Stream hydrological and ecological responses to climate change assessed with an artificial neural network

An artificial neural network (ANN) was used to evaluate the hydrological responses of two streams in the northeastern U.S. having different hydroclimatologies (rainfall and snow+rain) to hypothetical changes in precipitation and thermal regimes associated with climate change. For each stream, historic precipitation and temperature data were used as input to an ANN, which generated a synthetic daily hydrograph with high goodness-of-fit (r2 > 0.80). Four scenarios of climate change were used to evaluate stream responses to climate change: + 25% precipitation, -25% precipitation, 2 x the coefficient of variation in precipitation regime, and +3”C average temperature. Responses were expressed in hydrological terms of ecological relevance, including flow variability, baseflow conditions, and frequency and predictability of floods. Increased average precipitation induced elevated runoff and more frequent high flow events, while decreased precipitation had the opposite effect. Elevated temperature reduced average runoff. Doubled precipitation variability had a large effect on many variables, including average runoff, variability of flow, flooding frequency, and baseflow stability. In general, the rainfall-dominated stream exhibited greater relative response to climate change scenarios than did the snowmelt stream. Stream ecosystems are at risk for changes due to climate change because ecological processes are strongly influenced by seasonal patterns of precipitation, runoff, and temperature (Carpenter et al. 1992; Allan 1995). If historical hydrological and thermal regimes in streams are modified by anthropogenically altered climate change, then ecosystem alteration is to be expected. Hydrological modifications may result either from changes in average conditions or from changes in the distribution and timing of extreme events such as floods and droughts. Evaluating the extent to which stream hydrographs are modified by scenarios of climate change can provide important information on the relative sensitivity of stream ecosystems to potential climate change. Modeling stream hydrological response to climate variation can be performed with a variety of techniques. If detailed watershed and climate data are available for parameterization, one can use mass balance models, such as hydrological budget models (e.g. Gleick 1987). However, for many stream systems, detailed watershed data are lacking, making the mechanistic modeling of hydrological response to climate difficult, Further, traditional empirical models (e.g. regression models) may not perAcknowledgments We thank P. Mulholland and C. P. Hawkins for constructive reviews of an earlier version of this paper. This research was funded in part by the U.S. Environmental Protection Agency through cooperative agreement CR8 16540010 (N. L. Poff and J. D. Allan). Completion of this project was also supported in part by a National Science Foundation grant (DEB 92-l 5019) to N. L. PoK form well due to structural constraints (e.g. linearity) and paucity of data. Neural network analysis is a recently developed modeling technique that may be useful in simulating hydrological response to climate change in basins with limited data. We used this technique in examining the hydrographic response of streams in Maryland and New York to various scenarios of climate change. The responses of several hydrological variables were examined, such as measures of flow variability, and timing and frequency of low and high flow extremes. That such hydrological descriptors can influence ecological processes and patterns in stream ecosystems has received attention (see Minshall 1988; Poff and Ward 1989, 1990) and some limited empirical support (e.g. Horwitz 1978; Poff and