ioai2025-athome-satellite-images.py

import datasets
import numpy as np
import pandas as pd

_DESCRIPTION = """
GOES-16 ABI (Advanced Baseline Imager) satellite image dataset with multi-spectral imagery and corresponding labels.
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.
Data provided by NOAA and NESDIS.
"""

_HOMEPAGE = ""
_LICENSE = ""

# URLs for the data files - using HuggingFace repository URLs
_URLS = {
    "data": "https://huggingface.co/datasets/Silicon23/ioai2025-athome-satellite-images/resolve/main/data/dataset.npz",
    "metadata": "https://huggingface.co/datasets/Silicon23/ioai2025-athome-satellite-images/resolve/main/data/metadata.csv"
}

class Goes16Dataset(datasets.GeneratorBasedBuilder):
    """GOES-16 ABI satellite image dataset with multi-spectral imagery."""

    VERSION = datasets.Version("1.0.0")

    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name="all",
            version=VERSION,
            description="All resolutions combined (128x128 and 256x256 images)",
        ),
        datasets.BuilderConfig(
            name="128x128",
            version=VERSION,
            description="128x128 resolution images only",
        ),
        datasets.BuilderConfig(
            name="256x256", 
            version=VERSION,
            description="256x256 resolution images only",
        ),
    ]

    DEFAULT_CONFIG_NAME = "all"

    def _info(self):
        if self.config.name == "all":
            # For "all" config, use flexible features that can handle both resolutions
            features = datasets.Features({
                "image": datasets.Sequence(datasets.Sequence(datasets.Sequence(datasets.Value("float32")))),  # Variable size [16, H, W]
                "label": datasets.Sequence(datasets.Sequence(datasets.Value("uint8"))),  # Variable size [H, W]
                "i": datasets.Value("int32"),
                "j": datasets.Value("int32"),
                "start_time": datasets.Value("string"),
                "end_time": datasets.Value("string"),
                "ind": datasets.Value("int32"),
                "size": datasets.Value("int32"),  # This field indicates the actual resolution (128 or 256)
            })
        elif self.config.name == "128x128":
            # For 128x128, use fixed-size arrays for better caching compatibility
            features = datasets.Features({
                "image": datasets.Array3D(shape=(16, 128, 128), dtype="float32"),
                "label": datasets.Array2D(shape=(128, 128), dtype="uint8"),
                "i": datasets.Value("int32"),
                "j": datasets.Value("int32"),
                "start_time": datasets.Value("string"),
                "end_time": datasets.Value("string"),
                "ind": datasets.Value("int32"),
                "size": datasets.Value("int32"),
            })
        else:  # 256x256
            # For 256x256, use fixed-size arrays for better caching compatibility
            features = datasets.Features({
                "image": datasets.Array3D(shape=(16, 256, 256), dtype="float32"),
                "label": datasets.Array2D(shape=(256, 256), dtype="uint8"),
                "i": datasets.Value("int32"),
                "j": datasets.Value("int32"),
                "start_time": datasets.Value("string"),
                "end_time": datasets.Value("string"),
                "ind": datasets.Value("int32"),
                "size": datasets.Value("int32"),
            })

        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            homepage=_HOMEPAGE,
            license=_LICENSE,
        )

    def _split_generators(self, dl_manager):
        # Download the files (no extraction needed for .npz/.csv)
        downloaded_files = dl_manager.download(_URLS)
        
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "data_file": downloaded_files["data"],
                    "metadata_file": downloaded_files["metadata"],
                    "split": "train"
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "data_file": downloaded_files["data"],
                    "metadata_file": downloaded_files["metadata"],
                    "split": "test"
                },
            ),
        ]

    def _generate_examples(self, data_file, metadata_file, split):
        # Load data and metadata from the provided file paths
        data = np.load(data_file)
        metadata = pd.read_csv(metadata_file)
        
        # Filter metadata for the current split
        split_metadata = metadata[metadata['split'] == split]
        
        example_id = 0
        
        if self.config.name == "all":
            # For "all" config, load both 128x128 and 256x256 data
            for size in [128, 256]:
                # Filter metadata for current resolution
                size_metadata = split_metadata[split_metadata['size'] == size]
                
                # Get corresponding arrays
                X_key = f"X_{split}_{size}"
                Y_key = f"Y_{split}_{size}"
                
                X_data = data[X_key]  
                Y_data = data[Y_key]
                
                # Generate examples using metadata
                for _, row in size_metadata.iterrows():
                    ind = row['ind']
                    
                    if ind < len(X_data):  # Safety check
                        # Convert numpy arrays to lists for "all" config compatibility
                        image_array = X_data[ind].astype(np.float32)
                        label_array = Y_data[ind].astype(np.uint8)
                        
                        yield example_id, {
                            "image": image_array.tolist(),  # Convert to nested list
                            "label": label_array.tolist(),  # Convert to nested list
                            "i": int(row['i']),
                            "j": int(row['j']),
                            "start_time": str(row['start_time']),
                            "end_time": str(row['end_time']),
                            "ind": int(row['ind']),
                            "size": int(row['size']),
                        }
                        example_id += 1
        
        else:
            # For specific resolution configs (128x128 or 256x256)
            if self.config.name == "128x128":
                size = 128
            else:  # 256x256
                size = 256
            
            # Filter metadata for current resolution
            size_metadata = split_metadata[split_metadata['size'] == size]
            
            # Get corresponding arrays
            X_key = f"X_{split}_{size}"
            Y_key = f"Y_{split}_{size}"
            
            X_data = data[X_key]  
            Y_data = data[Y_key]
            
            # Generate examples using metadata
            for _, row in size_metadata.iterrows():
                ind = row['ind']
                
                if ind < len(X_data):  # Safety check
                    # Return numpy arrays directly for fixed-size configs
                    image_array = X_data[ind].astype(np.float32)
                    label_array = Y_data[ind].astype(np.uint8)
                    
                    yield example_id, {
                        "image": image_array,  # Return numpy array directly
                        "label": label_array,  # Return numpy array directly
                        "i": int(row['i']),
                        "j": int(row['j']),
                        "start_time": str(row['start_time']),
                        "end_time": str(row['end_time']),
                        "ind": int(row['ind']),
                        "size": int(row['size']),
                    }
                    example_id += 1