Plant stand count — the measure of how many plants per unit area have emerged and lived — sits at the foundation of crop management. Emergence is governed by seed quality, planting technology, soil conditions and weather, and an accurate count feeds directly into growth monitoring, yield prediction and resource allocation, including the high-stakes decision of whether a poorly established stretch of field should be replanted at all. The conventional method is brutally simple: field workers count plants by hand in selected plots. Beyond the labor involved, manual counting is slow, inefficient and incapable of delivering high spatial resolution across large fields, which means stand gaps and uneven emergence patterns often escape notice until they surface later in the yield map.

Uncrewed aerial vehicles fitted with high-resolution RGB cameras, working in concert with computer vision and deep learning, have already been deployed to count corn, cotton and rice, sweeping hundreds of hectares in a single flight and returning images at centimeter scale. Yet the models behind these systems carry two well-documented weaknesses. They are data-hungry: modern object detectors typically demand thousands of accurately labeled images, and large, precisely labeled farmland datasets are scarce and expensive to produce. And they are brittle. Farmland scenes shift dramatically in illumination, growth stage, shading, soil background and surface residue, and a detector tuned on one crop in one season frequently stumbles when pointed at another. Existing approaches split into two camps. Object detection draws a bounding box around every individual plant and excels in low-density stands, but degrades when canopies overlap or plants crowd together. Density estimation instead predicts pixel-wise density maps that are summed into a total count, a strategy that suits impenetrable canopies such as rice but has rarely been validated across crop types.

The Missouri team — Dinglun He, Fengkai Tian, Tianqi Yao and Jianfeng Zhou — attacked both weaknesses with transfer learning, the technique of pretraining a neural network on one large dataset and fine-tuning it on a small one so that hard-won knowledge carries over. Their framework unfolds in two phases. In Phase A, four object detectors were initialized from checkpoints pretrained on COCO, a generic archive of everyday photographs, then trained on a multi-year cotton dataset to absorb agricultural fundamentals: leaf edges, texture, seedling scale, row geometry and field backgrounds. The four architectures were chosen to be deliberately complementary. YOLOv8x sharpens feature extraction through a refined C2f module and an anchor-free decoupled head; YOLOv9e adds Programmable Gradient Information and a Generalized Efficient Layer Aggregation Network to fuse local detail with global structure; YOLOv10x dispenses with the computational overhead of non-maximum suppression altogether; and RT-DETR-X contributes a transformer-based hybrid encoder that applies self-attention to high-level features, giving it a different way of reading the spatial context of a crop scene.

The source data came from a 3.3-hectare cotton field at the university’s Fisher Delta Research, Extension, and Education Center near Portageville, Missouri, photographed across the 2022, 2023 and 2024 seasons with two platforms: a DJI Mavic 3M, whose 5280 × 3956-pixel camera was georeferenced in real time through a networked RTK satellite-correction service, and a Phantom 4 Advanced carrying a 5472 × 3648-pixel sensor. Flight heights of 10 to 15 meters produced ground sampling distances between 2.2 and 13.2 millimeters per pixel, with image overlap held between 70 and 80 percent. Corn and soybean imagery came from single-season 2024 flights over 16- and 22-hectare fields near Columbia, Missouri, in no-till, residue-heavy conditions that mirror commercial farming. All imagery was stitched in Agisoft Metashape into georeferenced orthomosaics, exported in the WGS 84/UTM zone 15N projection and cut with QGIS into fixed 640 × 640-pixel tiles so that plants would not straddle tile boundaries. From 1,277 annotated cotton tiles, 767 went to training, 255 to validation and 255 to testing; annotators boxed every confidently visible plant once, and ambiguous partial plants were excluded to keep training targets clean.

In Phase A, every detector learned cotton to a high standard. RT-DETR-X led with a test mAP@50 — mean average precision computed at a 50 percent intersection-over-union threshold — of 0.961, followed by YOLOv9e at 0.957, with hyperparameters selected through a multi-seed tuning protocol that reserved the test set for final reporting only. Then came the decisive experiment. For each target crop, the researchers imposed a strict annotation budget of just 30 original labeled training images, expanded to 90 stored images through offline augmentations such as flips and rotations, while validation and test sets of 20 images each remained fixed. Fine-tuning paired this small dataset with a grid search over learning rate and the proportion of frozen layers. On soybean, the cotton-pretrained RT-DETR-X topped the field at 0.894 mAP@50; on corn, YOLOv10x reached 0.948. Strikingly, no single architecture dominated both crops, which the authors read as evidence that cross-crop performance emerges from an interplay of source representations, target characteristics and tuning choices.

