Mapping bull kelp forest canopies with aerial imagery
A collaborative project between WA State Department of Natural Resources and the Northwest Straits Commission
Overview
Kelp forests are a vital natural resource that provide critical habitat to a range of marine life, fix nutrients for nearshore food webs, and play a foundational role in shaping the ecosystems of Puget Sound.
There is a growing concern that these forests are at risk of significant decline, and recent research has shown that areas such as South Puget Sound have lost most of their historic forest extent.
Many incredible projects are being conducted to monitor the distribution of kelp forests around Puget Sound, however additional data is needed to achieve the complete Sound-wide assessment necessary to discern long-term trends and identify areas that are declining and/or at the highest risk.
Zoom in and pan around in the imagery below to see the bull kelp forests along the shoreline between McCurdy Point and Fort Worden, as captured by a near-infrared camera from a manned fixed-wing aircraft.
This project represents a collaborative effort by the Washington State Department of Natural Resources and the Northwest Straits Commission to explore the use of aerial imagery to map bull kelp forest canopies and thereby fill gaps in Sound-wide data and/or to enhance existing monitoring efforts.
Platforms that were tested included unmanned aerial vehicles (UAVs or "drones") and fixed-wing aircraft, both carrying cameras that captured imagery in the visible and near-infrared spectra.
Move the slider and zoom in for a comparison of surveys with each platform at Ebey's Landing on Whidbey Island. On the left is near-infrared, green, blue (NGB) imagery captured from a fixed-wing aircraft flying at ~2k ft, and on the right is the smaller area but higher resolution (up to 5x) multispectral imagery captured by a UAV flown at ~400 ft.
Many tools and methods were explored as part of this project to assess how aerial imagery could be used to characterize kelp forest canopies. Pictured here is a slider comparing two examples: on the left is a hand-delineated forest perimeter using an 8 meter distance between plants requirement, and on the right is the output of a process known as "image classification", where kelp canopy is shown in green. These tools can be used to quantify the spatial extent and abundance respectively of the kelp present in a forest.
Overall, this project generated many findings that can help inform potential applications of aerial imagery for kelp forest mapping. The precise role that these platforms will play in kelp monitoring in Puget Sound going forward is a topic of ongoing development between the project partners, as further improvements are made to collection, processing, and analysis methods.
Please read on below for an in-depth look at the methods and results of this project, as well as more interactive maps with aerial imagery and data products generated from them!
This StoryMap is a companion to a report published by WA DNR summarizing the results of this project. A link to that can be found here .
Background
Both project partners are committed to research and monitoring of kelp forests in Puget Sound.
The NW Straits Commission's SoundIQ interactive platform hosts data on a diverse array of marine life in Puget Sound, including bull kelp. Click here to open SoundIQ in your browser!
The Northwest Straits Commission coordinates local ecosystem focused activities in northern Puget Sound, which includes supporting volunteer-based kelp canopy monitoring through Marine Resources Committees (MRCs) in seven counties. The information gathered from these efforts are hosted on SoundIQ, an interactive map-based interface for a variety of data on marine life in Puget Sound (see right).
The Washington State Department of Natural Resources (DNR) manages state-owned aquatic lands for the benefit of current and future residents of Washington State. DNR’s stewardship work includes long-term monitoring of kelp forests . Recent research efforts in South Puget Sound are presented in these StoryMaps about kayak-based monitoring efforts and the historic losses uncovered by a thorough review of historic documents.
The 2020 Kelp Conservation and Recovery Plan: link here !
WA DNR and the NW Straits Commission were both also co-authors of the 2020 Puget Sound Kelp Conservation and Recovery Plan, which summarized the state of kelp forests in Puget Sound and laid out potential paths towards more responsible stewardship of them in the future.
One of the six strategic goals it outlined was to "Describe Kelp Distribution and Trends" in greater detail than is currently available.
This strategic goal served as the primary motivation for the development of this project in order to assess whether aerial monitoring could provide meaningful data to enumerate those distributions and trends.
Project Design
Project goals
In order to assess whether UAV and fixed-wing aerial imaging platforms would be suitable to complement the NW Straits Commission's volunteer kayak survey program administered by county MRCs, the project partners devised a pilot project with the following goals:
- Collect imagery with drone and fixed-wing platforms
- Develop tools and methods for data collection, processing, and analysis
- Compare aerial imagery to MRC and DNR kayak survey data, and assess how the methods are commensurate and/or complementary
- Leverage existing resources & expertise within both organizations
Platform selection priorities
At least two platforms were to be chosen for this project, one UAV and one fixed-wing. Some important considerations when choosing these were that they be:
- Affordable
- Off-the-shelf products that do not require extensive customization to operate
- Capable of capturing imagery in both visible and near-infrared range, based on previous research showing the latter's importance to marine vegetation detection
UAV platform
The first of the two platforms is the DJI Phantom 4 Multispectral UAV. Some highlights of this device:
- Captures imagery in 5 spectral bands (blue, green, red, red edge, near-infrared)
- Flown at an altitude of ~80-120 meters above ground, resulting in a pixel resolution of ~4-6 cm
- Can cover an area up to 100 hectares in size within a single survey window
Fixed-wing platform
The second platform chosen was the MAPIR Survey3W NGB camera, which was attached to the underside of the wing of a fixed-wing aircraft. Some features of this camera:
- Captures imagery in 3 bands (near-infrared, green, blue)
- When flown at an altitude of 1,500-2,000 ft, results in a pixel resolution of 21-28 cm
- Can cover 10s of kilometers across multiple kelp forest sites in a single survey window
Study area
For this 2021 exploratory pilot project, a total of nine MRC monitoring sites were surveyed using aerial imagery. Every site was captured using the fixed-wing MAPIR NGB platform, and three sites were also surveyed using the P4 Multispectral UAV.
