Weed Mapping With a Drone

Recently a colleague and I used a DJI multi rotor drone to capture high resolution RGB imagery over a small 20ha project site in Sydney. With the aim to see if was possible to use the imagery to map the species composition, and volume of weeds within the site. As we did not have access to radiometric calibration targets, no atmospheric correction was performed on the raw imagery. A simple image based classification technique was then used to create a classified map from the high resolution RGB imagery.  Overall the interpretation worked well, however it did struggle in some areas, particularly trying to separate species of grasses. Manual re-coding of some sections of the image was needed. Volumes and the area coverage of weed species were then derived from the classified map. These were used to produce an estimate of the man hours and the volume of herbicide required to remove the weed species from the site.

Before people start jumping up and down about the project methodology, I would like to point out that yes I know ideally we should have performed some type of atmospheric correction on the imagery, and used a multispectral camera. However this was a simple test using only the available equipment that we had on hand, which was a Zenmuse RGB camera. I acknowledge that for this type of work a multispectral camera like the Parrot Sequoia or similar would have yielded more spectral information, and I hope to investigate one of these in the near future.

Why map weeds with a drone?

Normally I am not a big fan of simply doing something with a drone simply because you can. However in this case I think there are real potential benefits of using drones to help map weed species. There are two main benefits that drone surveys can provide. These are helping to provide an initial scope of works for a regeneration site, and then being able to provide ongoing monitoring of that site. Currently the initial scoping study of a regeneration site is performed by an experienced bush regenerator visiting the site and estimating the scope of works manually, by simply using experience as a guide. While this methodology works it is highly dependent on the experience of the estimator and presents opportunity for errors in the estimation. In this study we wanted to see if we could improve this process by augmenting the site visit with an initial site survey carried out by a drone.  From the imagery a ‘classified map’ can then be produced that shows the species composition of a site, and the volume in area that each species of weed occupies on the site. The volume information for each weed species can then be used to accurately predict the man hours required to manually remove that species and also the amount of chemical herbicide needed to treat the weed. Having access to accurate site metrics is not only helpful to bush regeneration companies bidding on projects, but is also useful to land managers like local councils, community and not for profit groups.

Equipment used

For this study we flew a DJI Matrice 100 with Zenmuse X5 camera and 15 mm lens at a height of 60 m AGL.  The Matrice 100 is DJI’s professional quadcopter and has an endurance of up to 30 minutes enabling it to cover reasonable areas in one flight. As far as UAV’s go its reasonably affordable, I think our unit as flown was about $6K AUD  worth of drone.  Note- This was the older Matrice 1 not the newer 2.


The flight plan and data capture was created with the drone deploy application, and was planned with 80/60 overlaps. As this was a small site of around 20 ha the imagery was able to be captured in a single flight, lasting around 15-20 minutes. Data was captured in JPEG format and a GPS string was written into the image header from the L1 GPS in the drone. I am going to guess that the GPS in the Matrice100 is similar in accuracy to that of the Inspire, which I tested in a previous post and found to be in the vertical accuracy range of around 3-5m. An initial orthophoto was created using the GPS airstations only, producing an orthophoto with a spatial accuracy of 1-2m. Using this image an initial image classification was run. Accuracy of this initial classification was assessed using visual interpretation methods. Believe it or not visual interpretation of aerial photography is a dying art, and is still a very effective method of mapping vegetating communities. After running a few classifications and fine tuning the performance, we then ground truthed a number of sample sites. A ground truthing site is like ground control but used to both check and guide the performance of the image classification. To collect a ground truth point, the Lat and long of each point is first soured from our orthophoto. We then use a hand held GPS to navigate to that point and record the vegetation that occurs there. Actual recorded vegetation at each truthing site is then compared to the classified vegetation class for that site, and the accuracy assessed.  For this study ground truth data was collected using a simple hand held GPS.

