PlantSeg: A Large-Scale In-the-wild Dataset for Plant Disease Segmentation

Tianqi Wei, Zhi Chen, Xin Yu, Scott Chapman, Paul Melloy, Zi HuangarXiv 2024 · OpenReview

Abstract

Plant diseases pose significant threats to agriculture, necessitating proper diagnosis and effective treatment to safeguard crop yields. To automate the diagnosis process, image segmentation is usually adopted for precisely identifying diseased regions, thereby advancing precision agriculture. Developing robust image segmentation models for plant diseases demands high-quality annotations across numerous images. However, existing plant disease datasets typically lack segmentation labels and are often confined to controlled laboratory settings, which do not adequately reflect the complexity of natural environments. Motivated by this fact, we established PlantSeg, a large-scale segmentation dataset for plant diseases. PlantSeg distinguishes itself from existing datasets in three key aspects: annotation type, image source, and scale. It features 11,400 images with detailed segmentation masks and 8,000 healthy plant images. Extensive technical experiments validate the high quality of PlantSeg's annotations, providing a critical foundation for developing and benchmarking advanced plant disease segmentation algorithms.

Introduction: The "Lab-to-Field" Gap

While AI has shown promise in identifying plant diseases, most existing work suffers from significant bottlenecks. Most datasets provide only coarse annotations like class labels or bounding boxes. For precision agriculture—such as targeted fungicide spraying—farmers require pixel-level semantic segmentation to determine the exact area and severity of infection.

Furthermore, popular datasets like PlantVillage consist of leaves photographed against uniform backgrounds in controlled settings. Models trained on these "staged" images often fail when faced with the complex lighting, occlusions, and background noise of an actual farm. PlantSeg bridges this gap by focusing on "in-the-wild" imagery.

Comparison between controlled laboratory images (PlantVillage) and in-the-wild images (PlantSeg).

The PlantSeg Dataset

PlantSeg is designed to provide the scale and complexity needed to train robust, field-ready models. The dataset includes:

  • 11,458 images with detailed disease segmentation masks.
  • 8,000 healthy plant images serving as negative samples.
  • 115 disease categories across 34 plant species, including staple crops (wheat, corn), profit crops (coffee, tobacco), and various fruits and vegetables.

Compared to existing datasets like PlantDoc or FieldPlant, PlantSeg is the largest "in-the-wild" resource providing high-quality segmentation masks rather than just bounding boxes. This granularity allows for the calculation of "mask ratios"—the proportion of diseased pixels to total pixels—which is a critical metric for severity scoring.

Examples of high-quality segmentation masks in the PlantSeg dataset across various plant types and diseases.

Methodology

The creation of PlantSeg followed a rigorous three-stage pipeline:

  1. Image Acquisition: Images were collected via keyword searches across major web engines for 115 diseases.
  2. Data Cleaning: Annotators and experts manually reviewed the collection to discard ambiguous, incorrect, or low-quality images.
  3. Expert-Verified Annotation: Using the LabelMe tool, 10 trained annotators drew detailed polygons around diseased regions. The process was overseen by two expert plant pathologists who reviewed the masks for consistency and biological correctness.

Key Insights and Benchmarking

Benchmarking several state-of-the-art models on PlantSeg revealed that the complexity of "wild" data presents a significant challenge. The study found that diseases causing structural changes—like the curling of stems or wilting of leaves—are significantly harder for current AI models to segment than simple flat lesions.

Key performance results include:

  • SegNeXt was the top performer, achieving a Mean Intersection over Union (MIoU) of 44.52%.
  • DeepLabv3 struggled, with MIoU as low as 17.24% on smaller backbones.
  • Backbone capacity is crucial; moving from ResNet-50 to ResNet-101 consistently improved performance, suggesting that plant disease features are visually complex enough to require high-capacity models.

Conclusion

PlantSeg moves plant disease diagnosis from simple classification to precise, pixel-level segmentation in real-world environments. While challenges remain—such as handling extreme morphological deformations and expanding taxonomic breadth—PlantSeg provides a critical foundation for developing AI that can function effectively in the messy, unpredictable conditions of a real farm.

Summary updated

Abstract

Plant diseases pose significant threats to agriculture. It necessitates proper diagnosis and effective treatment to safeguard crop yields. To automate the diagnosis process, image segmentation is usually adopted for precisely identifying diseased regions, thereby advancing precision agriculture. Developing robust image segmentation models for plant diseases demands high-quality annotations across numerous images. However, existing plant disease datasets typically lack segmentation labels and are often confined to controlled laboratory settings, which do not adequately reflect the complexity of natural environments. Motivated by this fact, we established PlantSeg, a large-scale segmentation dataset for plant diseases. PlantSeg distinguishes itself from existing datasets in three key aspects. (1) Annotation type: Unlike the majority of existing datasets that only contain class labels or bounding boxes, each image in PlantSeg includes detailed and high-quality segmentation masks, associated with plant types and disease names. (2) Image source: Unlike typical datasets that contain images from laboratory settings, PlantSeg primarily comprises in-the-wild plant disease images. This choice enhances the practical applicability, as the trained models can be applied for integrated disease management. (3) Scale: PlantSeg is extensive, featuring 11,400 images with disease segmentation masks and an additional 8,000 healthy plant images categorized by plant type. Extensive technical experiments validate the high quality of PlantSeg's annotations. This dataset not only allows researchers to evaluate their image classification methods but also provides a critical foundation for developing and benchmarking advanced plant disease segmentation algorithms.
BibTeX
@misc{wei2024plantseg,
  author = {Tianqi Wei and Zhi Chen and Xin Yu and Scott Chapman and Paul Melloy and Zi Huang},
  title = {{PlantSeg: A Large-Scale In-the-wild Dataset for Plant Disease Segmentation}},
  year = {2024},
  eprint = {2409.04038},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.04038},
}
Figure