Global Satellite Flood Detection - U-Net Ready Patches (sf11-p-256)

This repository contains a highly optimized, machine-learning-ready dataset compiled from various globally collected satellite images and diverse flood events. It has been strictly pre-processed to facilitate semantic segmentation tasks (e.g., U-Net, DeepLab) for surface water and flood detection using multi-modal satellite imagery.

🛠️ Data Processing Pipeline

To ensure stable training and prevent vanishing/exploding gradients caused by edge artifacts, the original high-resolution satellite chips were subjected to the following pipeline:

1. Patching: Images were cropped into smaller 256x256 patches to optimize GPU VRAM usage and allow for larger batch sizes.
2. Quality Assurance (NaN Filtering): Patches containing more than 30% missing data (nodata/NaN) were rigorously filtered out to prevent model collapse.
3. Data Imputation: For the accepted patches, minor remaining NaN or Inf values were safely imputed to 0.0.
4. Data Split: The dataset is split into train (70%), val (15%), and test (15%) sets to ensure proper model evaluation.

📂 Dataset Structure

The processed dataset contains three main modalities for each patch:

• images/: Sentinel-1 SAR backscatter data (VV and VH polarizations). Shape: (256, 256, 2).
• ndwi/: Normalized Difference Water Index calculated from Sentinel-2 optical bands (Green and NIR). Shape: (256, 256, 1).
• masks/: Binary ground truth masks where 1 indicates water/flood and 0 indicates dry land. Shape: (256, 256, 1).

📊 Global Normalization Stats

A pre-calculated stats.npz file is included in the root directory. It contains the global mean and std computed only on the training set to prevent data leakage. These stats should be used to normalize data during the PyTorch/TensorFlow DataLoader phase.

🖼️ Visual Samples

Below are 5 randomly selected flooded patches from the dataset demonstrating the alignment between Sentinel-1 (VV), NDWI, and the Ground Truth Mask.

🚀 How to Load

import numpy as np

# Load a specific patch
s1_patch = np.load('train/images/train_00001.npy')
ndwi_patch = np.load('train/ndwi/train_00001.npy')
mask_patch = np.load('train/masks/train_00001.npy')

# Load normalization stats
stats = np.load('stats.npz')
s1_mean, s1_std = stats['s1_mean'], stats['s1_std']

# Normalize S1
s1_normalized = (s1_patch - s1_mean) / (s1_std + 1e-8)