Case studies
Video demonstration of how the interactive map works.
High-resolution sample images, from a height of 8-12 meters.
(Clicking anywhere on the interactive map presented in the video will display high-resolution images taken at the given location. This feature allows us to analyze every square meter in detail without having to go out into the fields.)
Drone Tassel Detection
Results
Our goal was to develop a novel software that analyzes extremely high-resolution aerial images from very low altitudes (<12m), allowing for the accurate identification of detasseling errors . Fewer abandoned tassles and faster error correction can significantly improve seed quality and value .
Target audience : 1. Seed producers/breeders, 2. Landowners, 3. Agronomists, 4. Labeling companies
The accuracy of the technology
When talking about accuracy, it is important to distinguish what stage the corn is in.
Before Flowering
After Flowering
Data and artificial intelligence training:
-
During 2024, we surveyed 550 hectares. This included 19 fields in Hungary and Slovakia.
-
The dataset included images taken during each flowering period, from multiple heights (6-30 meters).
-
The distribution of hand-detasseled and machine-cut detasseled fields in the data set is approximately 75-25.
-
The data presented in the table are the results of 800'000+ annotations.
Accuracy can be measured in practice in two ways:
-
With detailed review and annotation of images by experts.
-
Manually checking on the boards with an RTK device - this means that the inspectors receive the map on their phones/tablets and inspect a specific area with the naked eye, then compare the images analyzed by artificial intelligence with reality.
The data shown below were validated using these methods.
Tassel Detection Accuracy
(at different heights)
(Note: The artificial intelligence sees the tassels thar are visible from above almost perfectly (99%+) at all flowering stages. The numbers shown in the table are lower than 99% because, due to the density of the corn fields, there are cases where the tassel is not visible from above. This phenomenon almost completely disappears after flowering. Nevertheless, in order to visualize truly accurate data, we must also take this into account.)
It can be seen that (not surprisingly) the lower the drone footage is taken, the more accurately the artificial intelligence recognizes the tassels. In our experience, it is not advisable to take footage lower than 8 meters, as it only marginally improves the results.
Nevertheless, to recognize the tassels before flowering, it is worth going below 12 meters (or lower).
Considering scalability and time constraints, we consider a height of 8-12 meters ideal . This still allows for a performance of 150 hectares/day per drone , and the accuracy remains very high.
It can also be seen that in the post-flowering stage, the accuracy would be sufficiently accurate even above 12 meters.