GOES-16 ABI Satellite Image Dataset

This dataset contains GOES-16 ABI (Advanced Baseline Imager) satellite images with multi-spectral imagery and corresponding labels for semantic segmentation tasks.

Dataset Description

The dataset contains training and test splits at two different resolutions (128x128 and 256x256). Each image has 16 spectral channels from the GOES-16 ABI instrument. The data is provided by NOAA and NESDIS.

Dataset Structure

The dataset is organized into the following configurations:

128x128: Images at 128x128 pixel resolution
- Train: 529 examples (~568 MB)
- Test: 51 examples (~55 MB)
- Total: 580 examples (~623 MB)
256x256: Images at 256x256 pixel resolution
- Train: 1,713 examples (~5.5 GB estimated)
- Test: 183 examples (~587 MB estimated)
- Total: 1,896 examples (~6.1 GB estimated)

Data Fields

Each example in the dataset contains:

image: Multi-spectral satellite image as a 3D array with shape [16, height, width]
- 16 spectral channels from GOES-16 ABI instrument
- Values are float32 type, typically in range [-3, 3]
- Height and width are 128 or 256 depending on configuration
label: Corresponding label/mask as a 2D array with shape [height, width]
- Values are uint8 type, typically binary (0 or 1)
i: Spatial coordinate i (int32)
j: Spatial coordinate j (int32)
start_time: Start time of the satellite observation (string)
end_time: End time of the satellite observation (string)
ind: Index within the original data array (int32)
size: Resolution size (128 or 256) (int32)

Data Source

The satellite data originates from:

Instrument: GOES-16 Advanced Baseline Imager (ABI)
Provider: NOAA (National Oceanic and Atmospheric Administration)
Data Center: NESDIS (National Environmental Satellite, Data, and Information Service)

Usage

Basic Usage

from datasets import load_dataset
import numpy as np

# Load 128x128 resolution data
dataset = load_dataset("Silicon23/ioai2025-athome-satellite-images", name="128x128")

# Access a sample
sample = dataset["train"][0]

# Convert to numpy arrays for processing
image = np.array(sample["image"])  # Shape: (16, 128, 128)
label = np.array(sample["label"])  # Shape: (128, 128)

print(f"Image shape: {image.shape}")
print(f"Label shape: {label.shape}")
print(f"Image data type: {image.dtype}")
print(f"Label data type: {label.dtype}")
print(f"Image value range: [{image.min():.3f}, {image.max():.3f}]")
print(f"Label value range: [{label.min()}, {label.max()}]")

Accessing Metadata

# Get observation metadata
print(f"Spatial coordinates: i={sample['i']}, j={sample['j']}")
print(f"Observation time: {sample['start_time']} to {sample['end_time']}")
print(f"Resolution: {sample['size']}x{sample['size']}")
print(f"Array index: {sample['ind']}")

Working with Different Resolutions

# Load different resolutions
dataset_128 = load_dataset("your-username/goes16-satellite", name="128x128")
dataset_256 = load_dataset("your-username/goes16-satellite", name="256x256")

# Compare samples
sample_128 = dataset_128["train"][0]
sample_256 = dataset_256["train"][0]

image_128 = np.array(sample_128["image"])  # Shape: (16, 128, 128)
image_256 = np.array(sample_256["image"])  # Shape: (16, 256, 256)

print(f"128x128 image shape: {image_128.shape}")
print(f"256x256 image shape: {image_256.shape}")

Data Processing and Visualization

import matplotlib.pyplot as plt

# Load a sample
sample = dataset["train"][0]
image = np.array(sample["image"])  # Shape: (16, 128, 128)
label = np.array(sample["label"])  # Shape: (128, 128)

# Visualize a specific channel (e.g., channel 0)
plt.figure(figsize=(12, 4))

plt.subplot(1, 3, 1)
plt.imshow(image[0], cmap='viridis')
plt.title('Channel 0 (Raw)')
plt.colorbar()

plt.subplot(1, 3, 2)
plt.imshow(label, cmap='gray')
plt.title('Label/Mask')
plt.colorbar()

plt.subplot(1, 3, 3)
# Create RGB composite (example - adjust channels based on your specific needs)
rgb_composite = np.stack([image[2], image[1], image[0]], axis=-1)
rgb_composite = (rgb_composite - rgb_composite.min()) / (rgb_composite.max() - rgb_composite.min())
plt.imshow(rgb_composite)
plt.title('RGB Composite')

plt.tight_layout()
plt.show()

Training Example

from datasets import load_dataset
from torch.utils.data import DataLoader
import torch
import numpy as np

# Load dataset
dataset = load_dataset("your-username/goes16-satellite", name="128x128")

# Convert to PyTorch tensors
def collate_fn(batch):
    images = torch.stack([torch.from_numpy(np.array(item["image"])) for item in batch])
    labels = torch.stack([torch.from_numpy(np.array(item["label"])) for item in batch])
    return {"image": images, "label": labels}

# Create data loaders
train_loader = DataLoader(
    dataset["train"], 
    batch_size=8, 
    shuffle=True, 
    collate_fn=collate_fn
)
test_loader = DataLoader(
    dataset["test"], 
    batch_size=8, 
    shuffle=False, 
    collate_fn=collate_fn
)

# Example training loop structure
for batch in train_loader:
    images = batch["image"]  # Shape: (batch_size, 16, 128, 128)
    labels = batch["label"]  # Shape: (batch_size, 128, 128)
    
    # Your training code here
    # model_output = model(images)
    # loss = criterion(model_output, labels)
    break

Dataset Statistics

128x128 Configuration

Total examples: 580 (529 train, 51 test)
Dataset size: 623 MB
Image dimensions: 16 channels × 128 × 128 pixels
Data types: float32 (images), uint8 (labels)

256x256 Configuration

Total examples: 1,896 (1,713 train, 183 test)
Dataset size: ~6.1 GB (estimated)
Image dimensions: 16 channels × 256 × 256 pixels
Data types: float32 (images), uint8 (labels)

Applications

This dataset can be used for:

Satellite image semantic segmentation
Weather pattern recognition and classification
Multi-spectral image processing
Earth observation studies
Remote sensing applications
Computer vision research on satellite imagery
Time series analysis of atmospheric conditions
Cloud detection and classification
Environmental monitoring

Data Format

The dataset is automatically downloaded and processed when loaded through the HuggingFace datasets library. The underlying data is stored in NPZ format with corresponding metadata.