Abstract:
How does climate change alter water quality? This question has long been overlooked such that the effects of climate change on water quality have remained ‘invisible’. As an example, the most recent IPCC report discussed extensively climate risks associated with water quantity such as floods and droughts, but barely discusses the risks of inland water quality in the context of global climate risks. In particular, river water quality is often considered as influenced more by human activities such as land use and less by climate change, therefore often overlooked in climate risk assessments. Evidence however has been mounting, on the changing river water quality changes in a warming climate. For example, continental-scale data reconstructed using deep learning approaches have shown widespread river warming and deoxygenation in US and in Central Europe, at an extent more than those observed in rivers.
Understanding and predicting river water quality is challenging, especially with intensifying climate extremes expected in the future. Complex processes and costly data collection contribute to data scarcity, making accurate predictions difficult. Traditional physics-based models often fall short in addressing these challenges. Deep learning has emerged as a promising yet underutilized tool for uncovering patterns in high-dimensional water quality data and for filling temporal and spatial data gaps. In this talk I will highlight the strengths and limitations of deep learning and physics-based methods, emphasizing its potential to advance the understanding and forecasting the future of river water-quality.
Bio:
Dr. Li Li (李黎) is the Barry and Shirley Isett professor in the Dept. of Civil & Environmental Engineering at Penn State University. Her group works at the intersections of hydrology, biogeochemistry, ecology, and environmental engineering. She asks questions on how climate and human perturbations (e.g., land use) regulate hydrological and biogeochemical processes at earth surface and subsurface, ultimately shaping water quality. Her group uses big data, machine learning tools, and process-based reactive transport models to understand and forecast patterns and processes that drive temporal trends and spatial patterns of water quality from watershed to continental scales. In particular, her group’s recent work on river warming and deoxygenation has increased the awareness of the often-overlooked impacts of climate change on river water quality. She has been co-hosting a monthly seminar Women Advancing River Research since pandemic, promoting woman scientists’ work across the globe. She earned her PhD in environmental engineering and water resources from Princeton University, and MSc and BSc in environmental chemistry from Nanjing University, China. Before joining Penn State, she worked at Lawrence Berkeley National Lab.
Summary:
Focus: modeling hydrology
Earth is covered in water but most is saline
River water is potable and most human settlements are near rivers
Sources of tap water:
Groundwater: 30%
Surface water (rivers, streams, lakes, reservoirs): 70%
Water goes through filtration plants, then distributed via municipal water pipes
Water contains many chemicals
Dissolved O2, Sediments, Nutrients (N,P), Carbon, Toxic metals, etc.
Need to be filtered out but many are not, especially in developing countries
4.4b people lack safe drinking water
Water treatment is expensive
US spends $10s b / year | .8% GDP | $400/year per US citizen
River water quality is increasingly threatened
Floods, droughts, etc. degrade water quality
Climate change impacts but but Issue doesn’t get much attention
Focus: using deep learning models to reveal water quality dynamics
Challenge: data scarcity on water quality data
There are stream flow gauges, but water quality only ~2% of the time (water quantity available 60% of the time)
Measurement bias: richer places measure more and extent of measurement varies over time with funding
Deep learning in hydrology
Use has jumped exponentially over past few years
Focus of analysis: Dissolved O2
Depends on flow speed, light, temperature, photosynthesis by water plants (absorb O2)
Approach LSTM model
Given sparse time series of river measurements and context:
Weather/climate
Nearby terrain: land use, rocks/vegetation, terrain
Fill in data gaps
Predict O2 measurements in rivers with no chemical sensors
Forecast future
Observations
Warming in > 87% rivers
Daster in urban areas (urban heat islands)
Slower in rural rivers
Deoxygenation in >70% rivers
Fastest in Agricultural rivers (higher concentration of nutrients N,P, which facilitate plant respiration in rivers)
Rivers warm and deoxygenate faster than oceans but slower than lakes and coastal areas (worst hit)
Main variables driving deoxygenation: Temperature > Day length > Radiation > Flow
Model performs better at sites with less variability and fewer extreme events
Performs poorly at low O2 concentration, concentration pulses
Temperature and O2 have the most data available. Inferences for more data-poor chemicals will require correlation to these, chemical models, etc.
Physics-based models
River water comes from subsurface water (60%) and precipitation
Source affects river water chemistry
External forcing from climate/weather and human effects
Aboveground: land cover
Belowground: geology
Reactive Transport Models
Meteorology+Hydrology+Ecology+Aquatic/Soil Chemistry
Structure: break up system in sub-components
Surface, Shallow Zone, Deep Zone
Model mass flow in a conserving way
Focus on carbon in water
Dissolved: Organic, Inorganic
Particulate
Soil CO2 gasses
Spatially explicit analysis of dissolved organic carbon
Depends strongly on weather
Produced in Summer but not released because plants collect water
In Fall, water released and carbon flushes out
Also, substantial “basin respiration”, where inorganic carbon leeches from deep underground
Enables process understanding via thought experiments
Can measuring river water chemistry enable measurement of subsurface water?
Groundwater chemicals (Ca): more concentration with lower streamflow
Surface chemicals (DOC): less concentration with lower streamflow
Summer: dominated by deeper water
Winter/Spring: dominated by shallow zone
Confluence between deep learning and physics-based models
Challenges: data availability, complex processes
Opportunities:
Physics-based models generating “data”
Interconnections between variables
Water quality under Climate Extremes
Few observations of extreme events (rare, events break sensors)
Atypical dynamics during extreme events