toolambr/s2_lc_leuven_tiled

RSVSW Remote Sensing Semantic Segmentation Dataset

Dataset Summary

This dataset is specifically curated for an educational practical session focused on deep learning for Earth Observation. It contains remote sensing imagery sourced via Google Earth Engine (GEE) and corresponding pixel-wise classification masks. The goal of this dataset is that it can be used for semantic segmentation tasks. Within the practical, we will use U-Net.

Data Instances

Each instance in the dataset represents a single paired sample containing:

• image: A PIL Image representing the optical remote sensing patch.
• mask: A PIL Image representing the semantic mask where pixel values correspond to the class index (e.g., 0 for background, 1 for water, 2 for vegetation, etc.).

Data Splits

The dataset was pre-split into three distinct subsets:

Split	Number of Images	Description
train	[418]	Used for calculating the loss and updating U-Net weights.
validation	[114]	Used to monitor generalization during training and tune hyperparameters.
test	[57]	Held-out set for final evaluation of the model's performance.

Usage for Students

To load this dataset directly into a Python environment (like Google Colab), use the datasets library. No authentication is required.

!pip install datasets -q

from datasets import load_dataset

# Load the entire dataset dictionary (contains train, val, test)
dataset = load_dataset("your-username/rsvsw-segmentation-data")

# Accessing a specific split
train_data = dataset["train"]

# Viewing a sample
sample = train_data[0]
display(sample["image"]) # View satellite image
display(sample["mask"])  # View ground truth mask