swaib/Multimodal_Monsoon_Indian_Dataset

Multimodal Monsoon Indian Dataset (1 km)

Overview

The Multimodal Monsoon Indian Dataset is a curated, high-resolution (≈1 km) Earth-observation dataset designed for monsoon precipitation classification and analysis across Indian states. The dataset integrates satellite imagery, climate reanalysis variables, and static geospatial features into a unified, state-wise structure suitable for multimodal deep learning.

This repository contains raw exported artifacts generated via reproducible Jupyter notebooks (one per state), preserving provenance and enabling flexible downstream preprocessing. The dataset is intended for research in:

• Climate AI
• Precipitation classification and nowcasting
• Multimodal & geospatial deep learning
• High-resolution monsoon analysis over India

---

Geographic Coverage

The dataset currently includes state-level data for:

• Assam
• Bihar
• Himachal Pradesh
• Karnataka
• Kerala

Each state was processed independently using the same notebook pipeline, ensuring methodological consistency across regions.

---

Spatial & Temporal Resolution

• Spatial resolution: ~1 km × 1 km
• Spatial extent: State administrative boundaries
• Temporal coverage: Monsoon-season periods (state-specific, aligned with data availability)
• Projection: As exported from source datasets (see metadata/)

---

Modalities

Each state folder contains multiple modalities aligned spatially and temporally.

1) Satellite-Derived Features

Examples include (depending on state and availability):

• NDVI (vegetation index)
• Land Surface Temperature (LST)
• Spectral bands and derived indices

Source: Satellite products accessed and exported via Google Earth Engine (GEE).

2) Climate & Reanalysis Variables

• ERA5-derived atmospheric variables (e.g., temperature, humidity, wind, precipitation aggregates)
• Gridded climate fields resampled/aligned to the target resolution

Source: ERA5 reanalysis.

3) Static Geospatial Context

• Elevation / topography
• Land-use / land-cover (LULC)
• Other static spatial priors relevant to precipitation modeling

4) Labels

• Pixel-wise or region-wise precipitation labels (classification-oriented)
• Label definitions, thresholds, and mappings are documented in metadata/

---

Directory Structure

.
├── data/
│   ├── Assam/
│   ├── Bihar/
│   ├── Himachal_Pradesh/
│   ├── Karnataka/
│   └── Kerala/
│       ├── *.tif        # GeoTIFF raster layers
│       ├── *.csv        # Tabular exports (where applicable)
│       └── masks/       # Label masks / precipitation classes
│
├── metadata/
│   ├── variable_sources.yaml
│   ├── label_definitions.yaml
│   └── state_metadata.json
│
├── scripts/
│   └── README.md        # Placeholder for GEE / preprocessing scripts
│
├── README.md
└── LICENSE

Note: File names and formats are preserved as exported to maintain provenance and traceability.

---

Notebook-Based Data Generation

Each state dataset was generated using a dedicated Jupyter notebook (e.g., Assam.ipynb) that follows the same high-level pipeline:

1. Define the state boundary and spatial grid
2. Query satellite and climate datasets via Google Earth Engine / reanalysis APIs
3. Export aligned raster layers at ~1 km resolution
4. Generate precipitation labels
5. Save outputs as GeoTIFF / CSV artifacts

The same notebook logic was reused (with state-specific parameters) for Assam, Bihar, Himachal Pradesh, Karnataka, and Kerala, ensuring methodological consistency across all regions.

---

How to Use

This dataset is provided as raw geospatial files, not as pre-extracted ML tensors.

Example (Python): Load a raster

import rasterio

with rasterio.open("data/Assam/ndvi.tif") as src:
    ndvi = src.read(1)

Users are expected to:

• Construct spatial patches or tiles
• Perform normalization / temporal aggregation
• Align modalities according to their experimental design

This design choice keeps the dataset model-agnostic and suitable for diverse research pipelines.

---

Dataset Viewer Note

Note: The Hugging Face dataset viewer is not applicable for this dataset since it consists of geospatial raster files rather than tabular samples. The displayed “rows” correspond to file-level metadata, not ML training instances.

---

Intended Use & Limitations

Intended use:

• Research and benchmarking of multimodal precipitation models
• Regional monsoon analysis
• Spatial deep learning experiments

Limitations:

• Raw exported artifacts; users must implement task-specific preprocessing
• Temporal coverage may vary slightly by state
• Not intended for operational forecasting without further validation

---

Citation

If you use this dataset in your research, please cite:

@misc{mazumder2025learningregionalmonsoonpatterns,
      title={Learning Regional Monsoon Patterns with a Multimodal Attention U-Net}, 
      author={Swaib Ilias Mazumder and Manish Kumar and Aparajita Khan},
      year={2025},
      eprint={2509.23267},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.23267}, 
}

---

License

This dataset is released under the MIT License. See LICENSE for details.