Colorado District 2 - NASA Earmark Project
Fusing In-Situ and Remote Sensing with Machine Learning Tools to support Measuring and Mitigating Land Management Impacts on In-Stream Water Quality and Community-Led Climate Financed Riparian Restoration
Project Scope
This project builds on the State of Colorado Legislature and Governor Polis commissioned, ‘Emerging Technologies to Improve Water Resource Management in Colorado,’ report written by the proposing team, and pointed to the opportunities to use quantification and monitoring technologies to better apply limited conservation funding. More public funding demonstrating the path forward, especially in regions like the Yampa and Poudre where regulatory pressure is already causing utility planning for expensive and energy-intensive upgrades, can only increase the ability for leaders and stakeholders to guide decisions towards the best holistic outcomes.
Further, these community based demonstration projects in Steamboat Springs and Fort Collins will provide the necessary market-making proof of concept to unlock private sector investment in the form of carbon credit programs. Our preliminary estimates suggest that carbon finance revenue could provide about 30% of the costs of watershed projects, thereby enabling hundreds of activities that are currently not economically viable. Across hundreds of watersheds and water management systems, we estimate that the US market potential for this methodology could be as much as 30 MtCO2e per year, or about a $680M market opportunity.
This research will serve as critical demonstrations of generating ‘environmental outcomes’ which can be used for Clean Water Act compliance or to achieve voluntary goals for improved watershed resiliency. City leadership as well as regional stakeholders would become much more aware of natural infrastructure implementation opportunities and details, and thus more likely to work with their regulators to find regional solutions which use less energy and carbon than further upgrades to gray infrastructure.
This work would provide the required community-based demonstrations needed to scale a viable carbon finance methodology that will catalyze and incentivize heightened investment in regional wildfire and watershed resilience by public and private water utilities seeking to comply with environmental regulations.
The proposed scope of work dovetails with a National Science Foundation funded project led by the private company, SweetSense Inc., with subcontracts to the University of Colorado Boulder, Colorado State University, Friends of the Yampa and the Coalition for the Poudre River Watershed. This proposal continues the initial deployment of sensors and data analytical technologies and supports the methodology, policy and practice advances required. The NSF funded project extends through August 2024, after which this NASA-funded project would continue these efforts.
The scope of this project includes:
Implement in-stream water quality monitoring to calculate environmental outcomes and support the design of nutrient, wildfire sediment and temperature-reducing streamside vegetation projects in the Yampa watershed.
Implement in-stream water quality monitoring to calculate environmental outcomes and support the design of nutrient, wildfire sediment and temperature-reducing streamside vegetation projects in the Cache de la Poudre watershed.
Fuse in-situ in-stream water quality data with NASA and other remote sensing indicators of land cover, land use and land use change to develop predictive tools to establish the impact of land management practices on river water quality.
Conduct lab and field-based evaluation and validation of a tryptophan-like florescence sensor designed for alarm threshold detection of fecal contamination in drinking and river water.
Leverage community and local government partnerships in Steamboat Springs and Fort Collins to co-design data-informed watershed solutions.
Apply data-informed carbon credit methodologies toward the design and initial deployment of watershed restoration solutions along the Yampa River.
Consistent with recent Colorado state-level legislation, support the Colorado Department of Public Health and Environment in developing data-informed processes, rules, and pilots that enable pre-permit watershed restoration solutions along the Yampa and/or Poudre and extend these capabilities across a relevant selection of communities in Colorado.
Sensor Design
Virridy has developed the Lume for the measurement of contamination risk in water. The Lume uses fluorescence combined with machine learning analytics to estimate contamination parameters with remote reporting to alert water stakeholders.
This sensor technology is the first to demonstrate a fully integrated in-situ, autonomous, internet-connected water contamination risk sensor. Compared to other technologies on the market, the Lume:
Outputs quantification of water quality levels and not simply relative fluorescent units;
Does not require regular calibration and cleaning;
Is designed continuous in-situ use;
Retails for considerably lower cost.
