
Algal Bloom Monitoring Pigeon Lake, Alberta
Earth Observation for Algal Bloom Monitoring 2017 - 2020

Pigeon Lake Study Area
Summary
Sentinel-2 and Sentinel-3, remote sensing data, is tested whether it can effectively monitor algal blooms on Pigeon Lake, AB. This is done by answering four questions including information about the temporal and spatial bloom variability. The questions include:
· What are the most common locations of the algal blooms in Pigeon Lake?
· What is the intensity, extent, severity, and duration of the blooms in Pigeon Lake?
· How different are the Sentinel-2 vs. Sentinel-3 imagery predictions of Chlorophyll-A concentrations?
· How does algal bloom severity correlate to different depths in Pigeon Lake?
Residents have noted an increase in the frequency and severity of these algal blooms, which has raised concerns over the deterioration of water quality (Hutchinson et al., 2018). When organisms such as cyanobacteria photosynthesize, they consume nutrients and produce harmful toxins that can negatively impact both humans and wildlife (El Farra, 2015). Therefore, it is of high importance to monitor the water bodies where blooms occur (Grendaitė et al., 2018; Pirasteh et al., 2020).
This project was completed by using rasters clipped by the Extract by Attribute tool to create Chlorophyll-A concentrations on the lake; these were downloaded using an algorithm in Google Earth Engine. To answer each of the questions spatial analysis was done on these rasters, this is fully outlined in the following section on analysis and results. ArcGIS Desktop tools were used in the analysis, these included: Cell Statistics, Extract by Attribute, Zonal Statistics at Table, Raster Domain, Create Random Points, and Extract Multi-values to Point. Along with this several tables and graphs were created in Microsoft Excel.
Some of the questions showed better results, questions 1 and 3. Overall the work showed that there was some success in remote sensing data being used in this way. There are however recommendations that further work should take place to fully understand the effectiveness of this monitoring method.
Rationale
Inland waters are a vital resource in Alberta (Government of Alberta, 2010). In recent years, Pigeon Lake, AB has suffered from the effects of algal blooms. Residents have noted an increase in the frequency and severity of these algal blooms, which has raised concerns over the deterioration of water quality (Hutchinson et al., 2018). When organisms such as cyanobacteria photosynthesize, they consume nutrients and produce harmful toxins that can negatively impact both humans and wildlife (El Farra, 2015). Therefore, it is of high importance to monitor the water bodies where blooms occur.
Traditionally, this type of monitoring involved in situ measurements and subsequent laboratory analyses of the collected samples (Grendaitė et al., 2018). These techniques require many hours of field and/or lab work, leading to an overall high amount of human effort. Furthermore, in situ sampling often has difficulty capturing the full temporal and spatial variability of algal blooms (Pirasteh et al., 2020).
Chlorophyll-A (Chl-A) concentrations have been identified as a direct indicator for the quantity (mass) of cyanobacterial algal bloom present in freshwater (El Farra, 2015; Grendaitė et al., 2018; Mollaee, 2018; Sayers et al., 2015). Algorithms obtained from Alberta Biodiversity Monitoring Institute (ABMI) can be applied to satellite imagery to retrieve Chl-A concentrations by pixel. Using Sentinel-2 and Sentinel-3 raster outputs of these algorithms, our team answered the following four questions in an effort to demonstrate how organizations can effectively use remote sensing data to monitor algal blooms while decreasing the need for extensive field work for ABMI:
1) What are the most common locations of algal blooms in Pigeon Lake?
2) What is the intensity, extent, severity, and duration of the blooms in Pigeon Lake?
3) How different are the Sentinel-2 vs. Sentinel-3 imagery predictions of Chl-A concentrations?
4) How does algal bloom severity correlate to different depths in Pigeon Lake?
