YOLO-Based Deep Learning for Citrus Fruit Detection, Counting, and Yield Estimation in Complex Orchard Environments: A Systematic Review

Imagine a grove of orange trees, their branches heavy with fruit, hidden beneath a canopy of leaves. For a human worker, counting every single orange in such a scene is a slow, exhausting task prone to error; one person might count a fruit twice, while another misses one tucked deep in the shade. For farmers, getting this count wrong means misjudging how much fruit they will harvest, which throws off everything from hiring labor to planning how many trucks are needed. To solve this, scientists have turned to cameras and computers, teaching machines to "see" the fruit just as a human would. The most popular tool for this job is a type of software called YOLO, which stands for "You Only Look Once." It is a fast, efficient system designed to scan an image and instantly spot objects, making it ideal for robots or drones that need to make split-second decisions while moving through an orchard. But as these tools have become more common, the field has become crowded with hundreds of different studies, each claiming to have found a better way to count citrus. The question remains: are these systems truly ready for the real world, or are they just good at performing in a controlled lab?

A team of researchers set out to answer this by taking a comprehensive look at nearly two hundred scientific studies that used these fast detection systems to find and count citrus fruit. They did not simply read the headlines; they dug into the details of ninety specific studies, applying a strict set of rules to see what was actually happening behind the numbers. Their goal was to understand not just how well the computers performed, but how the researchers tested them and whether the results could be trusted. They found that the technology has indeed matured. When these systems are tested on images of fruit, they are remarkably accurate, correctly identifying the vast majority of oranges, lemons, and mandarins they see. On average, the systems found about ninety percent of the fruit that was actually there and were correct about ninety percent of the time when they said they had found something. This level of performance suggests that the core technology is working well enough to be useful.

However, the researchers discovered that the story is more complicated than the high success rates suggest. While the computers are good at spotting fruit in pictures, the way scientists have been testing them is often flawed. In nearly every study reviewed, the researchers ran their tests only once, without repeating the experiment to see if the results were consistent or just a lucky fluke. Furthermore, most of the studies relied on very small collections of images, often taken from just one orchard under perfect lighting conditions. This is like testing a car only on a smooth, empty track in the summer and then assuming it will handle a snowy mountain road just as well. The researchers found that when fruit is hidden behind leaves, or when the green fruit blends in with green leaves, the computers struggle significantly. They also noted that very few studies actually tested their systems on real robots or drones out in the field; most were tested on powerful computers in a lab, which is a very different environment from a dusty, unpredictable orchard.

The review also highlighted that the field has settled on a common set of tools. Most researchers are using the same few versions of the detection software and are making similar tweaks to improve them, such as adding special modules to help the computer focus on important parts of an image or making the software lighter so it can run on smaller devices. Yet, despite these shared methods, there is no single standard dataset that everyone uses to compare their results. It is as if every chef in the world is trying to prove they make the best soup, but no two chefs are using the same ingredients or the same recipe, making it impossible to know who is truly the best. The researchers found that only a tiny fraction of studies used advanced cameras that can see more than just visible light, such as sensors that detect heat or specific colors invisible to the human eye, which could help distinguish fruit from leaves more easily.

Ultimately, the study concludes that while the ability of computers to detect citrus fruit has reached a high level of accuracy in controlled settings, the field is not yet ready for widespread, reliable use in real-world farming. The technology itself is not the main problem; the issue lies in how it is validated. The researchers argue that future work must focus on testing these systems in diverse, difficult conditions and on actual hardware used in the field, rather than just in the lab. They call for a shared standard for testing so that different systems can be fairly compared, and for more rigorous statistical checks to ensure that the results are repeatable. Until these steps are taken, the promise of fully automated fruit counting and yield estimation will remain just out of reach, waiting for a more solid foundation of evidence to support it.