Our design approach includes:
Applying our experience in IOT systems design to create a compact, rugged device designed for continuous, remote use;
Using advanced LEDs and silicon photomultipliers to reduce cost, complexity and increase stability, and
A quantification analytical system that compensates for background noise, biofouling and signal drift through fusing a network of sensor data with remote sensing and other parameters (streamflow, rainfall, landcover, etc.) in order to accurately estimate fecal contamination risk and change.
Research Sites
Nutrient Runoff, Yampa River, Colorado
Accurate, high-frequency monitoring of total organic carbon (TOC) and nitrogen concentrations in rivers is crucial for water quality management, but often challenging to implement. In a study conducted in partnership with In-Situ Inc. and Friends of the Yampa, Virridy demonstrated a novel approach to estimating TOC concentrations by combining in-situ fluorescent dissolved organic matter sensors with high-resolution land use land cover (LULC) data to predict TOC and total nitrogen (TN) concentrations in the Upper Yampa River, Colorado. We found that integrating fluorescing dissolved organic matter (FDOM) measurements with machine learning (ML) and downscaled LULC yields more accurate predictions than traditional grab sample monitoring methods. To provide fine-temporal resolution estimates of TOC, we developed a gradient boosting machine model on data collected from July 2023 to July 2024, incorporating FDOM readings and LULC data derived from 10 cm resolution imagery. With this model, we achieved root mean square error of < 0.7 mg/L for TOC with overall prediction errors below 8% with spatial-temporal cross-validation from withheld data. Very high-resolution LULC data significantly improved model performance compared to standard 30 m resolution data. Additionally, we successfully developed an ML pipeline that can use sensor inputs to accurately predict TN concentrations in relevant categories of low (TN < 0.1 mg/L), medium (0.1 mg/L < TN < 0.45 mg/L), and high (TN > 0.45 mg/L) with overall accuracy of 65%. This approach enables real-time water quality predictions, facilitating rapid response to changing conditions and enhancing compliance with Clean Water Act standards, and allows for more comprehensive river monitoring applicable to diverse watersheds.
The Yampa River watershed spans 250 miles through Routt County, Colorado, maintaining one of the last near-unimpeded rivers in the western U.S. This 45 km study reach extends from Bear River headwaters through Stagecoach Reservoir to Steamboat Springs. This segment was selected due to its thermal impairment status (303(d) listed since 2019), mixed land use gradient (62% agricultural, 28% forested, 10% urban) and hydrological controls (two minimally invasive reservoirs accounting for 15% of annual discharge. Field sensors were co-located with USGS’s long-term monitoring stations to leverage streamflow data as well as placed above and below points of potential contamination including historic ranching properties, reservoirs, water treatment facilities, and urban areas.
Comparison of classified remote sensing image from Urban Sky (left) and NLCD (right) at a water quality monitoring site in Steamboat Springs, Colorado, providing a visual of the impact of higher resolution data where the Urban Sky data provides finer detail of water versus cultivated (i.e. agricultural) land. The legend in the right-hand corner displays the land use classifications used between both data sets.
We deployed a network of ten in-situ sensors from July 2023-April 2024 at 1-7 km spacing, maintaining 15-minute logging intervals. This high temporal data resolution captured fine-scale variations in water quality parameters that may have been lost by only collecting grab samples. Sensors were removed from the river and monitoring was suspended from mid-November to mid-April to prevent equipment damage due to low temperatures while still capturing the entirety of the irrigation season and the snowmelt runoff. FDOM, turbidity, temperature, conductivity, and chlorophyll-a were continuously monitored, chosen as key indicators of water quality that are measured by in-situ sensors. FDOM acts as a proxy for total organic carbon loading, while chlorophyll-a indicates potential harmful algal blooms related to nutrient loading, both of which are critical for assessing the impacts of agricultural practices and land use changes on water quality. These sensors included thermistors for water temperature as well as optical sensors for turbidity and FDOM. The optical sensors were equipped with automatic wipers to reduce fouling and help to maintain accuracy. Field cleanings and calibration checks of the in-stream sensors followed standard USGS techniques and methods. FDOM sensors excite at 375 nm and detect emissions in the range of 460-530 nm. We paired this high-frequency sensor data with weekly grab samples that were lab-analyzed for quantification of nutrients including Total Nitrogen (TN), TOC, Total Kjeldahl Nitrogen (TKN), Nitrate/Nitrite as N, Total Phosphorus (TP), and Potassium.
