AI Counts Wheat Seedlings From Drone Images Using Only Dot Annotations

The model, called LGD-PeakNet, needs only tiny dot marks at the visible centers of seedlings as its training signal, yet it delivers both accurate plant counts and the individual locations of seedlings across drone imagery, according to a study published in Smart Agricultural Technology.

The work addresses a persistent bottleneck in agricultural remote sensing. Seedlings photographed from above are small, weakly textured targets embedded in cluttered backgrounds of soil, weeds, shadows, and uneven illumination. Traditional object detection approaches, such as YOLO or Faster R-CNN, can locate individual plants but demand expensive bounding-box annotations and degrade when targets shrink or overlap. Density-map regression methods, by contrast, are cheap to annotate but produce continuous response surfaces from which individual plant positions cannot reliably be recovered, limiting their usefulness for spatial distribution analysis.

Annotators mark a single pixel at the visual center of each distinguishable seedling, drastically reducing labeling effort. The network then learns to map images directly to predicted point sets, with the number of predicted points serving as the count estimate. The architecture proceeds through three stages: feature representation, feature fusion, and prediction regulation.

The front end is a dual-branch feature extractor. One branch, the local representation enhancement branch, uses Space-Channel Reconstruction Convolution to sharpen fine-grained seedling cues while suppressing redundant responses from repeated background textures. The other branch, a global feature extractor, employs a lightweight Transformer encoder to model long-range spatial context that local convolutions cannot capture. An adaptive multi-scale aggregation module then fuses these outputs, and a global attention mechanism recalibrates channel and spatial responses to concentrate feature activity on seedling-relevant regions.

The most distinctive innovation lies at the prediction end. Conventional point-supervised counters assign a fixed number of candidate slots to every spatial location, a design that fails when density varies sharply across a field: dense clusters need more candidates to separate adjacent plants, while sparse areas accumulate redundant false responses. The team’s Adaptive Sparse Point Response module solves this with a density-aware strategy. A lightweight density branch, supervised by Gaussian kernels generated online from the point annotations, estimates local seedling density. A smoothed density response is then converted through a sigmoid mapping into the number of active candidate slots per grid cell, ranging from one to eight, and only the top-ranked slots are retained for prediction.

To test the framework, the researchers built a Crop Seedling Dataset from UAV imagery collected with a DJI Mavic 3 flying at five meters over wheat plots and ten meters over maize, yielding ground sampling distances of 0.13 and 0.27 centimeters per pixel. Orthomosaic images were cropped into 600 by 600 pixel patches, producing four subsets: low-density drilled wheat averaging 81 seedlings per patch, high-density drilled wheat averaging 110, manually sown wheat averaging 76, and drilled maize averaging 123. Patches from the same orthomosaic were kept within a single data split to prevent spatial leakage between training and testing.

On the low- and high-density wheat test subsets, LGD-PeakNet achieved mean absolute errors of 6.86 and 8.14 seedlings per patch respectively, lower than all comparison methods evaluated, including CSRNet, CANet, Bayesian Loss, PET, P2PNet, P2P-CNF, RS-P2PNet, and APGCC. Its point-level matching F1 scores at an eight-pixel tolerance reached 0.926 on low-density wheat and 0.904 on high-density wheat, again the highest among the tested point-supervised methods. Similar improvements appeared on the manually sown wheat subset, where uneven clustering is common, and on the maize subset, which the authors emphasize served as a supplementary evaluation scenario rather than an independent test of cross-crop generalization, since maize training data participated in model optimization.

Ablation experiments on the low-density wheat subset clarified where the gains originate. Removing the ASPR module raised false positives on an extremely low-density test set of patches with fewer than ten targets from 115 to 405, pushing the false positive ratio from 15 percent to 38 percent. Comparisons of slot configurations showed that simply enlarging a fixed candidate pool did not match the performance of density-constrained dynamic allocation, confirming that the benefit comes from adaptive capacity regulation rather than raw capacity. The module ablations also revealed that the adaptive multi-scale aggregation module alone could even worsen performance, and only paid off once complementary local and global features and attention recalibration were in place.

Beyond counting, the point-level outputs enabled field-scale spatial analysis. The researchers divided an orthomosaic of the drilled wheat plots into a 75 by 32 grid of patches, counted predicted seedlings per cell, and generated emergence-level maps at fine, medium, and aggregated scales. Low-value cells showed spatial clustering, including continuous low-emergence bands along the planting direction, offering cues for locating problem zones and guiding field inspection. The authors caution that the model’s performance depends on density estimation quality, which can falter under severe occlusion in extremely dense regions, and that their data cover a limited range of locations, varieties, growth stages, and years, so broader validation remains necessary.

The study points toward a practical recipe for agricultural counting under weak annotation: pair enriched local-global feature representation with a prediction mechanism that adapts to the density structure of the scene itself. With computational cost close to the classical P2PNet baseline, at 20.43 million parameters and comparable training and inference times, the framework suggests that drone-based seedling monitoring could become both more affordable to label and more informative in output, extending from simple headcounts to maps of emergence variability that farmers and agronomists can act on.

Subject of Research: A point-supervised deep learning framework for counting wheat seedlings and mapping their spatial distribution from UAV imagery.

Article Title: LGD-PeakNet: a density-aware point-supervised framework for UAV-based wheat seedling counting

Article References: Zhang, B., Zhao, W., Wang, X., Xie, C., Wang, S., Che, Y., Shi, L., Ma, X., Xiong, S., & Xi, L. (2026). LGD-PeakNet: a density-aware point-supervised framework for UAV-based wheat seedling counting. Smart Agricultural Technology, 15, Article 102594. https://doi.org/10.1016/j.atech.2026.102594

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102594

Keywords: UAV remote sensing, wheat seedlings, point-supervised counting, deep learning, density estimation, precision agriculture, crop monitoring, drone imagery, seedling emergence, computer vision, LGD-PeakNet, density-aware