List of sites in study area, and when they were surveyed using both platforms.
Survey imagery
Two views of images from a survey conducted at North Beach being mosaicked in Agisoft Metashape: from above (top) and at an oblique angle (bottom). When the images are combined they create "tie points" that represent a low resolution approximation of the environment being mapped, which can be seen in the bottom image.
Imagery captured during UAV and fixed-wing surveys were processed using the software Agisoft Metashape (seen to the right). The process of turning a large number two-dimensional images into three-dimensional scenes is known as "photogrammetry" and involves both mosaicking the images together and georeferencing them to ground control points to achieve a geospatially accurate result.
The final product of this process is a large continuous image of each survey known as an "orthomosaic," which serve as the basis for all further analysis.
Key findings:
Overall, it was found that the P4 Multispectral UAV and a low cost NGB camera affixed to a small fixed-wing aircraft were both able to capture imagery capable of being successfully processed into orthomosaics using the same photogrammetry workflow.
This is considered a significant finding in that it validates the use of these platforms as a potential way to capture meaningful data about kelp forests in Puget Sound.
Explore below to see examples of these products generated at the three sites where both methods were deployed: North Beach, Edmonds, and Ebey's Landing.
NOTE: Please see the top left corner of each scene to toggle on map legends and to navigate using bookmarks.
Fixed-wing surveys using the MAPIR Survey3W NGB:
UAV surveys using the Phantom 4 Multispectral drone:
How do methods compare?
Between the two different imagery platforms
Shoreline from McCurdy Point to Fort Worden, which contains the North Beach County Park kelp monitoring site. The Phantom 4 Multispectral UAV was able to cover a smaller area (top) than the fixed-wing platform carrying the MAPIR Survey3W (bottom), albeit with higher resolution.
As part of this project, the partners were interested in the relative strengths of UAVs and fixed-wing aircraft for monitoring kelp forests given differences in costs, accessibility, and logistics involved in deploying each.
Key findings:
The central tradeoff between the two platforms is pixel resolution vs. spatial coverage.
The fixed-wing platform captured lower-resolution pixels over a larger area, and therefore would be most suitable for regional large-scale distribution mapping.
The UAVs were able to capture significantly higher resolution imagery but were confined to areas approximately 100 hectares or less per survey. This makes them more suitable to detailed analysis of the distribution and population dynamics within individual sites.
Below is a slider to enable visual comparison between two surveys at North Beach: the fixed-wing NGB on the left, and the P4 multispectral on the right.
Notice the difference in pixel resolution between the two and how that impacts the visibility of bull kelp in the imagery. This is more apparent as you zoom in.
Also zoom out to see the difference in their respective spatial coverage. It took the UAV about an hour to survey the smaller site-level area, whereas the fixed-wing platform was able to capture the entire stretch of shoreline in about 10 minutes.
Comparison between imagery and kayak-based surveys
In order to compare the imagery to kayak-based surveys a hand-delineation of the bed perimeter was performed. This involved projecting an 8 meter grid onto the imagery and tracing all plants that fell within that distance of each other.
Below are two interactive viewers comparing the kelp bed perimeters captured by MRC volunteer kayak surveys to those generated using the hand-delineation method.
Please experiment with turning on and off separate layers using the "Layers List" widget in the top right of each viewer next to the Legend. It's recommended to view each imagery layer and it's corresponding perimeter to the kayak-based survey line independently to see how they compare.
You also can change the basemap, measure distances and areas, and choose your own slider using the icons in the top left of each viewer next to the zoom buttons.
Interactive viewer of imagery and bed area delineations at North Beach.
Interactive viewer of imagery and bed area delineations at Edmonds.
Key findings:
Kelp bed area calculated for the kayak perimeters and hand-delineated digital equivalents at both sites.
Calculating the area contained within the kayak-based and hand-delineated bed area boundaries enables us to assess whether they are comparable.
In general, there was agreement between methods about the location and distribution of the kelp bed at each site.
