The work, published in the journal Smart Agricultural Technology, tackles a deceptively hard computer vision problem. Wheat seedlings at the one-to-three-leaf stage are small objects with slender leaves that point in every direction. Neighbouring plants frequently overlap, blurring the boundaries between individuals, while soil patterns, shadows and the occasional weed can masquerade as narrow green leaves. Conventional convolutional neural networks, which process images through local filters, struggle to capture the long-range spatial relationships needed to separate densely packed plants. Simply making a network bigger does not solve the problem either; the researchers found that a much heavier model, ResNet50, with 25.34 million parameters, delivered no meaningful accuracy gain over far smaller architectures on this task.

WheatNet’s designers, led by Yafeng Han and colleagues, built their detector on YOLO11n, one of the smallest members of the widely used YOLO real-time detection family, and then re-engineered all three of its main components: the backbone that extracts features from the image, the multi-scale neck that fuses features at different resolutions, and the loss function that guides training. The result is a model with just 3.83 million parameters and a computational cost of 9.6 billion floating-point operations, yet it outperformed seven established detectors, including the transformer-based RT-DETR-N, which is nearly three times larger and three times slower.

The backbone upgrade is the most technically ambitious piece. The team adapted Vision-RWKV, an architecture originally designed to model global sequences efficiently, and enhanced it into what they call a Gated-Enhanced Vision-RWKV, or GE-VRWKV. The module flattens the image’s feature map into a sequence and models interactions among all spatial positions, which helps the network understand how densely distributed seedlings relate to one another. Two additions make it wheat-ready: an efficient channel attention mechanism that recalibrates feature channels to emphasise responses associated with seedling edges, leaf tips and elongated leaf textures, and a learnable scalar gate that controls how strongly the enhanced features flow into the residual pathway. Placed at the two feature scales most relevant to small seedlings, this backbone alone lifted the model’s mAP@0.5 score by 1.29 percentage points over the baseline.

The second innovation, Rotation-Enhanced Area Attention, addresses the fact that wheat leaves extend horizontally, vertically and diagonally. The module runs four lightweight directional branches in parallel, corresponding to zero, 45, 90 and 135 degrees. Horizontal and vertical responses come from one-dimensional depthwise convolutions, while the diagonal branches use a clever trick: instead of learning rotated kernels, they cyclically shift the feature map along the diagonals before applying a small convolution, capturing oblique structure at a fraction of the computational cost. An attention branch then estimates, at every spatial location, how much each of the four directions should contribute, and the branches are fused by weighted summation. Visualisations of these attention weights show the 90-degree branch responding strongly to vertical leaves and the diagonal branches lighting up in oblique and overlapping leaf regions, exactly as the design intended. The module added only 0.10 million parameters yet improved mAP@0.5 by 1.47 points over a neck with no attention at all.

The third component targets a subtle but consequential problem in small-object detection. When a bounding box covers only a few pixels, a shift of even one pixel can cause a large change in the intersection-over-union metric normally used to score localisation, making training unstable. WheatNet combines two complementary losses in equal measure: Wise-IoU, which adaptively down-weights low-quality and ambiguous regression samples, and the normalised Wasserstein distance, which treats bounding boxes as distributions and measures their similarity in a way that is far less sensitive to tiny coordinate deviations. A sensitivity analysis over five weighting combinations confirmed that a 50-50 split performed best, and the joint loss improved mAP@0.5 by 1.16 points over the baseline without adding a single parameter.

None of this would have been possible without a purpose-built dataset, and the team’s careful methodology here is itself instructive. They collected more than 5,000 candidate images from experimental fields at four sites in Henan Province across two growing seasons, using smartphones and digital cameras held just 10 to 50 centimetres above the ground. Preliminary experiments comparing shooting angles showed that low-angle views of 15 to 30 degrees beat top-down and high-angle shots by a wide margin, because they reveal leaf edges, stem bases and the boundaries between neighbouring plants more clearly. Annotation strategy mattered just as much: drawing boxes around each seedling’s complete visible structure, rather than just its main body, raised mAP@0.5 from 77.1 percent to 89.1 percent. After removing duplicates with cryptographic and perceptual hashing, the final WheatSC dataset contained 2,895 images with 35,642 annotated seedling instances, expanded to more than 158,000 instances through randomised augmentation of the training split alone.

The headline numbers are striking. On the internal test set, WheatNet achieved 95.22 percent precision, 95.72 percent recall, 92.70 percent mAP@0.5 and 82.82 percent mAP@0.5:0.95, beating its YOLO11n baseline by 3.98 and 4.86 percentage points respectively. For counting, the metric farmers ultimately care about, it recorded a mean absolute error of 0.972 seedlings per image, a mean absolute percentage error of 8.59 percent, and a coefficient of determination of 0.893, with predicted counts clustering tightly around the one-to-one line against manual tallies. Crucially, the advantages held up under scrutiny: in a site-held-out experiment where the model was tested entirely on images from a location it had never seen, every paired difference remained statistically significant under bootstrap analysis, and across three independent training runs the standard deviations stayed below one percentage point.

Subgroup analyses revealed both the model’s strengths and the field’s remaining challenges. Performance declined for every model under dense planting, heavy occlusion, challenging lighting and complex backgrounds, but WheatNet outperformed the baseline in every single subgroup, including a dense-scenario mAP@0.5 of 91.81 percent against the baseline’s 89.63 percent. To demonstrate practical deployment, the researchers converted the model to ONNX format and built a WeChat mini-program prototype that marks each detected plant with a bounding box and reports an estimated count; on an iPhone 17 Pro, a single image was processed in roughly 200 milliseconds. At 156 frames per second on the team’s GPU test platform, the model also leaves headroom for edge devices and drones.

The authors are candid about limitations. The dataset primarily represents weed-controlled plots at the one-to-three-leaf stage, so performance under heavy weed infestation or at later growth stages remains untested, and images from the same site or date may not be statistically independent, meaning the evaluation units are observational rather than true biological replicates. Future work, they write, should span additional regions, seasons, cultivars and more severe occlusion, with systematic deployment tests on mobile and edge hardware. Even so, WheatNet marks a meaningful step toward a long-sought goal in digital agriculture: turning the humble seedling count, a task that once demanded hours of bent-backed labour, into a two-hundred-millisecond snapshot taken with a phone in the field.

Subject of Research: Deep learning-based detection and counting of wheat seedlings in field images

Article Title: WheatNet: A lightweight network for wheat seedling detection and counting in complex fields

Article References: Han, Y., Liu, B., Wang, X., Zhao, H., Li, X., Gao, G., Chen, L., & Guo, X. (2026). WheatNet: A lightweight network for wheat seedling detection and counting in complex fields. Smart Agricultural Technology, 15, Article 102648. https://doi.org/10.1016/j.atech.2026.102648

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102648

Keywords: WheatNet, wheat seedlings, deep learning, object detection, precision agriculture, plant phenotyping, YOLO, computer vision, crop counting, lightweight neural network, smart farming, dataset annotation