Date of Award

Fall 2026

Abstract

Stormwater runoff conveys nutrients, sediments, and pollutants to receiving waters. Conventional monitoring produces temporally sparse observations that introduce substantial uncertainty in pollutant load estimates. Ultraviolet-visible (UV-Vis) spectrophotometry offers a continuous, high-resolution alternative but lacks a standardized calibration framework, formal comparison with conventional protocols, and field-validated operational guidance for stormwater applications. This dissertation addresses these three gaps.The first chapter developed and compared multivariate regression models (principal component regression, partial least squares regression, and artificial neural network) for predicting the concentrations of total nitrogen, nitrate, total Kjeldahl nitrogen, total phosphorus, reactive orthophosphate, and total suspended solids from UV-Vis absorbance spectra. The calibration dataset combined laboratory-prepared standards spanning the full analytical range of Environmental Protection Agency approved reference methods with field-collected stormwater samples fortified at randomized concentration levels, with chloride and turbidity included as controlled interferents. The artificial neural network employed a two-hidden-layer feedforward architecture with hyperbolic tangent activation nodes in the first layer to capture nonlinear spectral relationships, distinguishing it from the strictly linear principal component and partial least squares regressions. Model performance was evaluated using ten-fold cross-validation with the Olkin-Pratt estimator adjusted coefficient of determination (R_adj^2) and root mean square error of prediction (RMSEP). All three models produced strong predictions for nitrogen species, with R_adj^2 between 0.88 – 0.99 and RMSEP values between 0.11 and 0.32 milligrams per liter with the linear, TKN models having the lowest performance. For phosphorus species, linear methods yielded moderate prediction metrics with R_adj^2 below 0.62 and RMSEP about 0.80 milligrams per liter, but the artificial neural networks achieved R_adj^2 exceeding 0.99 and RMSEP below 0.07 milligrams per liter which was a 63% improvement in R_adj^2 and 92% improvement in RMSEP. For total suspended solids, artificial neural networks reduced the prediction error by 38% relative to the lower PLSR linear method, from approximately 63 to 39 milligrams per liter. The nonlinear flexibility of artificial neural networks was critical for resolving complex spectral interactions among target analytes, particularly phosphorus, turbidity-induced scattering, and dissolved interferents. The second chapter evaluated whether conventional sampling protocols produce event mean concentration estimates statistically equivalent to high-resolution UV-Vis derived continuous pollutographs. The benchmark pollutographs for comparison were obtained from one-minute resolution monitored stormwater control measures across New England at three different sites. Measured UV-Vis absorbance spectra were converted to concentrations of total nitrogen, nitrate, total Kjeldahl nitrogen, total phosphorus, reactive orthophosphate, and total suspended solids using an artificial neural network model, from which flow-weighted event mean concentrations were calculated as the benchmarks. Grab sampling, time-paced autosampling, and flow-paced autosampling were each simulated and evaluated with arithmetic and flow-weighted event mean concentration calculations using Monte Carlo resampling and exhaustive enumeration across 28 storm event coverage levels ranging from 0.1% to 100%. Statistical equivalence was assessed using a two one-sided test (TOST) procedure with a 10% margin of the log error ratios of the simulated EMC values to the benchmark EMCs. Flow-paced autosampling achieved equivalence at the minimum coverage evaluated for all six pollutants. Flow-weighted grab and time-paced protocols achieved equivalence at coverage thresholds of 0.3% to 2.0%, depending on the pollutant. Arithmetic event mean concentration calculation failed to achieve equivalence at any coverage level for any pollutant, producing a persistent systematic bias that did not diminish with increasing sampling intensity. Arithmetic protocols underestimated nitrogen by approximately 15% and overestimated phosphorus and total suspended solids by approximately 25% and 35%, respectively. Root mean square error analysis demonstrated that practically defensible coverage thresholds were approximately 5% to 10% for most pollutants under flow-weighted protocols, corresponding to 38 to 76 samples for a median storm event duration of 12.7 hours. The third chapter developed field-validated operational guidance for UV-Vis deployment in stormwater environments by integrating controlled laboratory evaluations with multi-site, multi-season field experience. A multifactorial analysis of variance was used to evaluate the effects of ambient light and hydraulic conditions on spectral measurement sensitivity, finding that the mean predicted concentrations varied within less than 3% of the grand mean across all tested conditions. Repeated measurements over a three-month interval showed no operationally meaningful instrumental drift. Power budget analysis demonstrated that a UV-Vis system with an automated cleaning brush achieved battery runtimes 10 to 17 times longer than a portable autosampler on an equivalent battery capacity. Lifecycle cost analysis across nine monitoring scenarios showed that the UV-Vis system reduced annualized monitoring costs by factors of approximately 5 when compared with a composite-bottle autosampler program and a factor of 69 when compared with a 24-bottle autosampler program. On a cost-per-parameter-measurement basis at a five-minute measurement interval, the UV-Vis typical scenario cost $0.06 per parameter-measurement, compared with $2.20 for the composite autosampler and $32.53 for the 24-bottle autosampler, representing reductions of approximately 38-fold and 566-fold, respectively. Collectively, these three studies establish that UV-Vis spectrophotometry, when calibrated with appropriate multivariate regression models, provides a technically defensible, operationally practical, and economically advantageous alternative to conventional discrete sampling for stormwater pollutant monitoring.

Document Type

Dissertation

First Advisor

Thomas Ballestero

Second Advisor

James Houle

Third Advisor

Anne Lightbody

Department or Program

Civil and Environmental Engineering

Degree Name

Doctor of Philosophy

Download

Share

COinS