AerialYield-B2D: A Greenhouse Blueberry Dataset with Five-Stage Ripeness Masks and Fruit Counts

In the quiet, climate-controlled rows of a modern greenhouse, the path to a perfect blueberry harvest is not a straight line. It is a visual puzzle. A grower does not simply look for a single blue fruit; they must assess a cluster where green, pink, purple, and deep blue berries often grow side by side, hidden among leaves and shadows. This visual complexity is the daily reality of agriculture, where the decision to pick a fruit depends on its specific stage of ripeness. For decades, researchers have relied on human eyes to make these judgments, but the rise of automated robots and computer vision systems demands something more precise: a way to teach machines to see exactly what a human sees. To do this, computers need a library of examples, a vast collection of real-world images where every single berry is marked and categorized by its color and maturity. Without such a resource, the technology remains blind to the subtle differences that define a ripe crop.

This is the challenge addressed by a new resource called AerialYield-B2D, a carefully curated dataset designed to help computers learn the language of blueberry ripeness. Created by a team of researchers at Khalifa University in the United Arab Emirates, this collection brings together 514 real photographs taken inside a commercial greenhouse in Al Ain. The images capture the messy, natural reality of the crop, showing thousands of berries in five distinct stages of development: green and unripe, pale pink, turning purple, fully ripe, and over-ripe. Unlike previous attempts that might rely on simplified drawings or isolated fruits, this dataset presents the fruit exactly as it grows, tangled in clusters and partially obscured by foliage. The researchers did not just take pictures; they spent countless hours manually tracing the outline of every single berry in the photos, creating a digital map that tells a computer exactly where each fruit is and what color it is. In total, they annotated more than 30,000 individual berries, providing a level of detail that allows machines to practice distinguishing between a fruit that is ready to eat and one that is still waiting to ripen.

The creation of this dataset was a deliberate effort to capture the full spectrum of how blueberries are actually viewed in a farming environment. The team gathered images from two different sources to ensure the data was robust. The majority came from close-up photographs taken with smartphones, which capture the fine details of the fruit's skin and the wax that coats it. To add variety and simulate the view from a robot or a drone moving down the rows, they also included frames extracted from high-definition videos and footage captured by a small flying drone. This mix ensures that the computer models trained on this data can handle different angles, lighting conditions, and distances. The researchers were careful to record the source of every image, noting whether it came from a phone or a drone, so that future users can test if their systems work equally well on all types of pictures. The entire collection was processed with strict attention to detail, correcting the orientation of the photos so that the digital maps align perfectly with the real fruit, and verifying that no images were missing or corrupted.

What makes this work particularly valuable is its honesty about the difficulties of the task. The researchers openly acknowledge that the data reflects the natural imbalance of a real crop. In the greenhouse, green berries vastly outnumber the rare, over-ripe ones, and this uneven distribution is preserved in the dataset rather than artificially fixed. This means that anyone using the data to train a computer will have to grapple with the same challenge a farmer faces: finding the rare, perfect fruit among a sea of unripe ones. The team also clarified what the dataset is not. While the name includes the word "yield," the images do not provide measurements of weight or total harvest volume. Instead, the numbers provided are strictly counts of how many berries appear in each picture. This distinction is crucial, as it sets clear boundaries for what the data can tell us: it is a tool for counting and identifying, not for weighing the final harvest.

To prove that the data is useful, the researchers tested it by feeding it into standard computer vision models. They asked these models to learn from the training images and then predict the ripeness of berries in images they had never seen before. The results showed that the models could successfully identify the different stages of ripeness, though they struggled more with the rare colors, such as the pale pink or the over-ripe orange berries, simply because there were fewer examples of them to learn from. The team also tested the system's ability to count the total number of berries in a single image, achieving a level of accuracy that suggests the dataset is a solid foundation for future work. By releasing the images, the detailed maps, and the code used to process them, the team has provided a reproducible benchmark. This means other scientists can use the exact same images and rules to test their own ideas, ensuring that progress in this field is built on a shared, reliable foundation.

The release of AerialYield-B2D represents a significant step forward in the intersection of agriculture and artificial intelligence. It moves beyond the theoretical and provides a tangible, real-world resource that mirrors the complexity of a living crop. By focusing on the specific, visual nuances of blueberry ripeness, the dataset offers a clear path for developing tools that can assist growers in making faster, more accurate decisions. It does not promise to solve every problem in farming, nor does it claim to replace human judgment entirely. Instead, it offers a precise, verified set of examples that allows machines to learn the visual language of the blueberry, one annotated berry at a time. In doing so, it turns a chaotic, natural scene into a structured lesson, helping to bridge the gap between the human eye and the computer's lens.