Multi-Temporal Landsat Data Automatic Cloud Removal Using Poisson Blending

Changmiao Hu,Zheng Zhang,Lian-Zhi Huo,Ping Tang

doi:10.1109/access.2020.2979291

Changmiao Hu, Zheng Zhang + Show 2 more

Open Access

https://doi.org/10.1109/access.2020.2979291

Copy DOI

Abstract

Cloud and cloud shadow are common issues in optical satellite imagery, which greatly reduce the usage of data archive. As for the Landsat data, great advances have been made on detecting cloud and cloud shadow. However, few studies were performed on Landsat cloud removal for large areas. To facilitate land cover dynamics studies with high temporal resolution, we present an automatic cloud removal algorithm in this paper. Specifically, For Landsat Collection 1 Level-1 surface reflectance products, the algorithm first builds a cloud mask from the Quality Assessment (QA) band, and then reconstructs cloud-contaminated portions based on multi-temporal Landsat images with temporal similarity. To further eliminate radiation differences between cloud-free and reconstructed regions, a Poisson blending algorithm is adopted. Besides, the efficiency of gradient-domain compositing is accelerated by the quad-tree approach. Experiments have been performed to process more than 50,000 Landsat 8 Operational Land Imager (OLI) images covering China from 2013 to 2017, which yield promising results in terms of radiometric accuracy and consistency for experimental images with cloud coverage less than 80%. The produced Landsat time series images with cloud removal can be further used for analyzing land cover and land change dynamics in China, and the proposed algorithm should be easily employed to produce cloud-free Landsat time series for other areas.

Highlights

Remote sensing imagery has been widely used for different applications, especially with recent advancements on machine learning algorithms [1]–[8]
To achieve automatic cloud patching for large amounts of Landsat images efficiently and effectively, we propose an automatic cloud removal algorithm based on Poisson blending for multi-temporal Landsat data
To further quantitatively compare the two approaches, root-mean-square error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) index, are used to evaluate results, and the results are shown in the Table 1

Summary

Introduction

Remote sensing imagery has been widely used for different applications, especially with recent advancements on machine learning algorithms [1]–[8]. Equipped with many advanced data analysis tools, we still face another great challenge in optical satellite image analysis: cloud. Cloud and cloud shadow are common issues in optical satellite imagery, which limit the power of optical images and increase the difficulty of time series analysis. Cloud detection and removal has always been an important issue in remote sensing image processing. Cloud removal techniques are mainly based on single or small-scale images due to limited remote sensing data sources and limited cloud detection accuracy. Cloud and cloud shadow hinder further processing of Landsat time

Methods

Results

Conclusion