Data Capture
Prior to the team’s involvement, ABMI had already successfully identified and adapted band math algorithms from existing research in order to extract Chl-A concentrations from Sentinel imagery. Utilizing Google Earth Engine, the team used ABMI’s codes to apply those algorithms to Sentinel-2 and Sentinel-3 image datasets and clip the results around Pigeon Lake (see Appendix A). Dates captured fell within the ice-free periods each year (May 1st to October 31st) from 2017 to 2020. Of the four years of data collected, the team obtained 49 useable Sentinel-2 images and 104 useable Sentinel-3 images.
Implementation, Analysis and Results
Question 1 Implementation
To determine the most common location of the algal blooms in Pigeon Lake, the mean Chl-A concentration was determined for the lake. This was done using the Cell Statistics tool to create a mean raster image for each separate year and one for all four years (python script see Appendix B), for both the Sentinel-2 and Sentinel-3 data. This was performed on the cleaned raster images that contained the Chl-A concentrations.
Validation of this procedure was done by visual inspection. Cell Statistics were run twice, to determine both the sums and the means. These results were then compared. Upon inspection, and consultation with ABMI, the sum vector raster displayed some linear characteristics associated with the removal of data due to cloud clipping. Though there were some of the same linear characteristics associated with the mean they were less visible and therefore, this method was chosen.
Question 1 Analysis and Results
The mean for all four years shows similar results. Most algal blooms were located in the northern part of the lake with lower concentrations in the south. Sentinel-2 results showed some edge effects, where concentrations are higher along the eastern edges of the lake. This is most likely not shown in the Sentinel-3 as the lake was buffered 300 m in from the edge as part of the data capture due to the large cell size.
Means Locations for both Sentinel-2 and Sentinel-3
Question 2 Implementation
To determine the algal bloom indices, we used the definitions outlined by the Government of Canada’s EO Lake Watch program (Environment and Climate Change Canada, 2020). Before answering question 2, we determined the bloom flag for each raster image (see python script Appendix B). A bloom flag includes all pixels within the image considered to be in bloom. To determine intensity, extent, and severity we used a 10 µg/L threshold, where all pixels with a Chl-A concentration greater than 10 were considered in bloom. We extracted these bloom flag pixels to new rasters using the Extract by Attribute tool.
Intensity
EO Lake Watch defines bloom intensity as the average Chl-A concentration within the total area flagged as an algal bloom. To calculate bloom intensities, we used the Zonal Statistics as Table tool to find the mean concentration (in µg/L) for each raster image (see Appendix C for python script).
Extent
EO Lake Watch defines bloom extent as the total area of the pixels flagged as in bloom (in Km2). To calculate bloom extent we used the Raster Domain tool on each bloom flag raster image. This extracted a polygon for each image which allowed us to calculate the total bloom area (see Appendix D and E for Sentinel-2 and Sentinel-3 python scripts respectively).
Question 2 Bloom Extent Example - Sentinel-2, 2018
Severity
EO Lake Watch defines bloom severity as bloom intensity multiplied by bloom extent (μg/L Km2). This was a straightforward calculation using the results from our bloom intensity and extent outputs.
Duration
EO Lake Watch defines bloom duration as the number of days a pixel is flagged as in bloom between June and October. For the purposes of this project, we summarized the duration for the whole lake rather than by individual pixel. Using a 10 μg/L resulted in the inclusion of essentially all image dates that we captured. After consulting with our ABMI mentor, as well as Rolf Vinebrook, a stakeholder from the University of Alberta Biological Sciences department, we adapted our methodology to use a 30 μg/L Chl-A concentration threshold instead. According to Vinebrook, 30 μg/L is a standard threshold of oligotrophic- to mesotrophic states in lakes. In other words, the threshold between low algal productivity to moderate productivity. To calculate duration, we determined the number of days between the first and last dates each year where images contained bloom flags greater than 30 μg/L. Additionally, if image dates between the first and last dates of the season did not pass the 30 μg/L threshold, then the blooms on either side of that date were considered separate blooms. Lastly, the actual number of images that passed the 30 μg/L threshold each year was recorded. This approach to duration most closely aligns with the EO Lake Watch definition. However, it is dependent upon the image dates captured each year and therefore not a reliable metric.
