State Water Resources Research Institute Program
Project ID: 2009MT199B
Title: Student Fellowship: Bayesian Uncertainty and Sensitivity Analysis for Complex Environmental Models, with Applications in Watershed Management
Project Type: Research
Start Date: 3/01/2009
End Date: 12/31/2009
Congressional District: At-Large
Focus Categories: Non Point Pollution, Agriculture, Surface Water
Keywords: agriculture, pesticides, nutrient loading, irrigation, nitrogen, Yellowstone River, watershed management, water resource degradation, Bayesian inference
Principal Investigator: Mashamba, Able
Federal Funds: $ 1,500
Non-Federal Matching Funds: $ 0
Abstract: The Buffalo Rapids Irrigation District is a 45,580 acre sub-watershed located in Custer, Prairie, and Dawson counties in eastern Montana (MT) and paralleling 64 miles of the lower Yellowstone River. The district includes 155 full-time crop and/or livestock producers with an average of 164 acres per full time farm. Issues of assessment of drinking water, pollution and prevention, and watershed management and conservation are integral to the future agricultural sustainability and economics of the watershed.
Potential causes of water-resource degradation in the watershed include;
Multi-disciplinary research is currently underway to study the effects of implementing improved watershed management and engineering practices (better known as BMPs/BEPs) aimed at reducing or eliminating water degradation in the case study catchment. As part of this inter-disciplinary effort, this research proposes using a well researched complex model (the Soil and Water Assessment Tool, SWAT) to simulate and analyze the effects of implementing these improved management scenarios on water quantity, quality and use in the watershed. The reliance on irrigation delivery for crop production in the watershed provides an opportunity to model the effect of irrigation practices and develop scenarios for improved water management. Results and insights obtained from the BMP and BEP scenario research are expected to inform the interests of farmers, local and federal authorities, financiers, researchers and other stakeholders in Buffalo Rapids and similar watersheds. The challenge of the scarcity of water quality data in the watershed could hinder reliable conclusions. Even then, the potential, or lack of it, for SWAT to investigate management practices for watersheds similar to the case study will have been evaluated.
Beyond watershed management improvement analysis, this research will also contribute towards resolving challenges arising from the use of distributed or semi-distributed hydrologic models. Distributed hydrologic models are a class of complex environmental models that are finding increasing use in studying the impacts of climate change, water pollution, land use changes, human settlements and watershed management on the well being of large watersheds. Their ability to represent the heterogeneity of physical watershed processes as they are found in geographic space gives them great explanatory power (Moreda, Koren et al. 2006; Hansen, Refsgaard et al. 2007). Unfortunately their great level of detail presents two major challenges. Firstly, distributed models require a much larger input dataset as compared to other environmental models (Liu and Gupta 2007). Such data may not be readily available, which makes model calibration, validation and uncertainty estimation difficult. Secondly, distributed models inherently require more parameters than lumped models ( Moreda, Koren et al. 2006). This makes them harder and time consuming to calibrate due to the larger parameter sets that calibration algorithms must search through for optimum parameter values. Also, more parameters increases the model's prediction uncertainty (Kuczera, Kavetski et al. 2006). Quantifying a model's prediction uncertainty is crucial in improving the objectivity, credibility and hence acceptability of simulation output (Montanari 2007). Yet challenges abound in developing fast uncertainty assessment frameworks for distributed models.
Automated methods based on Bayesian inference have been demonstrated to be faster and more informative than classical statistical methods for uncertainty analysis of complex models (Kuczera, Kavetski et al. 2006; Yang, Reichert et al. 2007; Huard and Mailhot 2008). In response to the foregoing, this research will develop a Bayesian based uncertainty and sensitivity analysis framework suitable for use with complex environmental models. The Bayesian framework will be tested on the case study SWAT model developed for the watershed.
Thus the three challenges for this research will be to (i) investigate watershed management practices that are beneficial to the case study while simultaneously testing the developed approaches on a practical situation, (ii) develop a fast Bayesian MCMC error estimation approach suitable for use with complex environmental models and (iii) develop a suitable sensitivity analysis approach that can be used with the error estimation method.
Progress/Completion Report, 2009, PDF