Integrating Hyperspectral Remote Sensing, Models and In-situ Observations with Data Fusion to Estimate Nutrient Loadings in the Shenandoah River Basin

Abstract

Estimating and predicting water variables such as chlorophyll, colored dissolved organic matter, turbidity, phosphorus and nitrogen are of paramount importance due to their strong influence on water resource quality. This study focuses on the Shenandoah River which has gained much attention as it flows into the Potomac River, has been overly polluted in the past and contains many contaminated sites and classified by the Virginia Department of Environmental Quality as impaired. This research is aimed to demonstrate the feasibility of combining remotely sensed water quality observations with water quality modeling using data fusion techniques for an efficient and effective monitoring of water quality in the Shenandoah River. This study explores the hypothesizes that: Sensitivity and uncertainty from water quality remote sensing and water quality modeling can be improved through data fusion for a better prediction of water quality in the Shenandoah River Basin. To validate this hypothesis, we use a three-step approach; First, we investigate the spatial dynamics of water quality across the Shenandoah River basin with hyperspectral remote sensing and chemometrics to estimate chlorophyll-a (Chl), and colored dissolved organic matter (CDOM) using three band combinations and nutrients (total nitrogen and total phosphorous) in the Shenandoah River. This approach has been adopted because wavelengths with the blue-green and green algae peak reflectance are close together and make their differentiation more difficult. It has also been demonstrated that hyperspectral imagers allow for improved detection of chlorophyll and hence algae, as a result of acquired narrow spectral bands between 450 nm and 600 nm (Kirk, 1994; Maffione and Dana, 1997; Lee et al., 2002). Spectroscopic and field samples collected during spring of 2013 was used to perform chemometrics analysis, and results show that total phosphorous; nitrogen and turbidity can be predicted between 450 nm to 555nm and 670 nm to 710 nm, the range of wavelengths that indicated better predictability for spectroscopic analysis. Further, using Airborne Real-time Cueing Hyperspectral Enhanced Reconnaissance (ARCHER) imagery; partial least squares analysis was carried out for the 504 nm to 1000 nm spectral range, with up to 18 samples for each sample field location in the summer of 2014. The coefficients derived from chemometrics applied to the ARCHER data to map and predict nutrients show that if residual surface reflectance are well accounted for, airborne hyperspectral data can be very effective in estimating water quality. Cross-validation also showed that ARCHER retrieval method has transferability ability to other locations, which significantly makes a prediction for areas that do not have data relatively easy. The second step uses the Environmental Protection Agency (EPA), Water Quality Analysis Simulation Program (WASP) to simulate Shenandoah Basin water quality and to conduct an uncertainty analysis of the state variables and uses available data from monitoring programs to validate the model. The purpose was to simulate nutrients, dissolve oxygen and chlorophyll-a dynamics, and examine the influence that changes in the parameters that affect these variables would have on water quality in the river. The results show that greater complexity and increased model sensitivity decreased the error in the output simulations with model predicted dissolved oxygen values tending to be close to the measured data while total nitrogen and phosphorus tended to be overestimated or overestimated. The third step used a Kalman Filter (KF) to merge the results of the first two stages, with the goal of determining an improved water quality estimate. We assess the accuracy of our model to find out if the Kalman filter merging improved our estimation; the analysis was verified and validated by comparing the merged data to the field measurement to evaluate the degree to which the estimation has improved the predicted variable of interest. Our merged and lab analysis cross validation showed some improvement in the results with very high coefficients of determination for some variables. Field observations are used as explanatory variables to evaluate the sensitivity of the analysis to observations, in an attempt to examine how changes in field observation affect final KF analysis. The statistical confidence test was also conducted, and results show that with sufficient field observation, a better fusion result will be obtained. This rersearch is a first delivery attempt to use finer resolution ARCHER airbone data to provide an assessment of water quality retrieval with an understanding of the relationship that exist between the distinctive kinds of geographic setting that characterize the spatial variability of in the water quality parameters.