The new framework, described in the journal Smart Agricultural Technology, tackles a problem that has long frustrated agricultural engineers. Sugar beet is one of Europe’s most valuable arable crops, with France and Germany leading production according to the United Nations Food and Agriculture Organization, yet a meaningful share of every harvest is degraded by mechanical damage between the soil and the truck. Because the damage happens inside a machine moving at full field speed, farmers have had no practical way to see it, measure it, or respond to it. The researchers, led by Konstantinos Gkountakos of the Centre for Research and Technology Hellas, set out to give harvesters something close to a nervous system: cameras, neural networks, and the ability to distinguish a healthy beet from a broken one in real time.

The journey from field to tank inside a modern harvester is short but violent. First, defoliators gently remove the leafy tops; cut too high and leaves remain to regrow, cut too low and the head of the beet, which is rich in sugar, is shaved away. Next, lifting shares pry the roots from the ground, ideally without cracking them. Then come the cleaning turbines, cylindrical drums that spin the beets against one another to knock off soil, stones, and debris. Residual soil inflates transport costs and accelerates erosion, but overly aggressive cleaning shatters and abrades the roots. Finally, a metallic sieve belt carries the cleaned beets into the tank. At the end of this conveyor, just before loading, the researchers saw an opportunity: a natural inspection point where every beet passes the camera, and where a damaged crop could still trigger adjustments to the lifting and cleaning mechanisms upstream.

To train and test their system, the team created something the field has lacked: a genuinely realistic dataset, which they named VerBeet. Using an industrial Basler camera with 8.3 megapixel resolution mounted in a white housing above the sieve belt of a Vervaet Q-621 harvester, they captured 694 frames over five harvesting days between early December 2025 and early January 2026, spaced 60 to 75 seconds apart to avoid overlapping content and to sample entire fields. Because daylight shifted with the harvester’s position relative to the sun, producing frames that ranged from dark to glare-bright, they installed three 40-watt LED lights on magnetic mounts, providing 13,200 lumens of consistent illumination. Two human annotators then spent roughly 17 hours drawing bounding boxes around every beet, cross-validating each other’s work, ultimately labeling 5,158 individual beets across three categories: undamaged, breakage, and abrasion.

The class distribution itself tells a story about mechanized harvesting. As expected, undamaged beets dominate with 3,685 instances, but 1,293 beets showed breakage, structural damage to the root from lifting or cleaning, and 180 displayed surface abrasion. The imbalance between the two damage types proved instructive: abrasions, the researchers found, are often hidden by clinging soil or simply invisible from a single camera angle, whereas a broken beet announces itself regardless of perspective. This matters because the ultimate goal is not just detection but attribution, knowing whether damage originated in the topping, lifting, or cleaning stages, so the harvester can adjust the right component rather than guessing.

The framework itself operates in two stages. A state-of-the-art object detector first localizes every beet in each frame; the researchers benchmarked lightweight members of the YOLO family, including YOLOv11n, YOLOv12n, and the newest YOLO26, against the transformer-based RT-DETR. Each cropped beet is then passed to a convolutional neural network classifier, drawn from EfficientNet, DenseNet, and MobileNet architectures, that decides whether the root is whole, broken, or abraded. Speed was non-negotiable: YOLOv11n processed more than 160 frames per second on the VerBeet test set, while MobileNetV2 classified beets at around 130 instances per second on modest consumer-grade hardware, confirming the pipeline could run on the edge devices that a real harvester would carry.

Two clever training tricks pushed performance further. The first is a size-aware weighted loss. Because the camera always views the belt from the same distance, a beet’s apparent size in the image reveals how much of it is visible; occluded or partially buried beets appear smaller and are harder to classify. The loss function weights each training sample by the inverse of its normalized area, so smaller, harder instances contribute more to learning, with weights clipped between 0.1 and 1.0 to keep training stable. The second trick is generative: the team used FLUX.2-4B Klein, a lightweight open-source diffusion model, to synthesize augmented training images in which each beet is placed on a randomized background. This counteracts overfitting to the unchanging metallic belt behind the beets and, crucially, multiplies scarce examples of the rare abrasion class. Every synthetic image was manually inspected, and hallucinated samples where the model invented or erased beets were discarded.