GBM prediction of TOC from high frequency FDOM data by month from July 2023 to April 2024. Including watershed-level LULC data (labeled as Ag. in Buffer) into the model improved RMSE to less than 0.7 mg/L for TOC, an average prediction error of 11%. The percentage of agricultural land at a watershed level (derived from NLCD data) was found to be a significant predictor of in-stream TOC.
Lab-measured TN across sensor locations is plotted here across monitored time periods (data was not collected when the river was frozen from December 2023 to March 2024). Values are color-coded by monitoring location and marked by correct classification of low, medium, or high TN. Actual thresholds are marked with a dashed line at 0.1 mg/L (low) and 0.45 mg/L (high). While the threshold model performed well for low and medium TN values, it failed to accurately predict values of high TN.
Boulder Creek, Colorado
Globally, over 4.4 billion people drink fecally contaminated water. In the United States, half of America’s rivers are impaired in part by fecal contamination and do not meet the Clean Water Act standards. Fecal contamination stems from treated and untreated wastewater, organic decomposition, and agricultural runoff. Socioeconomically marginalized communities are often the most negatively impacted by poor water quality. Monitoring of the source, concentration, risk and trends of fecal contamination exposure is a major limitation toward improved water quality management both in the United States and globally.
Fecal contamination in water sources is a significant public health concern, being a major contributor to waterborne diseases. Detection methods utilizing sensors, particularly those based on tryptophan-like fluorescence, offer real-time monitoring capabilities that allow for rapid identification of contamination events. Tryptophan fluorescence serves as a biomarker indicative of fecal pollution and is often linked to the presence of fecal indicator bacteria (FIB), such as E. coli and enterococci, which are associated with human and animal waste. Hyperspectral imaging techniques have demonstrated potential in detecting fecal contamination by analyzing fluorescence emitted by tryptophan and similar compounds found in fecal matter. For example, studies have applied hyperspectral fluorescence imaging successfully to identify fecal contamination on various surfaces, including agricultural products. These methods utilize the specific fluorescence signature emitted by tryptophan when exposed to certain wavelengths, enabling differentiation between contaminated and uncontaminated samples, even at low contamination levels (Gerba, 2013). Moreover, real-time water quality sensors significantly enhance the monitoring of microbial contaminants, addressing challenges posed by traditional sampling methods. Advanced optical sensors have proven effective for continuous monitoring of E. coli in water distribution systems, indicating their potential for detecting fecal contamination events. Recent research utilizing fluorescent optical sensors has demonstrated sensitivity for identifying microbial contamination at notably low bacterial concentrations, underscoring their applicability in public health surveillance. The integration of these technologies into a wireless sensor network framework allows for comprehensive evaluations of water quality. Continuous data streams from these sensors enable timely responses to detected contamination, facilitating public health interventions before disease spread occurs. Furthermore, emerging approaches that employ machine learning algorithms for analyzing sensor data show promise in enhancing predictive capabilities related to fecal contamination trends over time.. In summary, combining tryptophan-like fluorescence with advanced sensor technologies provides a reliable method for detecting fecal contamination in water. This integration also strengthens public health measures by allowing proactive responses to potential outbreaks associated with contaminated water sources.Â
The Lume uses tryptophan-like fluorescence (TLF) combined with machine learning analytics to estimate E. coli contamination with remote reporting to alert water stakeholders.
The Lume sensitivity demonstrates performance in distinguishing between between DI and E. coli concentrations in a 5% wastewater effluent dilution which was enumerated with membrane filtration and contained 10 CFU/100mL, which signifies intermediate risk contamination (p-value <0.01, t = 4.2). The R2 between E. coli concentrations in the wastewater effluent dilutions and sensor output was 0.92. There was a drop in sensor output in the range of 1000 CFU/100mL. The drop in sensor output can be attributed partly to IFE, but could also be a result of light scatter from particles. The absorbance data collected on a benchtop spectrophotometer shows increasing absorbance as the concentrations increase. The calculated corrected fluorescence due to IFE increases the R2 between E. coli and sensor output to 0.95.