Question 2 Analysis and Results
Intensity
Graphs of the output intensity data revealed patterns of intensity spikes showing up in the latter half of the ice-free period each year, with greater intensities in 2017 and 2020 than in 2018 and 2019. Notably, 2018 had a very small spike in bloom intensity, while 2020 had not one but three bloom intensity spikes. Additionally, the average intensities summarized by year were higher for the Sentinel-2 imagery in each case, and over all 4 years collectively.
Sentinel-2 and Sentinel-3 Intensity Graphs
Extent
Each year, at some point, the extent of the algal blooms grows to nearly the full area of the lake (97 Km2). However, the timing of the largest extents varied from year to year. Additionally, the largest extent lasted for different periods of time each year. While there are large extents throughout 2020, there is only a brief period of 2018 with a large extent and lower extents the remainder of the year. When summarized by year, no clear pattern emerged in average extents when comparing Sentinel-2 and Sentinel-3 results.
Sentinel-2 and Sentinel-3 Extent Graphs
Severity
Overall, the severity of blooms was skewed each year with the more severe blooms occurring later in the season. More severe blooms were present in 2017 and 2020, while the year with the least severe blooms was 2018.
Sentinel-2 and Sentinel-3 Severity Graphs
Duration
Ultimately our duration results were inconclusive. Almost every Sentinel-2 image we captured had at least a few pixels with greater than 30 μg/L concentrations. The Sentinel-3 imagery provides more information, with multiple blooms and die-offs each year, with the exception of 2020. Notably, 2019 only had 3 images considered to be in bloom, each separated by other image dates which didn’t pass the 30 μg/L threshold.
Sentinel-2 and Sentinel-3 Duration Graphs
Question 3 Implementation
To determine the difference in predictions between the Sentinel-2 and Sentinel-3 data, the SAIT team compared the Chl-A concentrations at random points throughout the lake. The first step was to select Sentinel-3 rasters that were taken on the same day (and near the same time) as the Sentinel-2 rasters, totalling 15 sets of images. The Create Random Points tool in ArcMap was used to create 78 random points within Pigeon Lake. However, the lake boundary shapefile was slightly larger than the raster datasets. Therefore, visual inspection was required to delete points falling outside the raster datasets. We then used the Extract Multi-values to Point tool, which extracted the cell values of Chl-A concentrations for all 30 rasters at those points. The resulting feature class was exported as a table and saved as an Excel spreadsheet. The data were inspected for null values (from points within cloud-clipped polygons or outside the extent of the Sentinel-3 raster) and those points were disqualified. Several Pearson correlations and charts were constructed.
Question 3 Analysis and Results
A Pearson correlation using all 15 pairs of data (1023 observations(n)) revealed a strong positive correlation (r=0.79). Five of the 15 pairs had at least one sensor with partial cloud cover. A correlation was also done on these five sets (n=315), and the correlation was even stronger (r=0.89).
A correlation was also run for each year, and the correlations were strong for each year except for 2017. (2017: r=-0.12). The poor correlation for 2017 was due to one date (September 29). After consulting with our mentor, we learned the Sentinel-3 algorithm was calculating a concentration for dead algae (when it should not be). If the dataset was not included for 2017, the correlation is 0.70.
Sentinel-2 vs. Sentinel-3 Predictions Map
Sentinel-2 vs. Sentinel-3 Predictions Graphs
Question 4 Implementation
To understand the correlation between Chl-A concentration and lake depth, several analytical techniques were used. The Cell Statistics tool was used to derive the Chl-A concentration sums for each year, as well as all years together. The results of this tool created 5 new raster files (for both Sentinel-2 and 3). After this step, the team needed to utilize a tool that added the Chl-A concentration sums of the new rasters, as well as the lake depth values, to a randomly created points feature class. The Create Random Points tool was used to create 100 randomly generated points that lie within the Pigeon Lake boundary. However, the lake boundary shapefile, was slightly larger than the raster datasets. Therefore, visual inspection was required to delete points falling outside the raster datasets. Once the points were created, the Extract Multi Values to Points tool was used to extract the lake depth and Chl-A values for each individual point. The random points attribute table was then converted to an excel spreadsheet using the Table to Excel tool. The newly created spreadsheet was then used to derive the correlation between lake depth and Chl-A concentration sum, by using the built-in excel correlation function.
