A breeder in the sky: scoring flowering with fewer flights

In the world of agriculture, timing is everything. For a plant breeder trying to develop a new variety of corn, knowing exactly when a crop will flower is critical. This moment determines whether the plant can survive a local frost, how long it has to grow before harvest, and whether it will thrive in a specific region. Traditionally, finding this answer has been a laborious, human task. Because flowers do not appear on a fixed calendar date but rather depend on weather and genetics, researchers must walk through their fields every few days, watching for the first signs of pollen or silk. They must repeat this process for weeks, from the earliest blooming plant to the latest, to get an accurate average for each plot. This constant monitoring is slow, expensive, and limits how many fields a team can study at once.

To solve this, scientists have turned to drones, or unmanned aerial vehicles, which can snap high-resolution photos of entire fields from the sky. However, simply taking pictures is not enough. Previous attempts to use these images to predict flowering time often required a drone to fly over the same field dozens of times, capturing images every single day or two during the blooming season. This created a massive amount of data that was difficult to process and still required frequent, costly flights. The challenge remained: could a computer learn to tell exactly when a corn plant flowered from just a handful of photos, perhaps even just one, without needing to be flown over constantly?

A team of researchers has now answered this question with a new approach that changes how the problem is viewed. Instead of teaching a computer to simply decide if a plant is "flowering" or "not flowering" in a single snapshot, they trained an artificial intelligence model to act as a timekeeper. They asked the model to look at a photo of a corn plot and estimate how many days away the plant was from its peak flowering moment. This could be a negative number, meaning the plant had not yet flowered, or a positive number, meaning it had already passed that stage. By treating flowering as a continuous timeline rather than a simple yes-or-no switch, the model could learn from the subtle changes in the plant's appearance both before and after the bloom.

The researchers built this system using a massive collection of data gathered over six years. They combined more than 200,000 drone images taken from field trials in seven different states, covering over 1,200 different types of corn hybrids. These images were paired with precise records of when each specific plot actually flowered, as determined by human observers walking the fields. The team used a powerful type of AI known as a Vision Transformer, which is designed to understand complex visual patterns. They showed the computer images of corn plots along with the exact time difference between when the photo was taken and when the corn flowered. Over time, the model learned to recognize the visual cues that signal a plant is approaching its flowering date, such as the orientation of the upper leaves or the subtle extension of the stem, long before the flower itself is visible to the human eye.

The results were surprisingly effective. The team tested their model on field experiments it had never seen before, including locations and years that were completely new to the system. Even when the model was given just a single image of a field taken within two weeks of the flowering date, it could predict the flowering time with an average error of only about three days. This level of accuracy is comparable to the precision of human observers walking the fields. When the researchers averaged predictions from two or three flights, the error dropped even further, often to less than two days. This suggests that breeders could achieve highly accurate results with far fewer drone flights than previously thought necessary, potentially reducing the cost and effort of monitoring crops by a significant margin.

The study also revealed how the computer "sees" the plants. By visualizing which parts of the image the model focused on, the researchers found that the AI's attention shifted dynamically as the plants grew. Before the flowers appeared, the model concentrated on the center of the plant where the flower would eventually emerge. As time passed, its focus moved to the surrounding leaves and the overall shape of the canopy. This indicates that the model is not just looking for a flower; it is understanding the entire developmental journey of the plant. It can detect subtle physiological changes that happen days before a human would notice anything different, and it continues to track the plant's state long after the bloom has occurred.

One of the most significant findings was that the model worked well even when the data came from different years, different locations, and different types of cameras. This suggests the system is robust enough to be used in new environments without needing to be retrained from scratch with local data. The researchers also tested whether using a temperature-based measurement of time, known as growing degree days, improved the results. While this method did offer a slight edge in accuracy, the model performed remarkably well using simple calendar days as well. This means that the system does not require complex weather data to function, making it easier to deploy in a wide variety of settings.

The implications for plant breeding are substantial. By reducing the need for daily flights and manual scoring, this method allows researchers to monitor much larger populations of crops with fewer resources. It opens the door to studying how different genetic varieties respond to flowering time across a wider range of environments, which is essential for developing crops that can adapt to a changing climate. The study demonstrates that a single drone flight, or perhaps just a few, can provide the same critical information that once required weeks of human labor. The model does not replace the breeder but rather acts as a powerful tool that extends their reach, allowing them to see the invisible rhythms of plant growth from the sky.