SpecTf: Transformers enable data-driven imaging spectroscopy cloud detection.

Area of Science:

• Earth Science
• Remote Sensing
• Artificial Intelligence

Background:

• Visible-shortwave infrared (VSWIR) imaging spectrometers offer vast potential for global Earth system process quantification.
• Reliable cloud screening is a critical challenge, as traditional spatial and temporal methods are limited by cloud variability and temporal coverage.

Purpose of the Study:

• To develop an advanced cloud detection method for VSWIR imaging spectroscopy data.
• To address the limitations of existing cloud screening techniques using a novel deep-learning approach.

Main Methods:

• Introduced the Spectroscopic Transformer (SpecTf), a deep-learning architecture tailored for spectroscopy.
• SpecTf performs cloud detection using solely spectral information, treating spectral measurements as sequences.
• The model was evaluated against the baseline approach for the Earth surface Mineral dust source InvesTigation (EMIT) instrument.

Main Results:

• SpecTf significantly outperforms the current EMIT baseline cloud screening method.
• SpecTf achieves comparable performance to other machine learning methods with significantly fewer parameters.
• The model's attention mechanism provides interpretability, highlighting physically meaningful spectral features.

Conclusions:

• SpecTf offers a robust and efficient solution for cloud screening in VSWIR imaging spectroscopy.
• The model demonstrates potential for cross-instrument generalization, enabling instrument-agnostic algorithms.
• This work advances automated data processing for global environmental monitoring using remote sensing.