AgroMind

AgroMind is an agricultural remote sensing benchmark for evaluating large multimodal models on agricultural scene understanding. It contains 28,482 question-answer pairs paired with 20,850 images drawn from nine public datasets plus proprietary global parcel data, spanning 13 task types across 4 evaluation dimensions:

• Spatial Perception — localization, relationship determination, boundary detection
• Object Understanding — classification, pest/disease diagnostics, growth status assessment
• Scene Understanding — comparison, counting, area statistics
• Scene Reasoning — visual prompt reasoning, anomaly detection, climate classification, yield prediction

This dataset has been standardized to the HF image_text_to_text format and converted to Parquet with image bytes embedded directly in the dataset. All subsets (Agriculture, CropHarvest, Fruit, Leaf_diseases, Oil_palm_trees, Pest, Rural, Trees, corn, crop) have been concatenated into a single flat dataset — use the origin_dataset column or the id field (agromind_{subset}_{index}) to filter by source subset. Images are organized by subset under images/{subset}/ within each row's raw_metadata.

This dataset is indexed on https://project-agml.github.io/ as part of the AgML python library.

---

Schema

Every record shares the same columns so heterogeneous subsets concatenate cleanly.

Column	Type	Description
id	string	Unique row identifier of the form agromind_{subset}_{index}
images	list[Image]	List of images embedded as bytes, aligned to {"type": "image"} placeholders in messages
messages	list[dict]	Conversational VLM format with role and content fields
origin_dataset	string	Source dataset name for provenance tracking
raw_metadata	string	JSON-encoded string of all original source fields not folded into messages; always includes file_names pointing to images/{subset}/filename

messages structure

[
  {
    "role": "user",
    "content": [
      {"type": "image", "text": None},   # placeholder aligned to images list
      {"type": "text", "text": "<question>"}
    ]
  },
  {
    "role": "assistant",
    "content": [{"type": "text", "text": "<answer>"}]
  }
]

raw_metadata structure

raw_metadata is a JSON string (not a native struct) — this is intentional, as it allows concatenate_datasets() to work across subsets whose original source fields differ in type. Restore fields with json.loads(row["raw_metadata"]). It always contains file_names for image path traceability:

{
  "file_names": ["images/Agriculture/filename.jpg"],
  # ...original source fields preserved verbatim
}

---

Usage

from datasets import load_dataset

# Load the full dataset
ds = load_dataset("Project-AgML/AgroMind")

# Stream without downloading
ds = load_dataset("Project-AgML/AgroMind", streaming=True)

# Filter by subset using origin_dataset
agriculture = ds["train"].filter(lambda x: x["origin_dataset"] == "Agriculture")

# Or filter using the id field (format: agromind_{subset}_{index})
crop_harvest = ds["train"].filter(lambda x: x["id"].split("_")[1] == "CropHarvest")

Accessing images

Images are stored as embedded bytes and decoded to PIL automatically:

row = ds["train"][0]
image = row["images"][0]          # PIL Image, ready to use
print(image.size)                 # (width, height)
print(image.mode)                 # RGB

Restoring raw metadata

import json

meta = json.loads(row["raw_metadata"])
print(meta["file_names"])         # ['images/Agriculture/filename.jpg']

---

Citation

@misc{li2025largemultimodalmodelsunderstand,
      title={Can Large Multimodal Models Understand Agricultural Scenes? Benchmarking with AgroMind},
      author={Li, Qingmei and Zhang, Yang and Mai, Zurong and Chen, Yuhang and Lou, Shuohong and Huang, Henglian and Zhang, Jiarui and Zhang, Zhiwei and Wen, Yibin and Li, Weijia and Fu, Haohuan and Huang, Jianxi and Zheng, Juepeng},
      year={2025},
      eprint={2505.12207},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.12207}
}