MSDD: Maize Seedling Detection Dataset

• MSDD is a curated aerial RGB image dataset for maize seedling detection, stratified into single, double, and triple plant groupings.
• It benchmarks state-of-the-art object detectors like YOLOv9, YOLO11, YOLOv12, and Faster-RCNN under diverse field conditions.
• The dataset addresses challenges including class imbalance, annotation ambiguity, and variable imaging conditions for precision agriculture.

Searching arXiv for MSDD and closely related maize seedling datasets to ground the article in the cited literature. The Maize Seedling Detection Dataset (MSDD) is a curated high-resolution RGB aerial image dataset introduced for seedling maize detection and benchmarking, with primary use in maize stand counting, early-season crop monitoring, yield prediction, and in-field management (Kharismawati et al., 18 Sep 2025). Its central design choice is explicit stratification of maize groupings into single, double, and triple plants under heterogeneous field conditions, so that object detectors can be evaluated not only on ordinary emergence patterns but also on clustered planting outcomes associated with stand-count errors. Within the reported benchmarks, MSDD is used to compare YOLOv9, YOLO11, YOLOv12, and Faster-RCNN, and it is positioned as a public resource for detection research in precision agriculture rather than as a species-classification corpus or a 3D phenotyping dataset (Kharismawati et al., 18 Sep 2025).

1. Dataset identity and problem setting

MSDD is defined as a dataset for maize seedling detection from aerial RGB imagery, with an emphasis on stand counting. In this context, stand counting determines how many plants germinated and is linked in the source paper to decisions such as replanting or adjusting inputs. The dataset was assembled to address the scarcity of curated resources for this task and to support benchmarking under real field variability rather than controlled laboratory conditions (Kharismawati et al., 18 Sep 2025).

A defining property of MSDD is that its labels are not limited to presence or absence of maize. Instead, each annotated object is categorized as a single plant, a double plant cluster, or a triple plant cluster. This makes the dataset suitable for studying failure modes that arise when multiple seedlings are spatially adjacent or visually entangled. A plausible implication is that MSDD is structured to expose both classical object-detection performance and the agronomically specific problem of counting grouped seedlings, which is more demanding than simple crop-versus-background discrimination.

The paper describes the dataset as the first large-scale, high-resolution, and diverse dataset for maize seedling detection under varied field conditions. It further presents MSDD as a benchmark resource for current object detectors and as an operational substrate for near-real-time agricultural monitoring (Kharismawati et al., 18 Sep 2025).

2. Image acquisition and annotation workflow

MSDD is built from RGB aerial imagery acquired with DJI Phantom 4 and DJI Mavic 2 Pro platforms, with image resolutions of $4096 \times 2160$ px and $3940 \times 2160$ px, respectively (Kharismawati et al., 18 Sep 2025). Flights were manual and used the Autopilot or DJI Go 4 applications. Acquisition covered altitudes from 1 to 18 meters above ground level, and the paper reports that at 4.5 m the ground resolution is approximately 0.5 cm/pixel. Both nadir and oblique views were intentionally included.

The imagery was collected from Missouri maize genetic nurseries during 2019–2022. The fields include manual and machine planting regimes, and the plant material spans inbred, mutant, and elite hybrid lines. Frames were extracted from video at 24 or 30 fps with different sampling rates and then divided into $1920 \times 1920$ px patches for model input. Images were captured at different times of day and under sunny, cloudy, and windy conditions, so acquisition variability is part of the dataset design rather than an incidental artifact (Kharismawati et al., 18 Sep 2025).

The annotation protocol combines manual and semi-automatic stages. Initial frames were manually labeled in YOLO_mark with one bounding box per maize grouping. These manual labels were then used to train a preliminary YOLO model that generated tentative labels for subsequent frames, followed by manual correction. The workflow also used homography-based labeling, in which bounding boxes were projected between overlapping sequential frames using computed homography matrices. Labels are distributed either as YOLO text files with class and normalized bounding box coordinates or as COCO-style JSON for RCNN-based workflows. This annotation pipeline indicates that MSDD was curated not only for scale but also for annotation throughput and consistency (Kharismawati et al., 18 Sep 2025).

3. Composition, stratification, and scene variability

The dataset contains 163,921 annotated objects and is explicitly imbalanced toward the single-plant class (Kharismawati et al., 18 Sep 2025). The class structure is as follows:

Class	Definition	Share of annotated objects
Single	One seedling in a box	92.47%
Double	Two seedlings clustered	6.07%
Triple	Three seedlings clustered	1.45%

The train, validation, and test sets are separated by year and acquisition source. The training set uses 2019–2021 data and includes intentional planting of double and triple clusters in 2021. The validation set also uses 2019–2021 imagery but from videos excluded from training. The test set is composed of entirely new plants and fields from 2022. This split design is significant because it aims to reduce data leakage from overlapping videos and to test generalization across field instances rather than merely across frames (Kharismawati et al., 18 Sep 2025).

MSDD spans growth stages V2–V10 and includes variability in soil color and type, lighting, plant density, camera angle, plant orientation, and occlusion. The paper highlights uniform rows, gapped rows, and rows with single, double, and triple plantings. Wind-induced leaf deformation, overlapping leaves, and perspective distortion at image edges are reported as recurring visual factors. Benchmarking shows that detection is most reliable during V4–V6 stages and under nadir views. This suggests that MSDD is not only a detection corpus but also a controlled study of nuisance factors relevant to field deployment (Kharismawati et al., 18 Sep 2025).

