trust-tad/dual-validation-multispectral

Dataset Description

This is a Machine Learning-ready, multitemporal satellite dataset designed specifically for cloud gap imputation and stratified model evaluation. It acts as the foundational dataset for the framework introduced in "A Dual Validation Framework for Curating Machine Learning-Ready Satellite Datasets: A Scalable Pipeline and Stratified Analysis." The dataset is strictly curated from standard Analysis-Ready Data (ARD) to include intrinsic radiometric validation and is stratified using a novel composite Difficulty Index (DI). This index evaluates samples across three dimensions: spatial heterogeneity, temporal phenological variability, and cloud persistence.

Dataset Details

• Source Sensor: MODIS/Terra Surface Reflectance Daily L2G Global 500m (MOD09GA v6.1)
• Ancillary Data: ESA WorldCover 2021 (for spatial heterogeneity mapping)
• Study Area: Central Europe (Longitude 0°–20°E, Latitude 40°–60°N)
• Time Window: June 14 to July 3, 2021 (Julian days 165–184)
• Task: Masked Image Modeling / Cloud Gap Imputation
• Size: A Zarr data cube with dimensions of 20×8×4800×4800 (representing 20 days, 8 bands, and a 4800×4800 spatial grid). The first seven bands are surface reflectance, while the 8th band is a custom multi-label usability mask that encodes various atmospheric and quality conditions. The on-the-fly dataloader dynamically processes this mask to crop the cube into 224x224 spatiotemporal patches, utilizing 6 active reflectance bands for model input.
• Associated Models: trust-tad/dual-validation-multispectral

Data Splits & Sampling Strategy

To prevent temporal data leakage during model evaluation, the dataset employs strict sampling rules:

• Training Set: Generated using a sliding temporal window with a 1-day overlap to maximize data availability.
• Validation & Test Sets: Generated using strict non-overlapping temporal windows (a strict stride) and non-overlapping spatial patches.
• Target Frame: The clearest frame within a temporal window (usability $\tau \ge 0.85$) is selected and artificially masked to serve as the physical reconstruction target.

Dataset Structure

Samples are categorized into three difficulty strata based on the Difficulty Index (DI):

• Low Difficulty: DI < 0.50 (Minimal cloud cover, low variance)
• Medium Difficulty: 0.50 ≤ DI ≤ 0.64
• High Difficulty: DI > 0.64 (High cloud persistence, high spatial/temporal complexity)

Citation

(TBA)