chnoorlee/SegField-AgriSeg-5K

AgriSeg-5K: A Real-World Benchmark for Agricultural Semantic Segmentation

Dataset Description

AgriSeg-5K is a curated, high-resolution semantic segmentation benchmark designed for precision agriculture. It contains 5,000 densely annotated images captured from a commercial weeding robot operating in real vegetable fields across California, USA.

The dataset is specifically built to evaluate model robustness under challenging field conditions, including diverse crop morphologies, growth stages, and extreme illumination variations.

Key Statistics

• Total Images: 5,000 (RGB, 1024×1024 resolution)
• Splits: 3,000 (Train) / 1,000 (Val) / 1,000 (Test)
• Semantic Classes: 4 (Crop, Weed, Soil, Mulch)
• Crop Types: 5 (Lettuce, Broccoli, Tomato, Carrot, Onion)
• Growth Stages: 3 (Early Seedling, Mid-Growth, Pre-Harvest)
• Lighting Conditions: 4 (Dawn, Overcast, Direct Sun, Dusk)

Dataset Structure

The dataset is organized by a stratified 5 Crops × 3 Stages × 4 Lights = 60 condition matrix to ensure balanced evaluation.

agriseg-5k/
├── images/          # All 5,000 original images (RGB)
├── annotations/     # COCO Panoptic JSON format (RLE-encoded masks)
├── metadata.csv     # Per-image metadata: crop, stage, lighting, location
└── splits/          # train/val/test .txt files