Case Study 1 - Multi-Res Canola

When scouting fields on foot or in the air, there is always a tradeoff between coverage and detail.

To train our model, we collected both high resolution data for detailed accuracy, and lower resolution data to support better coverage. We've built a way to align data across cameras so we can train and predict across a range of GSD (Ground Sample Distance).

Out in the field, we want a comfortable altitude buffer with any drone that we use. To achieve this, our model has to perform well under a range of conditions and GSD.

Data Collection

To create a multi-resolution dataset we trigger multiple cameras on the Matrice 4E at each waypoint. We align the high resolution telephoto frame into the lower res ‘Med-tele’ frame so they share the same pixel space and can share labels.

Tele Footprint over Med-tele

Here’s a closeup of the two camera sources sharing the same pixel space.

Tele (left), Med-tele (right)

We down sample each stream by 4x so we end up with 4 total streams of data coming from a single waypoint mission.

• Tele 1x

• Tele 4x

• Med-tele 1x

• Med-tele 4x (only used to train rows)

By flying each field at a different altitude, we get a smooth GSD distribution in our training data.

Training GSD Distribution - Med-tele is only used for rows

Filtering

With any mission, we only want to use the 'good' photos for our analysis.

Other than GSD/stage, here are our constraints around data collection:

• Row spacing must be known (in or cm)

• Camera must be facing downward (nadir)

• Rows must not cross or meet at an angle

The row detector can handle slight curves, but scale accuracy depends on reasonable row spacing, so stay toward the interior of the field where rows are straight if you can.

PlantCounts has a 'Review' step to accept or dispute any photo based on how the row inference turned out, so you're only using credits on results with high scale accuracy.

Case Study 1 - Canola (Major, SK)

The first case study is a canola field we flew in the spring of 2025. This field was held out from training.

Methods

We followed our typical R&D collection approach:

• Tele and Med-tele at each waypoint

• 5 waypoints for each soil zone (10 zones)

• 1 waypoint dropped for poor row structure

• Staging between cotyledon and 4-leaf

• 12" row spacing

The Tele 1x labels were used to evaluate the common area in the Med-tele where the two frames overlap.

Results

• Tele 1x was the strongest overall detection stream

• Med-tele 1x had the best count fit by R²

• Tele 4x was the hardest of the three, but still held up well

Canola Prediction Performance

R² - How well do photo-level predictions match the pattern of truth?

R-squared for Canola Test Field

F1 - How well do we find plants without missing or misidentifying them?

F1 for Canola Test Field

MAPE (%) - How much do photos vary from the truth?

Mean Absolute Percent Error (MAPE) for Canola Test Field

MADE (plants/ft²) - How much do photos vary from the truth?

Mean Absolute Density Error (plants/ft2) for Canola Test Field

Predicted vs. Ground Truth Charts

Tele 1x (R² = 0.982)

Predicted vs. GT Counts for Canola Test Field - Tele 1x

Med-tele 1x (R² = 0.986)

Predicted vs. GT Counts for Canola Test Field - Med-tele 1x

Tele 4x (R² = 0.972)

Predicted vs. GT Counts for Canola Test Field - Tele 4x

Conclusion

Because this test field has similar soil and farming practices to the fields in our training set, these results are likely around the peak of what our model can do, and results will vary.

If you're outside of West Central SK, we have a true 'out of domain' dataset that we have evaluated against and will be sharing soon.

To increase coverage and robustness, we will continue to push accuracy at lower resolution by continuing to transfer labels across cameras during future missions.