amonshano/anonymousnips26

Agro

Dataset Description

Agro is a benchmark for evaluating on-the-fly category discovery in agricultural open-world environments. It contains over 100K labeled images spanning fungi, crop pests and diseases, and insects.

The benchmark is designed around three underrepresented evaluation conditions in existing category discovery benchmarks:

• Cluttered field imagery, where agricultural objects appear under natural backgrounds, occlusion, illumination changes, and non-canonical viewpoints.
• Long-tailed agricultural distributions, where common pests or diseases dominate while rare categories are underrepresented.
• Hierarchical biological taxonomy, where categories are organized across taxonomic levels rather than treated only as flat labels.

Agro consists of three complementary subsets:

• Agro-Fungi: fine-grained discovery under cluttered field imagery.
• Agro-Crop: long-tailed pest and disease discovery across crop domains.
• Agro-Insect: taxonomy-aware discovery using biological hierarchy labels.

Dataset Details

• Dataset name: Agro
• Task: On-the-fly category discovery in agricultural open-world environments
• Modality: Image
• Language: English
• License: Apache-2.0
• Size: Over 100K labeled images
• Domains: Fungi, crop pests and diseases, insects

Intended Use

Agro is intended for research on:

• Evaluating on-the-fly category discovery algorithms.
• Benchmarking old-category recognition and novel-category discovery under streaming protocols.
• Auditing category over-discovery and category calibration.
• Evaluating taxonomy-aware discovery under biological hierarchies.
• Studying robustness under cluttered agricultural imagery and long-tailed category distributions.

Out-of-Scope Use

Agro is not intended to serve as a complete deployment benchmark for agricultural decision-making. It does not fully model long-term temporal dynamics, geographic distribution shifts, sensor variation, expert-in-the-loop diagnosis, treatment recommendation, or downstream agronomic intervention.

Models evaluated on Agro should not be directly used for high-stakes agricultural diagnosis, pest management, treatment recommendation, ecological monitoring, or treatment planning without expert validation and additional field testing.

Dataset Structure

The dataset contains image records and labels for agricultural visual recognition. Each example includes an image and its corresponding label. Depending on the subset, labels may correspond to species, pest or disease categories, or biological taxonomy nodes.

The benchmark is organized into three subsets:

Subset	Main focus	Evaluation role
Agro-Fungi	Fungi images	Cluttered fine-grained discovery
Agro-Crop	Crop pests and diseases	Long-tailed agricultural discovery
Agro-Insect	Insect taxonomy	Taxonomy-aware discovery

Data Collection and Processing

Agro is constructed from a combination of field-collected imagery and curated public sources. Field imagery provides non-iconic agricultural scenes with natural clutter, illumination variation, occlusion, and growth-stage appearance changes. Public sources are used to increase coverage of rare taxa, pests, and disease categories.

Labels are normalized to reduce synonym, spelling, capitalization, and source-format inconsistencies. Low-quality, corrupted, duplicated, and visually uninformative images are filtered before constructing the benchmark splits.

Annotations

Agro uses category-level and taxonomy-level annotations depending on the subset:

• Agro-Fungi uses species or category labels.
• Agro-Crop uses pest or disease category labels.
• Agro-Insect uses hierarchical biological taxonomy labels.

The benchmark provides standardized splits and evaluation protocols for on-the-fly category discovery.

Personal and Sensitive Information

The dataset is composed of agricultural visual imagery focused on fungi, crop pests and diseases, and insects. It is not intended to contain personal or sensitive information.

Images were filtered to focus on agricultural and biological subjects. However, users should perform additional checks before downstream redistribution, deployment, or integration into applications.

Biases, Risks, and Limitations

Agro may reflect biases from field-collected imagery and curated public sources, including geographic coverage bias, species and crop prevalence bias, taxonomic imbalance, visual acquisition bias, and long-tailed category distributions.

Some rare pests, diseases, fungi, or insect taxa may be underrepresented. The visual conditions may not cover all agricultural regions, sensors, seasons, crop varieties, growth stages, or field environments.

Agro should therefore be used as an evaluation benchmark for recognition and discovery behavior, not as evidence of deployment readiness in real-world agricultural systems.

Data Use Cases

Suitable use cases include:

• Evaluating open-world and on-the-fly category discovery methods.
• Comparing category discovery algorithms under realistic agricultural conditions.
• Measuring old-category recognition and novel-category discovery performance.
• Studying category over-discovery and cluster calibration.
• Evaluating taxonomy-aware consistency in biological visual recognition.

Unsuitable use cases include:

• Automated agricultural treatment recommendation.
• High-stakes pest or disease diagnosis without expert review.
• Direct ecological decision-making.
• Real-world deployment without additional field validation.

Social Impact

Agro may support research on more reliable agricultural visual recognition systems, especially for discovering rare pests, diseases, fungi, and insect taxa under realistic open-world conditions. Potential benefits include improved benchmarking for agricultural AI and better understanding of model limitations.

However, benchmark performance should not be interpreted as proof of safe deployment. Agricultural, ecological, or treatment decisions should involve domain experts and additional field validation.

Synthetic Data

Agro does not intentionally include synthetic data.

Citation

Citation information will be added upon paper release.

Dataset Card Contact

Contact information will be provided upon release.