Φsat-2 Multispectral L0–L1 Dataset for Domain Adaptation and Onboard Deep Learning

Overview

This dataset contains multispectral Level-0 and Level-1 acquisitions from the Φsat-2 satellite, downloaded from the INSULA platform (https://phisat2.insula.earth/perception/). The Φsat-2 satellite is provided with a MultiScape100 CIS camera. The sensor is a multispectral push-broom imager primarily designed for earth observation applications as a primary payload for microsatellites like CubeSats. It is based on a CMOS imaging sensor and a 8-band multispectral filter in the visible and near-infrared (VNIR) spectral range.

Each acquisition includes 8 spectral bands, metadata, and manually curated scene-level annotations. The dataset provides acquisition-level train/validation/test splits and can support research in raw multispectral data analysis, domain gap studies, onboard AI deep learning applications (e.g., data compression, cloud detection).

The dataset is organized into two main directories (L0 and L1). There is a one-to-one correspondence between acquisitions in L0 and L1.

For distribution purposes, L0 and L1 data are released as two separate records (phisat2_L0, phisat2_L1). This record contains L1 files. 

L0 files, manual annotations and readme are available at: https://zenodo.org/records/18746355 

Level-0 Data (L0)

Each L0 acquisition folder contains:

8 multispectral bands (TIFF format)

8 thumbnail band images (TN)

1 RGB thumbnail

L0 images represent raw sensor-domain data converted from original binary packets.

Level-1 Data (L1)

Each L1 acquisition folder includes: 

L1/Session_<ID>/bands/ -> Spectral bands. It includes:

8 radiometrically processed bands

1 BC (Band Coregistered) image

L1/Session_<ID>/geolocation/ -> Geolocation file

L1/Session_<ID>/logs -> Detailed Level-1C processing reports (including geolocation, radiometric corrections, denoising, and band coregistration steps with timing and shift information)

Spacecraft Attitude and Orbit (L1/Session_<ID>/AOCS.json)

Session Metadata (L1/Session_<ID>/session_<id>_metadata.json)

Manual annotations

Each acquisition was manually labeled. The file acquisition_info.csv provides scene-level annotations for each acquisition. Columns Description:

session_id: Unique acquisition ID

artifacts : 1 if visual or histogram anomalies are present (e.g., stripe noise, band misalignment)

corrupted_bands: 1 if corrupted bands exist

corrupted_bands_id: Identifier of affected band (if applicable, [0-7])

tag: Comma-separated macro land cover labels

lat: Scene center approximate latitude (WGS84)

lon: Scene center approximate longitude (WGS84)

split: train / val / test

For land cover tags, possible tags include: land, urban, vegetation, desert, coastal_area, sea, inland_water, soil, snow, cloud. Tags are not mutually exclusive.

For dataset splitting, the split is acquisition-level and ensures no overlap between subsets. Each split contains balanced land cover classes. Split ratio: 70% training, 20% validation, 10% testing.