​Multiple Images Improve Lake CDOM Estimation: Building Better Landsat 8 Empirical Algorithms across Southern Canada

Picture
Figure 1. Locations of calibration (grey), validation (blue), and the additional 128 lakes without clear imagery taken within 30 days of sampling (purple) plotted on a map of Canada.
Picture
Figure 3. Map of standardized residuals for global regression model between in situ coloured dissolved organic matter (CDOM) and Landsat 8 B3/B4 and B2 for data obtained within a 30-day window of each other.Most standardised regression residuals fell within 1.96 standard deviations of the mean for this study, with little apparent spatial pattern in how positive and negative they were throughout the country.
Background

Coloured dissolved organic matter (CDOM) is the "tea-colored" part of dissolved organic matter that regulates light, temperature, and chemical safety in freshwater ecosystems. Despite its ecological importance, affecting everything from primary productivity to the cost of human drinking water purification, CDOM is rarely measured by public water-quality programs. In a country like Canada with nearly a million lakes, traditional in-person monitoring is physically and financially impossible at scale. While satellite remote sensing offers a solution, previous efforts were bottlenecked by the need for field crews to sample lakes on the exact day a satellite passed overhead. We saw a critical need to move beyond these restrictive timing requirements to build a truly national-scale assessment of lake CDOM.

Approach

We leveraged the high-precision sensors of the Landsat 8 satellite and the massive processing power of Google Earth Engine to analyze lakes across southern Canada. We used high-quality field data from the NSERC Canadian Lake Pulse Network, which sampled over 600 lakes during maximum summer stratification. Instead of relying on a single, "perfectly timed" image, we pioneered a median filtering technique. This involved taking the middle value of all available satellite observations for each lake over multiple summer seasons to eliminate "noise" caused by atmospheric haze or wildfire smoke. We then used random forest models to identify the most effective light frequencies, specifically the ratio of green to red light, to accurately estimate CDOM levels from space. 

Key Findings

We identified several breakthroughs that transform how satellite data is used for environmental monitoring:
  • Superiority of Multi-Image Medians: Models that utilized median band values from multiple years performed significantly better (adjusted R² = 0.70) than those restricted to a narrow 30-day window (adjusted R² = 0.45).
  • The "Four Image" Advantage: We discovered that model accuracy improves rapidly as more images are added, reaching an optimal point at approximately 12 images. Remarkably, selecting any four random summer images provided a better model than one carefully curated image matched to a specific sample date.
  • Expanded Monitoring Scope: By removing the need for tight timing, we increased the number of lakes available for study by 40%, effectively including remote water bodies that were previously excluded due to lack of same-day imagery.
  • Global Algorithm Consistency: We found no significant spatial patterns in our model errors, proving that a single global algorithm can work across the vast and diverse ecozones of southern Canada, mirroring our previous success with Secchi disk depth monitoring.

Impact

This research provides a powerful, automated toolkit for monitoring lake health on a continental scale. By proving that "temporal matching" is unnecessary, we have unlocked decades of existing field data for use in satellite calibration. This approach allows for the cost-effective monitoring of CDOM, which is vital for predicting how lakes respond to climate change, managing water treatment costs, and protecting aquatic habitats. Ultimately, we are making it possible to characterize nearly a million Canadian lakes, providing a scientific foundation for national freshwater policy.

Resources

Published Paper: Koll-Egyed* T, Cardille JA, Deutsch* ES. Multiple images improve lake CDOM estimation: building better Landsat 8 empirical algorithms across southern Canada. Remote Sensing. 2021; 13(18):3615. DOI : https://doi.org/10.3390/rs13183615