2013 Progress Report: Forecasting and Evaluating Vulnerability of Watersheds to Climate Change, Extreme Events, and Algal Blooms
EPA Grant Number:
Forecasting and Evaluating Vulnerability of Watersheds to Climate Change, Extreme Events, and Algal Blooms
Stevenson, R. Jan
, Hyndman, David
, Moore, Nathan
, Qi, Jiaguo
Michigan State University
EPA Project Officer:
June 1, 2012 through
May 31, 2017
Project Period Covered by this Report:
June 1, 2013 through May 31,2014
Extreme Event Impacts on Air Quality and Water Quality with a Changing Global Climate (2011)
Air Quality and Air Toxics
Our project is organized in 3 major tasks with 3 subtasks under each major task. The first major task was to identify historical algal blooms and statistically investigate how these relate to extreme events across a range of hydrologic regimes. The second major task was to use process-based hydrology and algae models to explore causal linkages between extreme events and algal blooms, and predict the influence of projected climate changes on algal biomass. The third major task was to develop statistical models that can be applied nationwide to examine vulnerabilities to extreme events under different management strategies.
We have developed models for chlorophyll a and Secchi depth using Landsat images that are accurate for application at a national scale.
The first two subtasks under Task 1 are:
Task 1a. Characterize historic water quality conditions in Tier 1 lakes.
Task 1b. Characterize historic land uses for Tier 1 watersheds.
We proposed using Landsat imagery from as long as 40 years ago to measure chlorophyll in lakes and characterize historic water quality conditions and land use in watersheds. These data would lay the foundation for relating land use, storm events and algal blooms in subsequent modeling. Due to problems with remote sensing algorithms for measuring chlorophyll a of phytoplankton in lakes using Landsat imagery, we have spent considerable time developing and testing algorithms that can be used a large spatial scales.
Traditionally, most remote sensing models of phytoplankton biomass use linear or polynomial regression and a few band and band ratios that would be expected to sense chlorophyll. In the last report we described our calibration and testing of these models in two settings: one for 447 lakes sampled in the 2007 National Lakes Assessment and the other for the 5 Great Lakes. We found poor cross validation with new data with existing linear and polynomial models. We then applied a new statistical method, boosted regression tree (BRT), which is a machine learning algorithm, to develop models for chlorophyll a using all bands and band ratios in Landsat and MODIS satellite images. The results were greatly improved with better precision of BRT models versus linear and polynomial models in the Great Lakes than the National Lakes Assessment, similar performance of models for chlorophyll and Secchi depth, and better performance for MODIS than Landsat images.
In the last report we described next lines of research. We decided not to follow up with one of those lines, evaluating methods for integrating results of Landsat and MODIS imagery, because MODIS pixels (1 km x 1 km) are too large for most inland lakes that we plan to study. Landsat is important because it has 30 m x 30 m resolution, which is small enough to fit within the deeper contours of small lakes so marginal wetlands and algae on lake bottoms do not affect phytoplankton chlorophyll estimates. We did follow up with the other line of research, in which we tested new Landsat image products that better account for atmospheric interference. We found that these products did not help much with chlorophyll a modeling. Specular reflectance from the water surface from winds seemed to be a major source of error in several lines of our work. So we plan to develop masks for specular reflectance that will enable eliminating pixels when specular reflectance (shimmers from small waves) that infers with seeing water color.
As we progressed with our remote sensing work, we became concerned that non-chlorophyll substances in water would also produce color that could interfere with estimates of chlorophyll and evaluation of storm effects on chlorophyll. Sediment in lakes often accompanies runoff after rains, and sediment produces color in water. Colored dissolved organic matter (CDOM) is produced by in-lake and terrestrial sources, which could interfere with chlorophyll estimates. We found a unique opportunity to study effects of sediment (measured as non-volatile suspended sediments) and CDOM on Landsat measurements of chlorophyll a in a 23-year study of 39 Missouri reservoirs by Jack Jones at the University of Missouri. The dataset was used to calibrate and validate chlorophyll a models using three representative algorithms: multiple linear regression (MLR), general additive models (GAM, non-linear and non-paramedic algorithm), and boosted regression trees using Landsat TM/ETM+ images. BRT and GAM had better performance than MLR. With all algorithms, we found little (< 10%) error related to suspended inorganic sediment and CDOM. We concluded that when using empirical algorithms such as MLR, GAM, and BRT to measure chlorophyll a in turbid waters, variation in NVSS and CDOM conditions should be less of a concern than other sources of error, such as atmospheric correction and specular reflectance. This also leads us to pursue masks for specular reflectance.
During the 2014-2015 project year, we also returned to the U.S. EPA’s National Lakes Assessment to compare different advanced statistical approaches for measuring chlorophyll a with bands and band ratios from Landsat imagery. Models for chlorophyll a based on Landsat imagery were better if they were developed with BRTs than GAMs, MLR, and artificial neural networks. We found that BRTs and GAMs worked better at MLR and artificial neural networks for Landsat measurements of chlorophyll a. Because of ease of use and interpretation, we plan to use BRT models in the future.
Task 1c is to statistically evaluate relationships among hydrology, climate, land uses, and algal biomass. During the 2014-2015 project year, we returned to the U.S. EPA’s National Lakes Assessment for a third time to develop statistical models of phytoplankton sensitivity to three global change variables: temperature, phosphorus, and nitrogen. Preliminary results of this research based on BRT models show strong interactions among temperature, phosphorus concentration, and nitrogen concentration with greatest sensitivity of phytoplankton chlorophyll a to changes in any of these variables when the other two variables are relatively high. Finally, we used the BRT models relating chlorophyll a to temperature and total phosphorus and nitrogen concentrations to predict expected chlorophyll a based on global change. We assumed by assuming a 4°C increase in temperature and 20 and 50% increases in TN and TP, respectively, over the next 50 years in US lakes, as has been predicted in the literature. Based on the BRT model and our global change assumptions, we can expect an average increase of 88% in algal biomass in US lakes in 50 years. We did not find strong evidence that increases in water temperature will increase relative abundance of cyanobacteria in lakes, compared to other types of algae.
There are three subtasks under Task 2:
Task 2a. Refine and test process-based hydrology and algal bloom models in Tier 1 watersheds with different climatic and hydrogeomorphic settings
Task 2b. Statistically downscale projected changes in climate for regional watersheds including Tier 1 lakes
Task 2c. Forecast future algal bloom frequency and algal biomass under projected climate change conditions.
No advances on the Tasks 2 subtasks was made during the last project year.
Task 3 is not scheduled to be implemented at this time.
We plan to finish Task 1 and Subtasks 2a, 2b, and 3a. We will make substantial progress on Subtasks 2c, 3b, and 3c, which will put us in position to finish Tasks 1 and 2, and then Task 3 by the end of the project period.
No journal articles submitted with this report: View all 6 publications for this project
Progress and Final Reports:
2012 Progress Report
2014 Progress Report
2015 Progress Report
2016 Progress Report