By short-cutting full community Floristic Quality Assessment, indicator species (Dufrêne and Legendre 1997; Siddig et al. 2016) could drastically accelerate field identification of high floristic quality and help to screen or verify potential reference-quality wetlands, including verification of predictions made remotely (e.g., Host et al. 2005). Unfortunately, the present results do not support using indicator species to rapidly discriminate high floristic quality wetlands in the US southern plains. In the same study area Bried et al. (2014) reported plant indicator species of potential reference-quality wetlands defined by multimetric vegetation criteria, including the floristic quality benchmark (≥ 20.0 FQI) used here. They reported some encouraging indicator values and false positive predictions but lacked a validation analysis (i.e. comparison with observed misclassification rates).
Several challenges may limit the performance of floristic quality indicator species. First and foremost, the species are being asked to indicate a level of quality derived from community data. We assumed species combinations (De Cáceres et al. 2012) would mitigate this disproportionality, but few species pairs and only one triplet combination met any specificity and sensitivity thresholds. Secondly, indicator species are always fixed to a target, but floristic quality is a “moving target” because theoretically many species compositions can result in the same level of quality. It seems indicator species would need strong associations or positive co-occurrence patterns with a large fraction of the assemblage to adequately represent floristic quality. Increased ecogeographic and environmental (e.g. hydrogeomorphic) stratification may reduce compositional variation and strengthen indicator performance (Brooks et al. 2006; Bried et al. 2019), but effective stratification may be difficult. In our study area wetlands are hard to geomorphically subclassify (Dvorett et al. 2012) and subject to strong climatic gradients and other natural heterogeneity (Hoagland 2000; Gallaway et al. 2019), adding to the challenges.
One candidate (Eleocharis compressa) appeared in every scenario and consistently showed among the highest indicator values. This species also possesses a high C value (6) relative to the low-biased C distribution across the dataset (Fig. 4). In western Oklahoma E. compressa can dominate in small depressions (often interdunal swales) on clay soils (Hoagland 2002), suggesting it may occur commonly enough for practice. Some candidates selectively outperformed E. compressa but most of the final pool struggled on one or both sides of validity. We used the most conservative false positive prediction (1 – Specificity UCL) when measured against observed rates, in part to counterbalance the more relaxed true positive measure (Specificity LCL). Even if we relaxed the false positive prediction (to 1 – Specificity) many of these indicators still would not pass the test, or they would run an unacceptable 20–30% risk of erroneously indicating a site as high quality.
More sample stratification may alleviate misclassifications while strengthening indicator values (Dufrêne and Legendre 1997, Bried et al. 2019). Overall, indicator value and validity improved after filtering the samples to a specific prairie ecoregion and geomorphic setting. Deeper stratification by vegetation types and hydrogeomorphic subclasses (e.g. Dvorett et al. 2012) could further strengthen indicators but requires adequate knowledge and acceptance of tradeoffs between accuracy and precision. Too much stratification may leave some strata poorly defined and lacking in indicator species or the data needed to extract and validate them.
Alternatives to traditional indicator value analysis (Dufrêne and Legendre 1997) might lead to improvements. Stapanian et al. (2013), for example, used classification and regression tree models to find indicator species of wetland vegetation quality in Ohio, USA. Their simplest model containing just two species predicted high-quality wetlands with 13% overall misclassification rate.
We tested species combinations (De Cáceres et al. 2012) but few combinations appeared in the candidate pool. Perhaps combining the single-species indicators post hoc would help mutually offset their deficiencies. For example, pairing E. compressa with Juncus torreyi, both relatively conservative (C = 6), may reduce misclassification error in the narrower spatial-environmental context (CGP depressions). Likewise joining E. compressa with Amorpha fruticosa, Leersia oryzoides, and/or Schoenoplectus pungens may strengthen performance in the broader context of prairie ecoregion wetlands. These latter species also have the advantage of being definitively recognizable throughout the growing season, unlike E. compressa which flowers early to mid-season and can look similar vegetatively to other Eleocharis species in the region (E. albida, E. montevidensis, E. tenuis). The potential drawback is whether such combinations will sufficiently occur in the target area, a problem mitigated ad hoc by preset sensitivity thresholds in the analysis of combinations (De Cáceres et al. 2012). Indeed, E. compressa and J. torreyi co-occurred in only 7.4% of our study sites, consistent with Hoagland (2002) who did not detect J. torreyi where E. compressa was most abundant. Similarly, E. compressa co-occurred with A. fruticosa, L. oryzoides, and S. pungens in 4.6%, 6.5%, and 7.4% of sites, respectively, and there was only one site where all four of these species co-occurred.
The present findings do not preclude the existence of high floristic quality indicators in other regions, especially where wetlands are well classified and training samples are given sufficient spatial-environmental context. Nor do our results negate potential for indicators in other ecosystems where Floristic Quality Assessment is commonly applied, namely forests and prairies. However, if wetlands tend to be less floristically diverse than forests and prairies (on average within climate zones) finding strong indicator species in those systems would seem even less likely for the reasons discussed above. Our study also cannot rule out potential for indicators at other levels of floristic quality. Indicators of low quality could be useful in avoiding a costly permitting process (Stapanian et al. 2013) or futile conservation investment, i.e. a site not worth protecting or restoring (a site “beyond repair”). Indicators of moderate quality could help direct protection and management effort to where there is both need and worth. Before broadly concluding that floristic quality indicator species do not exist for wetlands, we recommend exploring indicator potential in other regions and at other quality levels, and perhaps trying other statistical approaches (e.g. Stapanian et al. 2013).