AgricultureVision: Aerial Farmland Images with Expert Annotations, 2017-2019

Agriculture-Vision is a large-scale aerial agricultural image dataset created by Intelinair, Inc. It contains 94,986 high-resolution, 512x512 images sampled from 3,432 US farmlands between 2017 and 2019, each with four color channels (NIR, Red, Green, Blue). The dataset is split into 56,944 training, 18,334 validation, and 19,708 test images, with nine types of field condition patterns annotated by agronomy experts.

Use Cases

• Train semantic segmentation models to detect nutrient deficiency patterns in crops based on multi-band aerial imagery.
• Develop classifiers for identifying weed clusters and planter skips from expert-annotated field images.
• Apply weakly supervised learning methods using the added sequences of full-field imagery from the 2021 challenge.
• Analyze the impact of storm damage and water patterns on field conditions across multiple growing seasons.

Strengths

• 94,986 high-resolution (512x512) images provide substantial volume for model training.
• Nine specific, impactful field condition patterns (e.g., weed cluster, storm damage) are annotated by agronomy experts.
• Images contain four color channels (NIR, Red, Green, Blue), offering richer spectral information than standard RGB.
• A rigorous 6/2/2 train/val/test split from 3,432 unique farmlands prevents data leakage from the same field.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.
• Data may reflect geographic and temporal bias inherent to its collection from US farmlands between 2017 and 2019.

Provenance

Source

Intelinair, Inc., affiliated with a 2020 CVPR paper.

Collection Method

Aerial imagery captured from farmlands, with patterns annotated by agronomy experts.

Time Range

2017 to 2019

Freshness

Last updated date is unknown.

Geography

Numerous farming locations in the United States.