Few-shot learning enables large-scale annotation of trapped insect images

Every night, across farmlands in China and much of the world, light traps lure nocturnal insect pests toward glowing panels, where they are photographed, counted, and in many designs, eliminated by infrared heat. These intelligent pest-monitoring stations promise a revolution in agriculture: instead of blanket pesticide spraying, farmers could respond precisely to the pests that machine vision systems detect in real time. But beneath this elegant concept lies a stubborn bottleneck that has frustrated agricultural technologists for years. Before any artificial intelligence model can recognize a moth from a beetle, or distinguish a beneficial insect from a devastating crop pest, someone must teach it—and that teaching requires enormous, meticulously annotated datasets of insect images.

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.