Classifying the biological condition of small streams: an example using benthic macroinvertebrates

Abstract

Abstract. The ability to classify the biological condition of unsurveyed streams accurately would be an asset to the conservation and management of streams. We compared the ability of 5 modeling methods (classification and regression trees, conditional inference trees, random forests [RF], conditional random forests [cRF], and ordinal logistic regression) to predict stream biological condition (very poor, poor, fair, or good) based on benthic macroinvertebrate Index of Biotic Integrity data taken from the Maryland Biological Stream Survey. Predictor variables included land use and land cover (e.g., impervious surface, row-crop agriculture, and population density) and landscape measures (annual precipitation and watershed area). We included 1561 sites on small nontidal streams in the Maryland portion of the Chesapeake Bay watershed. We used 1248 sites (80%) as a training data set to build models and 313 sites (20%) as an independent evaluation data set. RF and cRF models most accurately predicted observed integrity scores in the evaluation data set, but we selected the cRF as the best model because of weaknesses in the RF model (e.g., biased variable selection). Percent impervious surface was the most important variable in the cRF model, and the probability that a site was in very poor or poor biological condition increased rapidly as % impervious cover increased up to 20%. When applied to predict stream biological conditions in all 7908 small nontidal stream reaches in the study area, the cRF model predicted that 33.8% were in fair, 29.9% in good, 22.7% in poor, and 13.6% in very poor biological condition. Our analyses can be used to manage and conserve freshwater and estuarine resources of Maryland and the Chesapeake Bay watershed. Model predictions for unsurveyed streams can help target field studies to identify high-quality streams deserving of conservation and impaired streams in need of restoration.Key words: stream condition, landscape-scale, land use, random forests, conditional inference, classification and regression trees (CART), ordinal logistic regression, prediction.Biological assessments have shown that many small streams have been impaired by anthropogenic stressors (Benke 1990, USEPA 2000. These results have generated much interest in conserving and managing these streams and their watersheds. However, small streams are numerous (Leopold et al. 1964), and assessing the biological condition of all small streams is logistically impractical and cost prohibitive. Models that reliably predict biological conditions at unsurveyed locations are needed. One way to estimate regional biological conditions in streams and rivers is to extrapolate the observed proportions of impairment classifications from a sample of streams to all streams in a landscape (e.g., if 30% of the sampled sites are impaired, then 30% of the sites in the region are impaired; USEPA 2006 Classification and regression trees (Breiman 1984, De'ath and Fabricius 2000, Loh 2008 are statistical techniques that might be useful for constructi...

Geoadditive regression modeling of stream biological condition

“…Boosting predictions of FIBI and BIBI indicators are similar to predictions obtained from other established statistical techniques (see, e.g., Maloney et al 2009), but spatial covariate patterns are detected and model estimates can be interpreted in a more meaningful way. This is possible because the structure of the proportional odds model allows for inspection and visualization of marginal predictor effects.…”

“…Following Maloney et al (2009), we used an ordinal scale to quantify BIBI and FIBI indicators (1 = "very poor site", 2 = "poor site", 3 = "fair site", 4 = "good site"). FIBI and BIBI indicators were regressed on sitespecific predictor variables using the P/O boosting algorithm introduced in Section 2.…”

“…Random forests have been shown to be one of the most accurate methods for predicting IBI indicators (Maloney et al 2009). Because the random forest method neither requires the proportional odds assumption nor the additivity of the prediction function specified in (3), it is less restrictive than P/O boosting.…”

“…Policy makers and land managers must assess the ecological effects of their decisions on streams, but also have to investigate the impacts of stream degradation on human health and the quality of life. For these reasons, a detailed understanding of the relationship between anthropogenic stressors and stream ecosystems is essential (Cushing and Allan 2001;USEPA 2006;Maloney et al 2009). To develop that understanding, ecologists and statisticians need to quantify how watershed characteristics affect stream biological condition.…”

“…Ideally, those models would both quantify the effects of anthropogenic stressors on stream condition and accurately predict biological condition at unsurveyed locations (USEPA 2006). In recent years, a wide range of statistical tools to characterize and to model the condition of small streams have been developed (see, e.g., Barker et al 2006;Collier 2009;Maloney et al 2009, or Cooper 2009 for recent studies in this field).…”