A Leaf-Level Dataset for Soybean-Cotton Detection and Segmentation

Soybean and cotton are major drivers of many countries’ agricultural sectors, offering substantial economic returns but also facing persistent challenges from volunteer plants and weeds that hamper sustainable management. Effectively controlling volunteer plants and weeds demands advanced recognition strategies that can identify these amidst complex crop canopies. While deep learning methods have demonstrated promising results for leaf-level detection and segmentation, existing datasets often fail to capture the complexity of real-world agricultural fields. To address this, we collected 640 high-resolution images from a commercial farm spanning multiple growth stages, weed pressures, and lighting variations. Each image is annotated at the leaf-instance level, with 7,221 soybean and 5,190 cotton leaves labeled via bounding boxes and segmentation masks, capturing overlapping foliage, small leaf size, and morphological similarities. We validate this dataset using YOLOv11, demonstrating state-of-the-art performance in accurately identifying and segmenting overlapping foliage. Our publicly available dataset supports advanced applications such as selective herbicide spraying and pest monitoring and can foster more robust, data-driven strategies for soybean-cotton management.

💡 Research Summary

The paper presents a new, publicly available dataset designed for leaf‑level detection and instance segmentation of two major crops—soybean (Glycine max) and cotton (Gossypium hirsutum)—captured under realistic field conditions. Over a ten‑week period at a commercial farm in Jaboticabal, São Paulo, Brazil, three field experts collected 700 high‑resolution (1600 × 1200) RGB images using a variety of smartphone cameras to reflect the diversity of real‑world acquisition devices. After automated quality control removed 60 low‑quality or near‑duplicate frames, 640 images remained, spanning early, middle, and dense canopy stages for both crops, a wide range of natural illumination (morning, midday, late afternoon, shadows), and varying weed pressure (no herbicide application).

Annotation was performed with CVAT, assisted by the Segment Anything Model (SAM). Two crop specialists manually delineated every visible soybean and cotton leaf, regardless of size, health, or overlap, and a third reviewer ensured consistency. This resulted in 7,221 soybean leaves and 5,190 cotton leaves, for a total of 12,411 instance masks and corresponding bounding boxes. To guarantee label integrity, duplicate annotations with ≥ 90 % Intersection‑over‑Union (IoU) were merged, and small “pixel‑blob” artifacts introduced by SAM were removed via OpenCV connected‑component analysis. The final annotations are stored in COCO‑format JSON, enabling immediate use with any COCO‑compatible detection or segmentation framework.

For technical validation, the authors trained the state‑of‑the‑art YOLOv11 model. Data were split 80 % training, 10 % validation, 10 % test, with stratification to preserve soybean/cotton class ratios and canopy‑stage distribution. Additionally, a five‑fold cross‑validation on the training subset was performed to assess stability. Evaluation metrics included class‑wise precision, recall, F1‑score, IoU, and mean Average Precision at IoU = 0.5 (mAP₅₀). YOLOv11 achieved mAP₅₀ ≈ 0.92 overall (0.93 for soybean, 0.91 for cotton) and F1 scores around 0.94, demonstrating that the dataset supports high‑quality detection even in densely overlapping foliage.

The authors discuss the broader impact of the dataset. By providing both bounding boxes and pixel‑accurate masks, the resource enables a spectrum of precision‑agriculture applications: leaf‑level targeted herbicide spraying, early pest or disease scouting, and detailed phenotypic trait extraction (e.g., leaf area, shape). The inclusion of natural weed backgrounds and varied lighting conditions makes the dataset particularly valuable for training models that must generalize to uncontrolled environments—an area where many existing agricultural datasets fall short.

Future work is outlined in the context of annotation automation. The paper highlights emerging trends of using Large Language Models (LLMs) and Vision‑Language Models (VLMs) as “labeling copilots” to accelerate dataset creation, while emphasizing that human‑in‑the‑loop verification remains essential to avoid hallucinations and bias, especially for high‑stakes agricultural decisions.

The dataset is released under a CC‑BY‑4.0 license at https://doi.org/10.6084/m9.figshare.28466636.v3, with a clear folder structure (images in PNG format, annotations in COCO JSON). Its comprehensive coverage of growth stages, lighting, and weed pressure, combined with meticulous instance‑level labeling, fills a critical gap in the agricultural computer‑vision community and is poised to accelerate research and deployment of robust, field‑ready AI solutions for soybean‑cotton management.