Sentinel-2 Cloud Masking: QA60 vs CloudScore+

Comparative analysis in Google Earth Engine over Barasat, India

Overview

This project evaluates two common Sentinel-2 cloud masking approaches—QA60 bitmasking and CloudScore+ (cloud probability–based)—to understand their impact on downstream analysis.

Using Google Earth Engine, I compare the masks over Barasat, India and quantify differences in cloud-free coverage, artifacts, and visual quality of the resulting imagery.

🧾 Project Details

Repository: S2 Cloud Masking: QA60 vs CloudScore+
Role: Remote Sensing / GEE Developer
Tools Used: Google Earth Engine (JavaScript), Sentinel-2 L1C/L2A, s2cloudless/CloudScore+, Export to Drive
Region & Period: Barasat, India — 2025
Duration: Aug 2025

Objectives

Implement QA60 and CloudScore+ masking pipelines in GEE.
Visually compare masked composites and quantify cloud-free area.
Document strengths/limitations and offer thresholding guidance.

Method Summary

QA60 (L1C): Bit 10 (opaque) & bit 11 (cirrus) used to build a binary mask.
CloudScore+ / Cloud Probability: Per-pixel cloud probability masked using a chosen threshold (e.g., 40–60%).
Compositing: Generated cloud-free median mosaics for both methods over the same date range/ROI.

Key Contributions

Produced reproducible GEE scripts for both approaches with clear comments.
Exported side-by-side maps and simple summary stats for transparent comparison.
Outlined practical tips for threshold tuning and QA checks.

Visuals

Original Sentinel-2 image of Barasat, India (unmasked).

QA60 masking result — clouds flagged using bitmask (opaque + cirrus).

CloudScore+ masking result — per-pixel cloud probability thresholding removes haze more effectively.

QA60 binary mask (white = cloud, green = clear).

CloudScore+ binary mask (white = cloud, blue = clear).

Insights

Coverage vs. Cleanliness: QA60 is fast and conservative but can leave thin clouds; CloudScore+ reduces haze with a well-chosen threshold, often yielding cleaner composites.
Tuning Matters: Probability thresholds (e.g., 40–60%) change outcomes noticeably. Lower thresholds retain more area but risk residual clouds; higher thresholds remove more haze but may over-mask bright surfaces.
Use Case–Driven: For precise indices or change detection, CloudScore+ tends to be more reliable; for quick looks or rough mapping, QA60 can be sufficient.

Conclusion

CloudScore+ generally provides cleaner, more flexible masking at the cost of threshold tuning, while QA60 remains a simple baseline. The choice should reflect project goals, tolerance for omission/commission errors, and available time for calibration.