The results reveal both the promise and the stubborn difficulty of real-world machine vision. On the controlled, laboratory-style Semantic Sugar Beet dataset, models reached F1-scores above 87 percent, but on VerBeet’s dusty, occluded, unevenly lit frames, the best binary damage detection, EfficientNetB3 with size-aware loss and diffusion augmentation, achieved an F1-score of 71.48 percent. Cross-dataset experiments drove the point home: models trained on the controlled dataset collapsed when tested in the field, with recall as low as 12.7 percent, while models trained on VerBeet generalized far better, reaching 74.6 percent mAP on the cleaner data. In the most demanding test, a field-disjoint end-to-end evaluation where the test set always contained beets from a harvesting day entirely unseen during training, the full pipeline of YOLOv11n and MobileNetV2-Size achieved a mean F1-score of 67.09 percent across five folds, distinguishing beets, breakage, and abrasion in realistic conditions.

The researchers are candid about limitations. The dataset came from a single harvester model, so cameras mounted at different angles or heights on other machines could shift the statistics the networks rely on. Abrasion remains the weakest link, with multiclass F1-scores hovering near 51 percent, limited by the scarcity of genuine examples and the visual similarity between abraded and broken surfaces. Future work, the authors suggest, will extend VerBeet to more harvesting conditions, beet varieties, and even other root crops such as potatoes, and will explore adapting modern vision foundation models to withstand occlusion, dust, and erratic light under real-time constraints on semi-autonomous harvesting platforms.

Still, the trajectory is clear. Agriculture is racing toward machines that perceive, decide, and react, and the sugar beet harvester, long a blunt instrument of force and throughput, is acquiring finesse. A future machine that detects rising breakage rates and automatically softens its lifting shares, or notices abraded beets and eases the turbine speed, would translate directly into tonnes of sugar saved and soil left in the field where it belongs. The VerBeet dataset, with its 5,158 honestly annotated beets captured in the mud and glare of actual harvest days, may well become the benchmark on which that generation of gentle machines is trained.

Subject of Research: Real-time computer vision detection of sugar beet damage during mechanical harvesting

Article Title: Sugar beet localization and damage detection during harvesting

Article References: Gkountakos, K., Pasios, S., Ioannidis, K., Demestichas, K., Vrochidis, S., & Kompatsiaris, I. (2026). Sugar beet localization and damage detection during harvesting. Smart Agricultural Technology, 15, Article 102560. https://doi.org/10.1016/j.atech.2026.102560

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102560

Keywords: sugar beet, computer vision, deep learning, YOLO, damage detection, precision agriculture, harvesting, VerBeet dataset, diffusion models, object detection, agricultural technology, neural networks

Creating those datasets the traditional way is a grind that consumes the patience of even the most dedicated entomologists. Each image captured by a light trap may contain dozens of insects at wildly different sizes, overlapping bodies, broken wings, and ambiguous orientations. An expert must draw a box around every insect, assign it to the correct species or category, and repeat this thousands upon thousands of times. The work is slow, expensive, and entirely dependent on scarce specialists whose time might otherwise be spent on actual research. A team of researchers from Zhejiang Sci-Tech University and collaborating institutions in China has now unveiled a solution that could dissolve this bottleneck, and their results, published in the Journal of Integrative Agriculture, suggest that the era of exhaustive manual insect annotation may be drawing to a close.

The team’s approach is a two-stage framework that separates the problem of dataset construction into detection and classification, tackling each with a purpose-built system. The first stage, called MLTIDD—short for multi-scale light-trapped insect detection and description—addresses one of the most technically thorny challenges in agricultural computer vision: the extreme disparity in scale between large and tiny insects within a single image. A large moth might occupy hundreds of pixels, while a minuscule midge may span barely a dozen. Standard object detectors, which rely on receptive fields tuned to typical object sizes, systematically miss the smallest specimens or fragment the largest ones.

To overcome this, the researchers fine-tuned Grounding DINO, a vision-language foundation model capable of open-vocabulary detection, meaning it can locate objects based on text descriptions rather than requiring exhaustive training on fixed categories. But foundation models alone were not enough. The team integrated SAM, the Segment Anything Model developed originally for general-purpose image segmentation, and SAHI, a slicing-aided hyper-inference technique that chops large images into overlapping tiles, runs detection on each tile, and then stitches the results back together. This combination allows the system to detect insects at multiple scales simultaneously: SAHI ensures that tiny insects magnified within tiles are not overlooked, while the fine-tuned Grounding DINO provides semantic understanding of what constitutes an insect even in cluttered, chaotic light-trap scenes.

