Evaluating Sentinel-2 gap filling techniques for cloud removal and data reconstruction

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Data Science

Background:

• Cloud cover frequently causes data gaps in Sentinel-2 imagery, impacting time-sensitive applications like water management and crop yield prediction.
• Existing gap-filling methods lack comprehensive performance comparisons, hindering optimal technique selection.

Purpose of the Study:

• To establish an evaluation framework for comparing spatial, temporal, spatio-temporal, and spatio-spectral gap-filling methods for Sentinel-2 imagery.
• To assess the effectiveness of these methods in restoring cloud-induced data gaps using only satellite-derived data.

Main Methods:

• Simulated cloud scenarios were used to evaluate gap-filling methods on Sentinel-2 visible, near-infrared, and shortwave infrared bands.
• Performance was assessed using coefficient of determination (R²), rRMSE, and bias.
• Methods included spatial, temporal, spatio-temporal (e.g., CLR, DL), and spatio-spectral (e.g., SSRF) approaches.

Main Results:

• Spatio-temporal methods, especially clustering and linear regression (CLR), exhibited the highest accuracy and robustness across diverse gap scenarios.
• Spatio-temporal Deep Learning (DL) performed well but required significant training and showed limited generalization.
• Spatio-spectral methods (SSRF) showed strong results in visible and NIR bands, while spatial methods like kriging struggled with larger gaps.

Conclusions:

• Spatio-temporal gap-filling techniques, particularly CLR, are most effective for restoring cloud-induced gaps in Sentinel-2 data.
• The study provides a comparative evaluation to guide the selection of appropriate gap-filling methods and identify areas for future research.
• All code is publicly available to ensure reproducibility and facilitate further advancements in satellite data gap filling.