Sentinel-2 and Sentinel-3 Lake Depth Map
Sentinel-2 vs. Sentinel-3 Lake Depth Graphs
Question 4 Analysis and Results
From the correlation values derived from the satellite imagery, the Sentinel-2 data shows that there is a positively moderate to negatively weak correlation between the Sentinel-2 imagery and Pigeon Lake depth. Conversely, the correlation values of Sentinel-3 imagery are much higher and range from positively strong to negatively weak. It is uncertain why the Sentinel-3 data is providing much higher correlation values, however, it could be deduced that the lower resolution of the Sentinel-3 data may affect the correlation values.
Conclusions
Based on the project results, there are several overall conclusions. The most common location of algae is in the northern part of the lake, with evidence in Sentinel-2 imagery of higher occurrence along the eastern edges.
The results of our algal bloom indices were interesting. Algal bloom intensities compare well between the two sensors, with the Sentinel-2 images resulting in slightly higher readings, especially in low bloom months where the bloom is along the perimeter of the shoreline. This is likely because a 300-metre buffer was applied to the Sentinel-3 data to prevent possible overlap of the larger pixels with the shoreline. Therefore, we suspect that the Sentinel-3 results missed some of the bloom along the lake’s edge.
The main pattern that emerges from the intensity metrics is that the more intense concentrations of Chl-A are occurring in the latter part of the ice-free period each year. As there was no significant pattern in the extent results for either Sentinel-2 or Sentinel-3 imagery, the intensity of Chl-A concentrations has a far greater impact on bloom severities each year. Notably, 2018 and 2019 had low intensity and severity results. This could be due to several factors including, but not limited to: dates of the image data captured and/or the number of images capture per year; environmental factors such as years with increased fire activity, or cold and wet summers with fewer warm days; and the impacts on mean bloom flag Chl-A concentrations (intensity) from the cloud clipping methods employed. Ultimately these indices are a good starting point for developing future metrics, with more input variables, for evaluating algal bloom characteristics with satellite imagery. The final index we explored was the duration of algal blooms. Ultimately, our duration results were inconclusive. This may be due to an inadequate approach to defining duration or to a need for more comprehensive data capture methods.
As for question 3; for the Sentinel-2 and Sentinel-3 images taken on the same day, there is a strong overall correlation in the Chl-A concentrations, and a strong correlation by year, except for 2017. The 2017 correlation became strong after removing an image that contained dead algae, which was mistakenly being calculated in the Sentinel-3 data.
In terms of Question 4 both the Sentinel-2 and Sentinel-3 data show a moderate or strongly positively to negatively weak correlation of algal bloom severity correlate to different depths. With Sentinel-3 data having much higher correlation values, which could be related to the lower resolution of the data.
These conclusions answer all of the four questions proposed by ABMI. Some of the questions show better results, but overall, the work shows success in remote sensing data being used to monitor algal blooms and answer important questions relating to temporal and spatial bloom variability, all while decreasing the need for extensive field work.
Recommendations
The methods of this project could be applied to other lakes in Alberta to compare results to this project. ABMI could examine further factors, other than lake depth, as to the location of algal blooms.
For bloom duration, we qualified a raster as in-bloom if just one pixel in the entire lake is over the threshold of 30 µg/L. This should be redefined as in-bloom if a percentage of lake area (say, greater than five percent) has pixels over 30 µg/L. This would disqualify many of the Sentinel-2 rasters that have questionably high pixel values right on the edge of the shoreline.