The performance gains were substantial. In experiments conducted with only limited annotated samples—a few-shot setting that mirrors the real constraints of agricultural research labs—MLTIDD achieved 79.6 percent average precision at the strict AP50–95 threshold and 90.8 percent average recall. Compared against a standard DINO baseline, this represented improvements of 7.0 and 4.7 percentage points respectively. When the complete pipeline was evaluated against the combined DINO and iBOT setup, the detection stage improved mean AP50–95 by 10.91 percentage points and average recall by 8.26 percentage points. For a field where even a few percentage points can determine whether a monitoring system is practical or prone to dangerous miscounts, these are meaningful margins.

Detection, however, is only half the battle. Once thousands of insect sub-images have been sliced out of the original photographs, each one must be classified. This is where the second innovation, InsectSSRL, comes into play. Standing for insect self-supervised representation learning, InsectSSRL is built on iBOT, a self-supervised vision transformer framework that learns by masking parts of an image and training the network to predict what is hidden. Crucially, it requires no labels at all during pre-training—instead, it exploits the massive trove of unlabeled insect sub-images that MLTIDD itself generates, turning the detection stage’s output into fuel for the classification stage’s learning.

InsectSSRL enhances its feature extraction through three proxy tasks, teaching the model to capture the fine-grained visual signatures that separate insect species: wing venation patterns, body segmentation, coloration, and texture. The result is a feature extractor that, when evaluated with a k-nearest-neighbors classifier—a test of how well the learned features cluster similar insects together—achieved 85.87 percent top-1 accuracy. The true measure of its value emerged in few-shot classification experiments. When a Swin-T vision transformer pre-trained with InsectSSRL was fine-tuned on just 5 percent of the InsectID dataset, it reached 80.35 percent classification accuracy. That figure exceeded an iBOT baseline by 2.08 percentage points and surpassed a model trained with conventional transfer learning from COCO, a generic object-detection dataset, by a striking 11.3 percentage points. The message is clear: features learned from insect imagery, in a self-supervised manner, transfer to insect classification far better than features borrowed from unrelated domains.

The complete pipeline works as a virtuous cycle. MLTIDD detects and crops insects from raw light-trap images. InsectSSRL consumes those unlabeled crops to learn robust representations, then supports a classification model that pre-classifies each insect. Human experts then step in—but only to correct and verify the machine’s suggestions rather than to label everything from scratch. Finally, the corrected labels are traced back to the original images, completing the annotation of the entire dataset. This human-in-the-loop design preserves the quality guarantee that only expert knowledge can provide, while eliminating roughly 80 percent of the time experts would otherwise spend on manual labeling.

That 80 percent reduction is the headline number with the most immediate practical consequence. Light-trap monitoring networks can generate millions of images per season across wide geographic areas. If every image demanded full manual annotation, scaling these networks to national or continental coverage would be economically impossible. By slashing expert workload to a verification role, the new framework makes large-scale, multi-class insect datasets achievable by ordinary research teams. Better datasets, in turn, mean better AI models in the field—models that can distinguish pest outbreaks from benign insect activity with the reliability that precision agriculture demands.

The implications extend well beyond pest monitoring. Light traps are also used in biodiversity surveys, ecological monitoring, and studies of insect decline, where automated species identification is equally hampered by annotation scarcity. A framework that can bootstrap a high-quality dataset from a small number of labeled examples and a mountain of unlabeled images is exactly the kind of tool those fields need. The combination of open-vocabulary foundation models, segmentation models, tiled inference, and self-supervised learning represents a template that could be adapted to any domain where objects are small, numerous, and taxonomically diverse—from plankton imaging to medical cell counting.

There are, of course, caveats worth noting. The framework’s classification accuracy, while impressive, still falls short of the near-perfect reliability that fully autonomous pest-control decisions would require, which is precisely why expert correction remains embedded in the pipeline. And the approach’s reliance on fine-tuned foundation models assumes access to computational resources that some labs may lack. Yet the trajectory is unmistakable. As corresponding author Professor Qing Yao of Zhejiang Sci-Tech University puts it, the method improved label quality, decreased expert workload, and increased annotation efficiency simultaneously—a rare trifecta in dataset engineering, where improvements in one dimension typically come at the expense of another.

What began as a practical problem—too many insects, too few annotators—has yielded a demonstration of how modern vision foundation models can be orchestrated into something greater than the sum of their parts. Grounding DINO finds the insects, SAM delineates them, SAHI ensures the smallest are not lost, and InsectSSRL learns to name them, all while human experts supervise from a comfortable distance. The moths still fly into the light. But now, the machines that watch them are teaching themselves to see.