Automated VNIR-only Cloud Detection for the WorldView Series and Similar Spacecraft

Abstract. The WorldView-110 imager on board the WorldView-3 spacecraft has spatial resolution that is an order of magnitude higher than Landsat or Sentinel platforms. This imager instrument has been in extensive use by the land applications community. It presents a unique opportunity to atmospheric scientists as well. However, cloud masking remains a challenge. Currently available cloud products from Vantor Holdings Inc. are limited in detail and quality, and cloud coverage is represented as a percentage for the entire satellite image scene. Additionally, Vantor does not make its cloud masking product public and does not disclose the methodology behind it. The limited spectral channel complement of WorldView-3 also presents a challenge in detecting clouds.

We have developed a unique method for detecting clouds in WorldView-3 scenes using only visible and near-infrared (VNIR) spectral channels. VNIR-only acquisitions dominate the available WorldView-3 data record. Unlike traditional algorithms, this pixel-level algorithm, Cloud Mask for Ultra-high Resolution Spaceborne Imagers (CM-URSI), does not rely on any sources of a priori surface classification. The algorithm detects clouds and clear sky correctly, including in challenging scenes, such as reflectively varied snow/ice present underneath thin cloud cover. Validation proved challenging due to the scarcity of reference datasets that match the high image resolution. In the limited situations where in-situ or lidar data existed for validation, CM-URSI results agreed with in-situ data 100 % of the time. Though this result is extremely encouraging, additional validation is necessary prior to any large-scale application.

In this paper, we show CM-URSI cloud masking results for a variety of scenes and initial validation. We describe WorldView data preparation and give some background on previous attempts at solving the WorldView cloud masking problem. We also discuss future directions such as extensions to commercial spacecraft with fewer channels than WorldView-3 and use in conjunction with computer vision machine learning methodologies.