Changes could be made to the Sentinel-3 script for downloading and applying the Chl-A algorithm. This script was made to automatically remove cloud cover; however, this function did not work perfectly, and the images had to be clipped further to remove cloud cover. It is unclear how much this imperfect methodology impacted results. Changes could be made to the code to fix this and the results reran.
We were unable to capture very many algal bloom images in July and August 2018. According to CBC News, 2018 was the worst forest fire season for BC on record (Lindsay, 2018). We did not pursue this line of inquiry any further as it was out-of-scope. However, in the future, it may be useful to examine how haze and smoke of fires impacts data capture and ultimately the analysis results.
Appendix
References
ABMI. (2021). Earth Engine Apps Experimental. Retrieved from https://abmigc.users.earthengine.app/view/pigeonlake-monitoring?mc_cid=14b8dbb3ec&mc_eid=35a2b83565
El Farra, L. (2015). A Gis Based Modelling Approach To Assess Lake Eutrophication [Concordia University]. Retrieved from https://spectrum.library.concordia.ca/979965/1/ElFarra_MASc_S2015.pdf
Environment and Climate Change Canada. (2020, September 1). Interactive algal bloom monitoring tool. EOLakeWatch: Satellite Observations for Lake Monitoring. Retrieved from https://www.canada.ca/en/environment-climate-change/services/water-overview/satellite-earth-observations-lake-monitoring/interactive-algal-bloom-monitoring-tool.html
Government of Alberta. (2010). Facts about Water in Alberta. Retrieved from https://open.alberta.ca/dataset/1832cd36-bbeb-4997-ae81-67d3eedfcfe5/resource/18a9d64b-bad8-413a-8c63-77a548ec9d88/download/4888138-2010-facts-about-water-in-alberta-2010-12.pdf
Grendaitė, D., Stonevičius, E., Karosienė, J., Savadova, K., & Kasperovičienė, J. (2018). Chlorophyll-a concentration retrieval in eutrophic lakes in Lithuania from Sentinel-2 data. Geologija. Geografija, 4(1), 15–28. https://doi.org/10.6001/geol-geogr.v4i1.3720
Harrell, Liz. (2020). Unsplash Photos for Everyone. Retrieved from https://unsplash.com/photos/yvn5UBRamAM
Hutchinson, N. J., Köster, D., Geiger, C., Hadley, K., & Smith, A. (2018). Development of an In-Lake Management Strategy for Pigeon Lake Phase 1: Problem Definition Report.
Lindsay, Bethany (2018, August 29). 2018 now worst fire season on record as B.C. extends state of emergency. CBC News, Retrieved from https://www.cbc.ca/news/canada/british-columbia/state-emergency-bc-wildfires-1.4803546
Mollaee, S. (2018). Estimation of Phytoplankton Chlorophyll-a Concentrations in the Western Basin of Lake Erie Using Sentinel-2 and Sentinel-3 Data [University of Waterloo]. https://doi.org/10.1080/07038992.2020.1823825
Pirasteh, S., Mollaee, S., Fatholahi, S. N., & Li, J. (2020). Estimation of Phytoplankton Chlorophyll-a Concentrations in the Western Basin of Lake Erie Using Sentinel-2 and Sentinel-3 Data. Canadian Journal of Remote Sensing, 46(5), 585–602. https://doi.org/10.1080/07038992.2020.1823825
Sayers, M. J., Grimm, A. G., Shuchman, R. A., Deines, A. M., Bunnell, D. B., Raymer, Z. B., Rogers, M. W., Woelmer, W., Bennion, D. H., Brooks, C. N., Whitley, M. A., Warner, D. M., & Mychek-Londer, J. (2015). A new method to generate a high-resolution global distribution map of lake chlorophyll. International Journal of Remote Sensing, 36(7), 1942–1964. https://doi.org/10.1080/01431161.2015.1029099