Date of Award

June 2019

Degree Type


Degree Name

Doctor of Philosophy (PhD)


Civil and Environmental Engineering


David Chandler


ANFIS vs SAC-SMA, development pattern, imperviousness distribution, real-time ensemble flood forecasting, seasonal runoff peak flow, urban hydrology

Subject Categories



Urban growth is a global phenomenon, and the associated impacts on hydrology from land development are expected to increase, especially in peri-urban catchments, which are newly developing catchments in proximity of growing cities. In northern climates, hydrologic response of peri-urban catchments change with the water budget and climatic conditions. As a result, runoff response of northern peri-urban catchments can vary immensely across seasons. During warm seasons, the evapotranspiration (ET) and infiltration rates are high, so urban floods are expected to occur during high intensity, low duration storm events. During cold seasons and below freezing temperatures, surficial soils are typically frozen and nearly impervious. In addition, the ET rate is low throughout winter. Therefore, the difference in runoff response between peri-urban and natural catchments is least in winter. Furthermore, winter snow redistribution by plowing and endogenous urban heat affect the snowmelt timing and frequency. Due to the limited availability of data on snow removal and redistribution activities in northern peri-urban catchments, cold-season hydrologic modeling for peri-urban catchments remains a challenging task in urban hydrology.

Research on the cold season hydrologic response of peri-urban catchments are mostly limited to Finland, Sweden, and Canada. The resulting research gap on seasonal change in hydrologic response of peri-urban catchments is common to many northern settings. In the first phase of this study, I use intensive discharge monitoring records at several peri-urban catchments near Syracuse, NY to calculate and compare seasonal runoff peak flows among several peri-urban catchments. These are selected to provide a range of drainage area and imperviousness to clarify the impact of urban development and catchment size on seasonal hydrologic behavior of peri-urban catchments.

It is well understood that greater peak flows and higher stream flashiness are associated with increased surface imperviousness and storm location. However, the effect of the distribution of impervious areas on runoff peak flow response and stream flashiness of peri-urban catchments has not been well studied. In the second phase of this dissertation, I define a new geometric index, Relative Nearness of Imperviousness to the Catchment Outlet (RNICO), to correlate imperviousness distribution of peri-urban catchments with runoff peak flows and stream flashiness. The study sites for this phase of the study include ninety peri-urban catchments in proximity of 9 large US cities: New York, NY (NYC), Syracuse, NY, Baltimore, MD, Portland, OR, Chicago, IL, Austin, TX, Houston, TX, San Francisco, CA, and Los Angeles, CA. Based on RNICO, all development patterns are divided into 3 classes: upstream, centralized, and downstream. Analysis results showed an obvious increase in runoff peak flows and decrease in time to peak as the centroid of imperviousness moves downstream. This indicates that RNICO is an effective tool for classifying urban development patterns and for macroscale understanding of the hydrologic behavior of small peri-urban catchments, despite the complexity of urban drainage systems. Results for nine cities show strong positive correlations between RNICO and runoff peak flows and stream flashiness index for small peri-urban catchments. However, the area threshold used to distinguish small and large catchments differs slightly by location. For example, for Chicago, IL, NYC, NY, Baltimore, MD, Houston, TX, and Austin, TX area threshold values of 55, 40, 50, 42, and 32 km2 emerged, runoff peak flows in catchments with drainage area below these values were positively correlated to RNCIO. This first phase of this study suggests that RNICO is a stronger predictor of runoff peak flow and stream-flow regime in humid northern and southern US study sites, compared to more arid western US study sites. This difference is likely due to the greater precipitation rates and greater antecedent soil moisture contents for humid climates. The extent of urban infrastructure is less likely to control the effectiveness of RNICO for predicting runoff peak flows and R-B flashiness index for the selected study sites, due to the relatively similar urban development level within the peri-urban study catchments.

Consistent forecast of peak flows across scales in flood hydrographs remains a challenge for most hydrologic models. Urbanization increases the magnitude and frequency of peak flows, often challenging the forecast ability for real-time flood prediction. Following advances in satellite and ground-based meteorological observations, global and continental real-time ensemble flood forecasting systems use a variety of physical hydrology models to predict urban peak flows. Artificial intelligence (AI) models provide an alternative approach to physical hydrology models for real-time flood forecasting. Despite recent advances in AI techniques for hydrologic prediction, ensemble stream-flow prediction by these methods has been limited. In addition, application of AI models for flood forecasting has been limited to large river basins, with very limited research on use of AI models for small peri-urban catchments. Flood forecasting in small urban catchments can be a critical task to urban safety due to the short time of concentration and quick precipitation runoff response. AI flood forecasting models typically apply upstream streamflow measurements to forecast downstream flood discharge. Therefore, the storm direction may change the flood travel time and time to peak, which challenges accurate flood forecasting. For example, if the storm direction is upstream through an AI model trained on the upstream gage data may fail to accurately predict peak flow magnitude and timing, at the outlet, this is due to the quicker runoff response of the downstream gage compared to the upstream station. There has been very limited focus on the impact of storm direction on peak flow response of urban catchments and available literature are limited to lab-scale prototypes and rainfall simulators. These may not fully represent real-world flooding scenarios. Therefore, the impact of storm direction on flood forecasting performance of peri-urban catchments is another important research gap in real-time urban flood forecasting.

In the third phase of my dissertation project, I initially assess the impact of storm direction on the flood forecasting performance of an Adaptive Neuro Fuzzy Inference System (ANFIS) at a peri-urban catchment in proximity of Syracuse, NY. Next, I compare the relative utility of physical hydrology and AI approaches to predict flood hydrograph in peri-urban catchments. For this comparison, I selected ANFIS, and Sacramento Soil Moisture Accounting Model (SAC-SMA) for real-time ensemble re-forecasting of streamflow in several small to medium size suburban catchments near NYC for Hurricane Irene and a smaller storm event. The SAC-SMA model is a physical hydrology model that was initially developed by Burnash et al. (1973). The National Oceanic and Atmospheric Administration (NOAA) selected the SAC-SMA lumped model as a comparison baseline for participating distributed hydrologic models in the Distributed Model Intercomparison Project (DMIP), which aimed to identify the most suitable model for National Weather Service (NWS) streamflow prediction across the US ( More importantly, the NWS is currently using the lumped form of SAC-SMA for ensemble flood forecasting across the US (Emerton et al., 2016). For these reasons, I chose to employ a lumped version of SAC-SMA in my dissertation project. SAC-SMA performed well for both large and small events and for lead times of three to 24 hours, but ANFIS predicted the Hurricane Irene flood discharge well only for short lead times in small study catchments. ANFIS had reasonable percent bias (PBIAS) for predicting the small storm event for all lead times, indicating the utility of ANFIS for small events. In addition, the accuracy of both SAC-SMA and ANFIS models for ensemble flood prediction did not change significantly with catchment size and imperviousness. Overall, results of the third phase of this study suggest that the lumped SAC-SMA model may be a reliable option for local urban flood forecasting for evacuation plan lead time up to 24 hours. Due to the uncertainties in future climatic conditions, my study emphasizes the importance of using physical hydrology models for real-time flood forecasting of large events in small urban catchments. This recommendation is based on the finding that the performance of data-driven models may greatly decrease with the storm scale if the training period includes storms of magnitude less than storms in the validation period.


Open Access

Included in

Engineering Commons