Comparison of Masking Algorithms for Sentinel-2 Imagery

Masking of clouds, cloud shadow, water and snow/ice in optical satellite imagery is an important step automated processing chains. We compare the performance masking provided by Fmask (“Function mask” implemented FORCE), ATCOR (“Atmospheric Correction”) Sen2Cor (“Sentinel-2 on a set 20 Sentinel-2 scenes distributed over globe covering wide variety environments climates. All three methods use rules based physical properties (Top Atmosphere Reflectance, TOA) to separate clear pixels from potential pixels, but they different class-specific thresholds. The can yield results because definitions dilation buffer size for classes cloud, shadow snow. Classification are compared assessment expert human interpreter using at least 50 polygons per class randomly selected each image. assignment considered as reference or “truth”. carefully assigned label visual true color infrared false images additionally bottom atmosphere (BOA) reflectance spectra. most part comparison done difference area classifications considered. This classification where Fmask, disagree. Results have advantage show more clearly strengths weaknesses than complete overall accuracy ATCOR, areas 45%, 56%, 62%, respectively. User producer accuracies strongly class- scene-dependent, typically varying between 30% 90%. Comparison complemented looking all give same result. Overall that “same area” 97% resulting 89%, 91% 92%