4. Benchmark protocol and reported detector performance

The benchmark evaluates YOLOv9, YOLOv11, YOLOv12, and Faster-RCNN. Faster-RCNN is instantiated with Detectron2, ResNet-50, and Feature Pyramid Network. The reported metrics include IoU, mAP at $\text{IoU} \geq 0.5$, [email protected]:0.95, precision, recall, F1, confusion matrices, class-wise analysis, growth-stage and camera-angle analysis, and inference speed (Kharismawati et al., 18 Sep 2025).

The paper gives the standard detection metrics

$\text{IoU} = \frac{|B_p \cap B_g|}{|B_p \cup B_g|}$

and

$\text{mAP} = \frac{1}{N} \sum_{i=1}^{N} AP_i .$

It also reports precision and recall in the standard forms

$$\text{Precision} = \frac{TP}{TP + FP} \qquad \text{and} \qquad \text{Recall} = \frac{TP}{TP + FN}.$$

For the single-plant class, performance is substantially stronger than for double or triple groupings. The benchmark reports single-class [email protected] up to 0.916, precision up to 0.953 for YOLOv12, recall up to 0.873 for YOLOv9, and F1 values of 0.905 for YOLOv9 and 0.854 for YOLOv12. Faster-RCNN attains the highest reported precision, 0.984, but with lower recall, 0.678. The abstract summarizes the single-class regime as reaching precision up to 0.984 and recall up to 0.873 across the tested models (Kharismawati et al., 18 Sep 2025).

Performance on multi-plant groupings is markedly lower. The paper attributes this to class rarity and ambiguous visual structure. As an example, YOLOv9 reaches double-class [email protected] of 0.233, while triple detection is reported as near zero, specifically 0.000. YOLOv11 and YOLOv12 improve triple performance only marginally. The detector comparison identifies YOLO11 as the fastest model, while YOLOv9 gives the highest accuracy for single plants. The abstract further reports that YOLO11 maintains inference at 35 ms per image, with an additional 120 ms for saving outputs. In aggregate, the benchmark indicates that current detectors are already effective for dominant singleton cases but remain limited on grouped seedlings (Kharismawati et al., 18 Sep 2025).

5. Technical limitations, ambiguities, and interpretation

The most prominent limitation reported for MSDD is class imbalance. Because 92.47% of annotated objects are singles, doubles and triples are under-represented. The paper directly links this imbalance to degraded multi-class accuracy. In addition, the appearance of clustered seedlings is described as irregular and often associated with planting errors, which makes the minority classes harder not only because they are rare but also because their visual morphology is less canonical (Kharismawati et al., 18 Sep 2025).

The paper also identifies annotation ambiguity as a challenge. Touching or closely spaced seedlings can be interpreted as either two singles or one double box, and this ambiguity affects both labeling and model evaluation. Occlusion is another major source of difficulty: mature plants can touch or overlap, while very young plants can be visually weak. Perspective distortion and scale variation caused by camera angle or lens geometry are reported to reduce robustness, especially away from nadir views. Transfer across views is only partial; models trained on nadir imagery generalize somewhat to oblique imagery but with measurable dropoff (Kharismawati et al., 18 Sep 2025).

Synthetic augmentation for doubles and triples was attempted, but the paper states that its usefulness was limited by the realism of the composites and suggests future use of GANs. A common misconception would be to interpret the presence of grouped classes as evidence that the stand-counting problem is solved for anomalous emergence patterns. The reported benchmarks do not support that conclusion. Rather, they show that MSDD exposes a persistent failure regime in rare and ambiguous multi-plant detections, which is precisely one of its research functions.

6. Relation to adjacent maize seedling resources

MSDD occupies a distinct position among maize seedling datasets because it is an aerial RGB object-detection benchmark with bounding boxes for single, double, and triple plant groupings (Kharismawati et al., 18 Sep 2025). This differs fundamentally from a controlled species-classification dataset such as the public plant seedling image database of Giselsson et al., which contains approximately 960 unique plants from 12 species, including maize (\emph{Zea mays L.}), across 407 images acquired in greenhouse conditions at roughly 10 pixels per mm (Giselsson et al., 2017). That earlier dataset uses species-level image labels, proposes a weighted average $f_1$ benchmark for classification, and does not provide bounding boxes or segmentation masks. It is therefore relevant to early-stage seedling classification under controlled imaging, but not directly to in-field maize seedling detection (Giselsson et al., 2017).

MSDD also differs from the 3D maize shoot point-cloud dataset of the stem-leaf segmentation study by Miao et al., which consists of 30 maize seedlings at V3–V8 stages scanned with a hand-held laser scanner and annotated by manual organ-level segmentation in MeshLab (Zhu et al., 2020). That resource is designed for automatic stem-leaf segmentation and phenotypic trait extraction from 3D point clouds, not for 2D aerial detection. Its reported outputs include stem and leaf instance labels, segmentation metrics such as mean precision 0.964 and mean recall 0.966, and six phenotypic traits with $R^2$ above 0.94 (Zhu et al., 2020). In methodological terms, it addresses organ segmentation and morphology reconstruction, whereas MSDD addresses field detection and stand counting.

These contrasts clarify the scope of MSDD. It is not a generic maize image collection, not a species benchmark, and not a 3D phenotyping corpus. Its specific contribution is to provide public, high-resolution, field-acquired, detection-oriented annotations under realistic variation, together with benchmarks on YOLO-family models and Faster-RCNN. The dataset and code are reported as publicly available through a Google Drive distribution and a GitHub repository, which supports both YOLO-format and RCNN-format use cases (Kharismawati et al., 18 Sep 2025).