenology
Multi-temporal RS imagery at high temporal resolution has the potential to capture phenological traits of rare plants, such as flowering, fruiting or leaf growth/fall (Campbell & Wynne, 2011;Turner et al., 2003). This phenological information can be advantageous for both RS approaches. Direct detection can benefit from phenological processes as long as the conditions mentioned in the previous subsection are met. However, multi-temporal imagery can only provide useful information when morphologically and phenologically similar target and cohabitating species display these features at different times. By contrast, single-date images provide similar information if the detection date captures the uniqueness of morphological characters derived from phenological features.
In the indirect approach, the spectral radiation associated with phenological features of rare plants can directly influence the information captured from remote sensors during the characterization process of their ecological niche, which is subsequently used to model rare plant occurrence. This fact has been defined as a source of unintentional bias when predicting potentially suitable habitats of plants, since the captured information is associated with their actual distribution (Bradley et al., 2012). However, this type of bias can be considered advantageous when RS-based predictions are used to locate actual rare plant occurrences. For instance, predictions of rare trees improved with the inclusion of multi-temporal predictors whose spectral information was directly influenced by their leaf phenological features (Zimmermann et al., 2007). Similarly, the detection of leaf phenological changes in the Watarase wetland allowed to accurately predict the occurrence of two of four rare plants studied (Ishihama et al., 2010).
The authors highlighted that one of these species, Ophioglossum namegatae Nish. & Kurita, because of its sprouting period (early spring) and rapid growth, could directly contribute to the spectral information captured in early May, which was one of the most important predictors for both species. Likewise, the flowering phenological stage of the endemic tree A. griffithii played an important role in predicting its distribution, since the EVI for the periods of June and July were the most influential predictors (Adhikari et al., 2018).

| Physiology
Plant spectral information is influenced by plant physiological traits such as concentration and distribution of biochemical components (Peñuelas & Filella, 1998), which can be identifiable and quantifiable based on their spectral absorption features (Asner & Vitousek, 2005;Blackburn, 1998;Sims & Gamon, 2003). The use of multispectral bands and physiological indices at high spatial resolution has been shown to significatively contribute to the detection of rare tree species (Liu et al., 2018;Omer et al., 2015). However, the detection of rare plants could also benefit from hyperspectral bands and LiDAR sensors, which have the capacity to assess plant physiological traits in more detail (Andrew & Ustin, 2006;Asner, Jones, et al., 2008;Ustin & Gamon, 2010).

| Ecological niche
The vertical position occupied by rare plants in their respective habitats (e.g. overstory or understory) influences their direct detection by RS. Active airborne sensors (e.g. LiDAR) are required to detect subcanopy plants Hernandez-Santin et al., 2019). On the other hand, the effectiveness of the RS indirect approach in characterizing species' ecological niches depends on two conditions. First, sensor spatial resolution must be adapted to habitat size. This is especially important for rare plants that are associated with small habitat patches (e.g. de Queiroz et al., 2012), which can remain indistinguishable if they are smaller than the RS imagery pixel (Luoto et al., 2002). Secondly, habitat specificity of rare plants has been shown to influence prediction accuracy (Buechling & Tobalske, 2011;Parviainen et al., 2013). Prediction accuracy of rare plants can increase when their habitat specialization increases (Hernandez et al., 2006), making habitats where they occur more distinct than habitats where they are absent. For rare plants with wider habitat specificity and few occurrences, insufficient prior knowledge seems to be the main limiting factor rather than capacity of RS to successfully discriminate between suitable and unsuitable habitats.

| REL ATING R ARIT Y FORMS WITH MODEL PREDIC TIVE PERFORMAN CE
The rare species classification developed by Rabinowitz (1981) (Fischer & Stöcklin, 1997;Matthies et al., 2004;Ouborg, 1993;Thomas, 1994). This can result in lower proportion of occupied suitable sites or "fidelity," which can lead to an increase in false positives and thus decrease model predictive accuracy. In contrast, species with larger populations are more resistant to stochastic events and therefore more stable over time, which would increase their probability of presence and fidelity.
Summarizing the previous ideas regarding the influence of Rabinowitz's rarity classification criteria on model predictive performance, predictive performance can be negatively influenced by geographic distribution range and positively influenced by habitat specificity and local population size ( Figure 2). Nonetheless, predicting widely distributed rare plants throughout their entire distribution range is not practical for their detection; thus, the effect of the geographic distribution range on model performance becomes negligible when prioritizing smaller areas for predictions. At this point and using the same classification as Rabinowitz (1981), better results in terms of predictive performance can be expected for predictable and endemic rarity forms showing large local population size ( Figure 2). However, because the assumption initially raised in this section has been only theoretically addressed, the veracity of these final conjectures must be tested empirically. to the rarity classification criteria, but this requires empirical testing.

| CON CLUS ION
In conclusion, RS is a powerful information source to generate predictions and guide the discovery of new rare plant populations.
New rare plant occurrences can subsequently be used as inputs for improving predictive models, to acquire better knowledge on ecological requirements and restrictions of species to help understand F I G U R E 2 Classification of rarity forms showing the potential changes in model predictive performance (solid arrows), based on the three criteria used for such classification: geographic distribution range, habitat specificity and local population size. The arrowhead indicates the direction of improvement (+) in predictive performance. Adapted from Rabinowitz (1981) the causes of their rarity, and to review and update when necessary their conservation status. With this synthesis, we have highlighted the strong potential of RS for the purposes of detection and prediction of rare plants, with practical applications for rare species conservation and management.

ACK N OWLED G EM ENTS
We greatly thank Sandra Fernández Fariñas and Milka for their valuable comments and support in the initial phases of this study. We thank Annie Desrochers for reviewing the first version of the manuscript. Also, we are very grateful for the constructive comments and suggestions provided by the anonymous reviewers, which greatly helped to improve the quality of our manuscript. We also thank Environmental Damages Fund, Environment and Climate Change Canada for funding this research.

PEER R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/ddi.13244.

DATA AVA I L A B I L I T Y S TAT E M E N T
Data sharing is not applicable to this article as no new data were created or analysed in this study.