CMIX: Cloud Mask Intercomparison eXercise

Kurzfassung

Cloud cover is a major limiting factor in exploiting time-series data acquired by optical spaceborne remote sensing sensors. Multiple methods have been developed to address the problem of cloud detection in satellite imagery and a number of cloud masks have been developed for optical sensors but very few studies have carried out quantitative intercomparison of state-of-the-art methods in this domain. Here, we summarize results of the first Cloud Masking Intercomparison eXercise (CMIX) conducted within the Committee Earth Observation Satellites (CEOS) Working Group on Calibration & Validation (WGCV). CEOS is the forum for space agency coordination and cooperation on Earth observations, with activities organized under working groups. CMIX, as one such activity, is an international collaborative effort aimed at intercomparing cloud detection algorithms for moderate-spatial resolution (10-30 m) spaceborne optical sensors. The focus of CMIX is on open and free imagery acquired by the Landsat 8 (NASA/USGS) and Sentinel-2 (ESA) missions. Ten algorithms developed by nine teams from fourteen different organizations representing universities, research centers and industry, as well as space agencies (CNES, ESA, DLR, and NASA), were evaluated within the CMIX. Those algorithms varied in their approach and concepts utilized which were based on various spectral properties, spatial and temporal features, as well as machine learning methods. Algorithm outputs were evaluated against existing reference cloud mask datasets. Those datasets varied in sampling methods, geographical distribution, sample unit (points, polygons, or full image labels), and generation approach (experts annotations, machine learning, or sky images). Overall, the performance of algorithms varied depending on the reference dataset, which can be attributed to differences in cloud definitions used when producing the reference datasets. Average overall accuracy (across algorithms) varied 80.0±5.3% to 89.4±2.4% for Sentinel-2, and 79.8±7.1% to 97.6±0.8% for Landsat 8, depending on the reference dataset. An overall accuracy of 90% yields half the errors than an overall accuracy of 80%. The study identified algorithms that provided a balance between commission and omission errors, as well as algorithms, which are cloud conservative (high user’s accuracy) and non-cloud (clear) conservative (high producer’s accuracy). With repetitive observations like those of Sentinel-2, it seems reasonable to favor non-cloud conservative approaches, with maybe the exception of very cloudy regions where every cloud free observation is critical. When thin/semi-transparent clouds were not considered in the reference datasets algorithms’ performance generally improved: overall accuracy values increased by +1.5% to 7.4%. It should be noted though that these clouds are commonly occurring and are often present in optical imagery. Within CMIX, we also developed recommendations for further activities, which include provision of a quantitative definition for clouds (targeting moderate spatial resolution imagery by Landsat 8 and Sentinel-2), generation of new reference datasets, and expansion of the analysis framework (for example, multi-temporal analysis and application-driven validation). Such intercomparison studies will hopefully help the community to improve the algorithms and move towards standardization of cloud masking. Given the importance of cloud masking in optical satellite imagery we encourage CEOS to continue the CMIX activities.