However, there were still significant differences in the area estimates that were generated from each. This indicates that further refinement of methods is needed before they can be be considered strictly comparable to each other.
Image classification
Another way that imagery can be analyzed into meaningful data metrics is a process known as "image classification." This takes raw imagery and converts it to different classes (e.g. kelp vs water) based on information contained within each pixel and groupings of them.
For example, you can see in the information below how the spectral properties of kelp pixels (top) differ from those of water pixels (bottom) from a visible-spectrum survey. The charts on the left are called histograms and represent the distribution of pixel brightnesses on a scale of 0 to 255 for each color channel, in this case red, green, and blue (RGB).
Color channel histograms of kelp pixels (top) and water pixels (bottom) from an RGB survey.
There are many ways that image classification can be conducted on imagery. Those who are interested in learning more about the object-based Random Forest classifier used in this project are encouraged to check out the full report !
Below are a pair of interactive viewers showing the results of classification at each analysis site using both the UAV and fixed-wing platforms. The UAV results are broken down into those using just the visible spectrum bands of the imagery, and those using the near-infrared spectrum as well.
These viewers behave the same way as those above.
Please experiment with turning on and off separate layers using the "Layers List" widget in the top right next to the Legend. It's recommended to compare each imagery layer and it's corresponding classified result using the slider widget!
Interactive viewer of image classification results at North Beach.
Interactive viewer of image classification results at Edmonds.
Distribution of randomly generated accuracy assessment points, showing correct or false classification results of each against the original imagery for the P4 RGB survey at North Beach.
The accuracy of classified results was quantified using a stratified random point approach, where the accuracy of the overall product is assessed by examining a subset of randomly distributed points (see to the right). This involves manually checking the classified result against the underlying image for each point.
Accuracy values of 85% overall for these points is typically considered the threshold for usability. Another metric called Cohen's kappa compares the accuracy result to a hypothetical random result on a scale of 0 to 1. Values of 0.60-0.80 are considered substantial, and >0.80 excellent.
Accuracy assessments of image classification results at both sites using three imagery sets.
Key findings:
It was found that accuracy of the image classifier varied between sites and platforms. RGB imagery classification consistently produced the lowest accuracy values.
It also was found that the accuracy of classification in the deeper "open water" region of each survey was consistently higher than that in the shallow subtidal zone up to 50 meters from the low tide line. The primary reason for this appears to be the presence of other marine vegetation found in the shallowest extents of each survey area.
Overall, these results indicate that there is promise in the use of image classification methods to characterize kelp canopy from aerial surveys. However, more work is needed to further develop and refine the methods before they could be deployed widely and in a repeatable way.
Conclusions
This project marks the first formal research effort to use multiple aerial imagery platforms to map the distribution and abundance of bull kelp forest canopies in Puget Sound, and to compare the results those platforms yield to ground-based kayak survey methods. As such, there were many lessons learned and insights gained throughout the project that will inform future work exploring the use of these technologies to monitor kelp forests in Puget Sound and beyond.
The major takeaways of this project can be summarized as follows:
- This was an exploratory project. The results are promising and indicate that aerial platforms can effectively capture imagery of bull kelp forest canopies capable of generating meaningful data products. However, many questions remain to be answered in terms of how these methods fit into the broader goals of monitoring kelp forests in Puget Sound.
- There are important differences between aerial imagery and kayak-based surveys in terms of the kind of data they produce. Both are valuable and could complement each other to create more complete pictures of kelp forests and their health.
- The methods developed in this project are resource intensive, primarily in terms of the effort required to process and analyze aerial imagery. Organizations that are interested in using drones and/or fixed-wing aircraft to survey kelp forests and other vital nearshore habitats should be encouraged to do so, but also to assess capacity for that work ahead of time.
- DNR is planning to keep developing these methods and surveying a subset kelp beds in Puget Sound every year as resources allow. DNR also hopes to continue collaborating with the Northwest Straits Commision and Marine Resources Committees to combine efforts and gather valuable data sets about our kelp forests!
- There's lots more aerial imagery that's already been collected by the project partners that could be processed and analyzed in this way. A future project could dive into those data sets to learn more about historic kelp distribution.
Acknowledgements
The project partners would like to acknowledge and thank the following people and groups who provided valuable assistance:
- The NW Straits Commission and county Marine Resources Committee staff and volunteers, who conducted and processed the kayak-based kelp bed perimeter surveys included in this report.
- Gregg Ridder, who provides his time and resources pro bono to collect aerial imagery of kelp forest canopies throughout Puget Sound. His dedication to community-based science and the aerial mapping of nearshore habitats is deeply appreciated.
- Andrew Ryan, who has been instrumental in the development of the Nearshore Habitat Program’s UAV mapping capabilities and who participated in planning, data collection and processing.
- DNR staff who provided field and office support including Julia Ledbetter, Melissa Sanchez, Lauren Johnson, Jeff Gaeckle, Lisa Ferrier, Tim Strickler, and Blain Reeves.