As stated above imagery was captured with 80 percent forward lap and 60 percent side lap. High overlaps were flown for two reasons.  Firstly high forward and side lap protects against any gaps in the data caused by platform pitching and rolling, and against any camera firing / cycle issues. Also flying high overlaps gives a true vertical perspective on the imagery, and serves to minimise any tree or vegetation lean. For this project a vertical perspective was paramount. As any tree lean could occlude smaller weeds like ground covers or small shrubs from view. As the saying goes if you can’t see it, then you can’t map it. On the day of the flight the imagery was captured under a high overcast, limiting harsh shadows leaving the imagery a little washed out with a lack of contrast. When trying to classify pixels using their spectral characteristics, a high overcast is not ideal especially if the overcast is not even. It does have some benefits however. Namely that it reduces the amount and depth of shadows cast by tall vegetation, allowing better observation of the midstory and ground cover vegetation. Our survey site was a small clearing ringed by some tall eucalypts on all sides. So had we flown under full sun these trees may have cast shadows across the survey area, potentially causing problems.

About the site

Firstly a disclaimer, I am not a botanist and not fantastic with my plant identification. I do have some basic skills but relied solely on the regeneration professionals for plant ID. So if I don’t sound like I know what I am talking about in this section, it’s because I probably don’t. From my limited knowledge I did manage to pick up that this is a fairly disturbed site, and does not have a lot of remnant native vegetation. It is mostly comprised of Crofton Weed and various grasses like African Love Grass, and Pampass Grass. Both the Pampass Grass and the Crofton Weed stuck out like the dogs proverbial in the aerial imagery, which made them easy to map. Grasses were a different story and were hard to separate.


RGB Mosaic
RGB mosaic of the project site, captured with DJI Matrice 100


So How Did We Go?

Ok is the short answer, but there are lots of areas were we could improve in future surveys. We were successfully able to map and estimate the volume for the two main weed species on the site which were Crofton Weed, and Pampass Grass. However we were not able to map most of the various grass species and smaller ground covers, as we could not separate them in the classification. Using a multispectral or Hyperspectral camera may give us the extra spectral fire power to be able to separate out these species, but I am not confident of this.

Overall for our application I think the trail was a success. Our weed species map had enough information to allow our regenerators to estimate the volume of weed species on the site and formulate an estimate of both man hours and chemical to remove them.

Also the spatial accuracy of our final map was not fantastic, being good to around 3-5m real world. For our purposes this level of accuracy was fine, but differing clients may require better. This would mean using an RTK enabled drone, or laying ground control prior to the flight, both of which would drive up the cost of the survey.

UAV Weed Map
Final classified weed map, showing the weed species and the expected amount of man hours and herbicide required for treatment.

Future Surveys

I think we coud have improved our classifcation results by using the following gear;

  • Multispectral camera (Parrot Sequoia, Tetracam, Micasense ect)
  • Atmospheric calibration targets or camera with in-built radiometric correction like the Parrot
  • PPK or RTK GPS, to enable a spatially accurate map to be produced with little or no ground control

We used an image based classification work flow in this project to create our final weed map. While an image based classification produces good results, it does require expensive software and is labor intensive. To get to a final map as shown in this study including photogrammetry would be about 2- 2.5 days labor. A far more cost effective method may be simple digitization workflow. Where the high resolution orthophoto is opened in a GIS package like global mapper and an operator simply traces around the target weed species creating a poloygon. Not quite as accurate as an image based classification but it would be cheaper faster a possibly faster and works with a standard RGB camera.

Future Applications

I can see real applications here for using high resolution imagery to aid in weed mapping in certain situations. It will only really work on sites that are reasonably open / disturbed with now canopy cover to occlude the mid and ground cover vegetation. For sites with thick low scrubby vegetation like coastal dunes that are hard to traverse on foot. I think this type of survey would be ideal, especially when combined with ongoing monitoring flights.

As with all things UAV the million dollar question is, just because it can be done with a drone, is it the most cost effective way of producing the answer. In our particular case is the extra cost involved with a UAV survey worth the extra information it provides. I think this can only really be answered on a site by site basis. For small sites, a simple site visit and walk around may be sufficient and a drone survey won’t add likely add much value. For a larger more complex site like a coastal dune with access issues, then a drone survey may be the most efficient method and provide savings.

Cheers Erron



4 Comments on “Weed Mapping With a Drone

    • Ivan, not sure as mapping grasses can be difficult due to the smaller leaf area. Assuming the tussock grass is different in apperance and spectral signature from the background species then yes it should be possible.

  1. Hi Errol, I’m looking at providing a service like this in the Monaro using machine learning on the imagery. Would love to hear more about your experience with this.

  2. Hi I am also working on weed classification.
    Do you know publicly available Multispectral dataset?

Leave a Reply

Your email address will not be published. Required fields are marked *