FranceCrops Float16 Dataset

Memory-efficient version of the FranceCrops dataset stored in float16 format.

Dataset Description

Each sample contains Sentinel-2 time series data (shape: 100, 60, 12) and corresponding crop type labels
Features stored in float16 format for reduced memory usage
Labels stored as int16

Usage

from datasets import load_dataset

# Load dataset
dataset = load_dataset("saget-antoine/francecrops-float16")

# Load with specific format (recommended for training)
# As float32
dataset = dataset.with_format('torch', dtype=torch.float32)
# Or as float16 for memory efficiency
dataset = dataset.with_format('torch', dtype=torch.float16)
# Or as bfloat16 for TPU compatibility
dataset = dataset.with_format('torch', dtype=torch.bfloat16)

Memory Usage

Using float16 format reduces the memory footprint by approximately 50% compared to float32, while maintaining sufficient precision for most agricultural monitoring tasks.

Citation

If you use this dataset, please cite the original FranceCrops paper.