Top Crop Manager

Features Canola
Eye in the sky counts canola plant stands

Research on drone imagery accurately counts seedlings.

November 16, 2023  By Bruce Barker

Kaylie Krys has developed a rapid and cost effective way to assess canola stand establishment with a drone. Photo courtesy of Kaylie Krys.

Throw away the hula hoop for counting canola stand establishment. University of Saskatchewan (USask) research is developing a way to count canola seedlings using unmanned aerial vehicles (UAV).

“Drone-based plant population estimates can benefit farmers by saving them time when scouting. It also carries the benefit of a data trail for spring emergent counts. Not only does the producer now have accurate and consistent counts but also an imagery file that can be looked back on in cases when the data may need to be revisited or investigated. This could be if a yield map was showing unexpectedly low yields in an area,” says Kaylie Krys, a master’s student at USask in professor Steve Shirtliffe’s Agronomic Crop Imaging Lab. “It also provides timely feedback in the spring, such as if reseeding is necessary due to plant death by flea beetles, seeding errors or poor emergence.”

The conventional method of assessing canola plant stands is to use a one-quarter metre or two-foot square hula hoop that is thrown in several areas of the field, and canola seedlings are counted inside the hoop. The current recommendation for manual plant counts is to take about 15 to 20 random plant counts in a field to obtain a representative sample.


“The consistency of the automated computer-model plant stand estimation is a huge aspect of this project. Human error can occur from field to field as the day wears on due to a number of factors, such as weariness, weather and different individuals counting,” says Krys. Krys became interested in how drones could be utilized in agriculture after taking an agriculture drone course by Landview Drones in the summer of 2020. She comes from a small mixed cattle and grain operation, and the use of drone and satellite imagery to collect on-farm data was exciting for her because it was a scalable practice that could be applied to farms of all sizes. Krys’ interest was piqued by Shirtliffe’s program because it had both field and imagery applications, was working at the field-scale and she could partner with canola producers across central Saskatchewan.

“An important part of this research for me was for it to be practical for producers. Producers do not need the newest piece of equipment with appropriate sensors for this data to be collected and applied,” says Krys.

How it works
Krys used a DJI Mavic 2 Pro drone with a standard RGB visible light camera that retails around $3,000. She developed flight plans using U.K. flight software from DroneAg called Skippy Scout. With this software, Krys maps the outline of the field and inputs the scout points that she wants to image. Fifty images per quarter section are taken, with each image size of 16 square feet (1.5 m²) taken 6.5 feet (two metres) off the ground surface. These images are then uploaded into our computer model sorted by field and analyzed for plant stand estimations per square metre and ground cover percent.

A computer model was developed and trained to recognize canola plants in the cotyledon to two-leaf stage. Krys says that the images are taken when canola seedlings are in the cotyledon to two-leaf stage because once the seedlings become larger, the computer model has a hard time telling the difference between plants if their leaves overlap.

The computer modeling was done using a convolutional neural network object detection methodology. This uses bounding boxes to determine if an object is present and the size/dimensions of that object. In this case, it was done by taking the images and manually drawing a bounding box around each canola seedling in the image, which sometimes numbered 200 to 300 seedlings per image. These hand annotations were also used to increase the accuracy of the model by training and testing it on a number of different images from different fields, says Krys.

The computer model analyzes each image, provides the plant counts for each image, and an Excel output for each field flown. The accuracy level, or F-1 score, has reached 0.91 (on a scale of 0-1), or about 91 percent accuracy. Krys says that if there is a situation where the F-1 score is lower due to a specific field, site-specific training can take place using the computer model to increase the accuracy for that field. This may occur if a field has a unique characteristic that the model has not seen before, such as plants in a different growth stage.

“Collaboration was key to make this project the success it has been so far. Erik Andvaag is a master’s candidate in computer sciences at USask who has been working on the computer model with me. It is the combination of his computer modeling skills and my agronomy knowledge that has gotten us to where we are today,” says Krys. She is also extremely thankful to the five producers across central Saskatchewan who provided over 50 fields to image in the two growing seasons during the short emergence stage of the canola.

Faster, more timely sampling
Time and consistency are two of the main advantages that UAV applied plant stand estimates can provide. In Krys’ research, she was able to image 50 sample points of 16 sq ft (1.5 m²) across a quarter section in approximately 30 minutes, depending on wind speeds. The flight paths can also be set up so that the UAV can be used to survey difficult parts of the field, while manual scouting can be conducted in areas closer to the field access if preferred.

Krys says the use of the automated count means there is consistency from the first field to the last. Other benefits include being able to check on the crop when access to the field is difficult, such as after a large rainstorm. UAV-based plant stand estimates can also be applied as a management practice for clubroot management protocols by limiting the number of times the field needs to be entered

Once Krys’ master’s program is completed, the hope is that the research will become available to canola growers. It is currently set up as a web interface that can access all aspects of the computer model over the internet, including the uploading of images, image processing, annotations, training image sets and viewing and downloading results.


Stories continue below