README.md

---
license: apache-2.0
language:
- en
tags:
- space-science
- cassini
- cosmic-dust
- time-of-flight-spectra
- mass-spectrometry
- saturn
- enceladus
- planetary-science
pretty_name: Cassini CDA Time-of-Flight Spectra
size_categories:
- 100K<n<1M
---

# Cassini Cosmic Dust Analyzer (CDA) Time-of-Flight Spectra Dataset

## Dataset Description

This dataset contains time-of-flight (TOF) mass spectra from the **Cosmic Dust Analyzer (CDA)** instrument aboard NASA's Cassini spacecraft. The CDA was designed to analyze the composition of cosmic dust particles in the Saturnian system and interplanetary space.

### Data Cutoff

**Important:** Only data until **2008-277** (October 3, 2008, day of year 277) are included in this dataset. Data recorded after this date are excluded due to a **sensitivity drop** of the CDA instrument. This sensitivity change occurred after a close flyby of Enceladus, during which Cassini flew through an ice plume—a cloud of ice particles ejected by Enceladus' geysers. This event likely introduced bias in subsequent measurements.

## Dataset Structure

```
data/
├── lvl1/
│   └── cda_qm_spectra_pre2008277_lvl1.parquet    # Raw spectra (774,953 samples)
├── lvl2/
│   ├── cda_qm_spectra_pre2008277_train_lvl2.parquet  # Labeled spectra (19,820 samples)
│   └── cda_qm_spectra_pre2008277_inf_lvl2.parquet    # Unlabeled spectra for inference (755,136 samples)
└── meta/
    └── cda_class_list.csv                         # Human-determined labels (34,850 entries)
```

## Data Levels

### Level 1 (lvl1)
Raw time-of-flight spectra with the following columns:

| Column | Type | Description |
|--------|------|-------------|
| `sclk` | float | Spacecraft Clock - unique identifier counting in seconds. Due to CDA's 1-second dead-time, SCLK serves as a unique event identifier |
| `f_desc` | int | Internal file descriptor (not relevant for analysis) |
| `event_id` | int | Internal event identifier (not relevant for analysis) |
| `spectrum` | array[float] | Time-of-flight spectrum array (1018 values). Each value corresponds to a time step covering ~6 microseconds total. Values are in **Volts** |
| `class` | int | Set to -1000 (unlabeled in lvl1) |
| `qi_ampl` | float | QI channel amplitude in **Coulombs**. The QI (Charge Induction) channel is mounted in front of the TOF spectrometer |
| `spectrum_length` | int | Length of spectrum array (1018 for all samples; shorter spectra are excluded) |

### Level 2 (lvl2)
Processed data split into training and inference sets:

**Training set** (`train_lvl2`): 19,820 samples with human-determined labels
- Same columns as lvl1, but with valid `class` labels
- Excludes `spectrum_length` column

**Inference set** (`inf_lvl2`): 755,136 samples without labels
- Same columns as training, but without `class` column
- Suitable for model inference/prediction tasks

### Metadata
The `cda_class_list.csv` file contains human-determined labels with:

| Column | Description |
|--------|-------------|
| `sclk` | Spacecraft Clock (identifier) |
| `utc` | UTC timestamp |
| `utc-doy` | UTC with day-of-year format |
| `class` | Human-determined spectrum classification |

**Note:** The metadata file includes labels for all classified spectra, including those after 2008-277 and compressed format spectra (e.g., 509-length), which are not present in the parquet files.

## Class Labels

| Class | Count (Train) | Description |
|-------|---------------|-------------|
| `Noise` | 12,077 | Noise/artifacts |
| `1` | 4,491 | Type 1 spectrum |
| `2` | 1,015 | Type 2 spectrum |
| `3` | 909 | Type 3 spectrum |
| `?` | 553 | Uncertain/unclassified |
| `4` | 263 | Type 4 spectrum |
| `5` | 186 | Type 5 spectrum |
| `3-Car` | 89 | Type 3 Carbon subclass |
| `3-Cl` | 47 | Type 3 Chlorine subclass |
| `3-OH` | 47 | Type 3 Hydroxyl subclass |
| `3-KNa` | 41 | Type 3 Potassium-Sodium subclass |
| `5-Na` | 41 | Type 5 Sodium subclass |
| `3-P` | 34 | Type 3 Phosphorus subclass |
| `3-K` | 22 | Type 3 Potassium subclass |
| `2-X` | 3 | Type 2 variant |
| `X` | 2 | Unknown/special |

### Important Note on Subclasses

The subclasses (e.g., `3-OH`, `3-Car`, `3-Cl`) were introduced later in the classification scheme. It is likely that spectra with these specific characteristics exist within the general `3` class samples that were labeled earlier. This may introduce **label bias** when training models, as some `3` class samples may actually belong to the more specific subclasses.

## Usage

### Loading the Data

```python
import pandas as pd

# Load training data
train_df = pd.read_parquet("data/lvl2/cda_qm_spectra_pre2008277_train_lvl2.parquet")

# Load inference data
inf_df = pd.read_parquet("data/lvl2/cda_qm_spectra_pre2008277_inf_lvl2.parquet")

# Load raw lvl1 data
lvl1_df = pd.read_parquet("data/lvl1/cda_qm_spectra_pre2008277_lvl1.parquet")

# Load metadata
meta_df = pd.read_csv("data/meta/cda_class_list.csv")
```

### With Hugging Face Datasets

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("your-username/cassini-cda-spectra")
```

## Citation

If you use this dataset, please cite the Cassini CDA team and relevant publications.

## License

This dataset is released under the [Apache 2.0 License](LICENSE).

## Acknowledgments

This dataset is derived from data collected by the Cosmic Dust Analyzer (CDA) instrument aboard NASA's Cassini spacecraft during its mission to Saturn (1997-2017). We thank the CDA team and NASA/ESA for making this data available for scientific research.