Sensor response from wastewater effluent dilutions graphed continuously. Boxes indicate the interquartile range and median, whiskers indicate maximum and minimum values except where outliers are indicated. The bar between 0 and 10 CFU/100mL indicate the significant differences between indicated concentrations calculated by a Students t-test.
Over a third of wastewater treatment utilities in Colorado discharge into rivers that do not meet Clean Water Act nutrient or bacterial water quality standards. In many cases, built infrastructure could be substituted with green infrastructure to reduce non-point source contamination. However, the deployment of these approaches has been limited in part due to the lack of technologies capable of directly monitoring water quality and attributing any change to improved watershed management practices. Continuous monitoring of nutrient and bacteria in the stream has been technologically limited. Currently, there are no specific technologies that can autonomously monitor water quality in situ and provide a direct enumeration of nutrient driven algae blooms or fecal bacterial contamination.
We are currently monitoring fecal contamination in Boulder Creek to support stormwater permit compliance efforts associated with the E. coli Total Maximum Daily Load (TMDL) for Boulder Creek, one of the only such TMDLs in Colorado. Sensor data is paired with sampling and geospatially derived land management practices in a machine learning model to extend local sensor readings into spatially interpolated quantified contamination estimates. The City of Boulder and the University of Colorado Boulder are together regulated by the CDPHE to monitor and control e. Coli contamination in Boulder Creek. This presently manifests as weekly e. Coli sampling at six locations. Virridy will deploy sensors at each of these six sites, conduct 3-4 weekly grab samples, and provide real time web-based sensor and weekly lab data to the City, CU and the public. \textit{The City of Boulder has articulated using the Lume data to support continuous monitoring of water quality and public communication of safe and unsafe times to recreate in Boulder Creek. CU Boulder has articulated using the Lume data to support improved understanding of when and where e. Coli contamination events occur, and further to support monitoring of the effectiveness of their routine, scheduled stormwater sewer cleaning.
Four Lume TLF sensors are installed along Boulder Creek, Colorado.
At each sensor site, periodic water samples are collected and enumerated for E. coli.
The prediction model is further informed through the incorporation of further remotely sensed and in-situ data sources as machine learning features. These include land use classifications (below) and land cover data, streamflow, andambient temperature.
Variable Importance Plot for Fused Remote and In-Situ Sensing Machine Learning Model.
Groundwater Pumps, Northern Kenya
The cyclical and intensifying droughts in northern Kenya, exacerbated by climate change, result in millions lacking reliable water access, with the government and international bodies providing only temporary, reactive emergency support.
Despite investment in borehole pumps, local communities and governments struggle with operation and maintenance due to a lack of funding, training, and viable service contracts, leading to high water point failure rates.
Launched by the Millennium Water Alliance, Virridy, and the University of Colorado Boulder, and primarily funded by USAID’s BHA, this program aims to create sustainable groundwater demand forecasts and incentivize efficient water system operations to prevent drought emergencies.
Drinking Water Filters, Rwanda
Virridy’s Amazi Meza program in Rwanda is deploying water treatment systems at primary and secondary schools for students who currently drink microbially contaminated drinking water. Diarrhea, associated with dirty drinking water, is still the leading cause of illness and death among school-aged children in Rwanda. Our program will reach about 600,000 students with clean drinking water services over the next several years and is expected to generate over 200,000 carbon credits by 2030.
The Virridy management team led the development and implementation of the first-ever United Nations Clean Development Mechanism and Gold Standard programs earning carbon credits for water treatment. Through these programs, tens of millions of dollars of private financing were leveraged to deliver household water filters to millions of people in Rwanda and Kenya, with revenue from carbon credits largely re-invested into education, repairs and replacements and resulting in significant health, economic and environmental benefits.
Chicago River, Illinois