The most eye-catching numbers came from the label-budget experiment, which compared the cotton-pretrained YOLOv9e against the identical architecture fine-tuned straight from its generic COCO checkpoint using nested subsets of 1, 3, 5, 10 and 30 target-crop images. With a single labeled soybean image, cotton pretraining raised mAP@50 from 0.189 to 0.860 — a leap that took the model from nearly blind to field-ready after seeing one photograph of one field. On corn, the same comparison improved accuracy from 0.135 to 0.527. The explanation lies in what the cotton checkpoint already knew: the apparent size of seedlings viewed from a drone, the geometry of planted rows, the visual texture of soil and crop residue, and the characteristic appearance of stitched orthomosaic imagery. The COCO baseline, by contrast, had to reconstruct all of that agricultural context from a handful of examples, leaving it exquisitely sensitive to whichever few images it was given. As annotations accumulated, the gap closed, and at 30 images the COCO baseline edged ahead by 0.005 mAP@50 on soybean and 0.003 on corn — a possible whisper of mild negative transfer, though the authors caution that the margins are tiny and rest on a single fixed data split, so statistically reliable negative transfer was not established.

Counting plants tile by tile is only half the job; the pipeline also converts detections into maps a grower can act on. For inference across full orthomosaics, the team employed Slicing Aided Hyper Inference, or SAHI, which slides a 640 × 640-pixel window with 20 percent overlap in both directions across the image, discards predictions scoring below 0.35 confidence and merges duplicated detections from adjacent slices through greedy non-maximum merging based on intersection over smaller area. Each surviving bounding box then reveals its position: the box center is transformed from image coordinates into projected map coordinates using the six-coefficient affine geotransform of the GeoTIFF, which encodes the raster’s origin, pixel size, rotation and orientation. Every detected plant becomes a georeferenced point in WGS 84/UTM zone 15N, assigned a unique identifier. Loaded into any GIS environment, these point layers can be aggregated by plot, row segment or management zone, exposing stand gaps and uneven emergence that no per-image score could reveal. The authors note candidly that because no independently surveyed check points were collected, the maps demonstrate deployment capability rather than validated absolute positional accuracy.

The entire workflow has been released as an open-source Python pipeline organized into five modules that mirror the study’s stages: orthomosaic preprocessing and image export, annotation and dataset organization, Phase A source-domain training on cotton, Phase B target-domain fine-tuning, and full-orthomosaic inference with georeferenced point export. All models ran on an NVIDIA RTX A6000 graphics card using stochastic gradient descent, with source-domain training capped at 300 epochs and fine-tuning at 500, early stopping guarding against overfitting in both regimes. For practitioners adapting the system to new fields, the authors flag the dials most likely to need adjustment: ground sampling distance, slice size and overlap, the detection confidence threshold, the prediction-merging threshold and the number of target-crop images labeled for fine-tuning. They also propose a pragmatic annotation strategy: label a small set of target-crop images first, check whether the transferred model already performs adequately, then direct additional labeling toward the conditions where it remains weak, such as dense emergence, heavy surface residue, shadows or unusual seedling morphology. The work was supported by the United Soybean Board and the U.S. Department of Agriculture’s National Institute of Food and Agriculture.

The researchers are forthright about the study’s limits. Excluding partially visible plants during annotation, while it improved consistency, left some visible-but-unlabeled regions to be treated as background during training — a rule that can bias detectors toward undercounting near tile boundaries and in densely emerged stands; future iterations may adopt clipped-object annotations, ignore regions or boundary-aware training procedures. The evaluation also rested on a single fixed set of data splits, and absolute positional accuracy awaits validation against independently surveyed plant locations. Even so, the central lesson stands: the true bottleneck in agricultural artificial intelligence is often not the algorithm but the labeling budget, and pretraining inside the agricultural domain multiplies the value of every labeled image a farmer or agronomist can afford to produce. A cotton field in Missouri, it turns out, taught a machine to count far more than cotton.