Using Drone Imagery & Machine Learning to Map Scotch Broom
For several years I have been interested in the advantages of using drones to collect high resolution imagery which can be used for vegetation mapping and monitoring. Imagery collected by satellite or airplane is incredibly useful for monitoring changes in land cover over time, but usually doesn’t provide enough detail to be able to distinguish between individual plant species. Imagery resolution is usually measured in meters of earth per pixel, so a 1m image has 1 meter (3.28 feet) of earth per pixel. The highest resolution imagery collected by airplane here in Nevada County was 6” resolution (6 inches of earth per pixel), and was collected in 2016. Comparatively, the imagery I’ve been collecting with my drone is typically between .5” and 1”, providing at least 6x more detail than the 2016 6” imagery, and nearly 40x more detail than the 1m imagery.
What can you do with detailed imagery? While less detailed imagery can be useful for measuring changes on the landscape, say for example to quantify how much forest was cleared for greenhouses, more detailed imagery can be used to identify and quantify species composition and distribution throughout a study area. While this can already be done on the ground using a handheld GPS unit, that is a time intensive activity. Aerial imagery collected by drone and processed using GIS software can automate this process for faster and more accurate results.
Study Plot with Manzanita, Broom, Pine, Wood Debris, and Bare Soil
To demonstrate the potential of this approach, I chose a site with a mix of native and invasive vegetation to fly and map. The site also happens to be on the campus of the school where I teach, and the math & science involved is all taught in our 6th grade curriculum.
Vegetation is considered native if it occurs naturally in the landscape & watershed where it is growing, and has important interactions with native insects and other animals. Invasive plants on the other hand are not native to the landscape they colonize, and once introduced, outcompete native plants slowly crowding them out. There are several issues with invasive plants, but one major impact is the decreased food or habitat for animal species that depend upon the native plant species they have co-evolved with.
Flowering Scotch Broom (Photo Credit: S. Oneto, UC Cooperative Extension)
Scotch and French broom are species in the pea family with bright yellow flowers. Because they are well adapted for the hot dry climate of the Mediterranean, they thrive in California and colonize whole hill sides (look for their bright yellow flowers in the springtime). As legumes, they also have a beneficial relationship with bacteria which allows them to grow in the nitrogen poor soils California native plants are adapted to. Broom plants also grow an abundance of seeds each year, which help them re-establish after they die or are removed.
As all plants are competing for resources, the colonization of broom is at the detriment of native grasses and shrubs such as manzanita. Broom has been identified as a plant of particular concern for the risk of creating “ladder fuels” or plant material in the understory which would allow a wildfire the opportunity to climb into the tree canopy above. Therefore, there is a particular effort in our community to identify, remove, and restore habitat which has been colonized by broom species. I believe that using drones and GIS software can be particularly effective for this effort to identify where broom is established, but has not yet fully colonized the habitat of native plants.
Several Manzanita Plants Surrounded by Broom (1” Resolution Imagery Collected by Drone)
Whether collecting new imagery, or using older imagery already collected, the first step is to scan the imagery for broom plants. Human beings are incredibly adept at pattern recognition, and once familiar with the shape and color of broom plants, it’s incredibly easy to identify them in high resolution imagery. For this case study I selected a huge patch of broom on campus that I have driven past every school day for the past two years. My goal was to use GIS software and the aerial imagery from the drone to identify how many species of plants were in the test plot, and calculate the percentage of the plot each species occupied.
Looking at the imagery we can see that there are several species of plants: the invasive broom, native manzanita, and also several pine trees along the outside of the plot. Using the draw polygon tool, it would be very simple, if not a little time consuming, to draw shapes around each of these individuals and use the “calculate geometry” function of GIS software to arrive at the percentage of the plot each species occupied. But for this case study I was also curious about the ability and accuracy of the GIS software to use machine learning algorithms (ML) to differentiate between each of these plant species. If the ML algorithms were accurate enough, this approach could be used to classify vegetation, and then other scripts could be used to identify habitats with specific species composition where restoration efforts could be more effective or impactful.
In a prior post I wrote about using the smart classification tool in Arc Pro (but also available in QGIS using 3rd party plugins) to identify or classify different land types in a wetland habitat. For this project I used this tool but also used the train samples tool to provide several examples of each species to the algorithm.
Providing reference data for the algorithm is as easy as drawing polygons around each vegetation type
Similar to how we teach our children what an object is by pointing it out and naming it, computer programs “learn” what objects are by being provided reference images. This could be a database of thousands of images of cats or stop signs or crosswalks. The computer program then creates its own algorithm identifying similarities between the images with cats or stop signs or crosswalks, and then tests this algorithm against the known set of images. This is done thousands (if not millions of times), until an efficient algorithm is created which can accurately identify whatever the algorithm was trained to identify. This is an important technology for many fields, but particularly in self-driving technology, which must identify a pedestrian, cyclist, or other vehicle and determine their speed, direction, and even likelihood of darting into the path of the car. If you’re interested in learning more about how machine learning algorithms are trained, here is a great short video by Youtuber CGP Grey.
Having provided the reference data, the smart classification tool takes a bit of time to scan the remaining pixels and classify them based on the reference set. Just as humans are adept at pattern recognition, with enough reference data the machine learning algorithm does a pretty good job classifying between the different types of vegetation, woody debris, and the bare soil. However, the manzanita and pine looked similar enough to the algorithm for manzanita to occasionally be classified as pine. As I was mostly interested in the broom distribution and coverage, that was acceptable to me for this case study. There are tools for reclassifying specific classes in specific areas of the map, for example any pine on the interior of the plot was actually manzanita and could be reclassified. I could not get the reclassify tool to work this time around but will revisit it in future projects.
Once classified, the final step to calculate percentage coverage is to calculate the area for each class (pine, manzanita, broom, wood debris, bare soil), and then divide by the total area of the plot.
Although the model had some trouble distinguishing between pine and manzanita, I’m fairly confident that the broom measurement is as close to accurate as possible. The study plot is 39,441 sq. ft, (a little less than an acre), so with the broom comprising 31,812 sq. feet, it covers 80% of the study plot (I selected this plot because of the high percentage of broom).
In addition to having a quantifiable metric which can be tracked year after year, our new model can also identify specific manzanita (darker green in the model above), where efforts can be focused. In future projects I hope to continue to apply this approach to larger parcels. This project looked at a single acre, but this approach can be applied to 50 or even 100+ acre plots to identify specific areas where broom is located and where multiple native species are located and can be bolstered with restoration efforts.
