Assessing the Impact of Threshold on Cloud Masking Algorithms for Downstream Tasks

This study presents an entropy-based threshold sensitivity analysis of cloud masking (CM) algorithms, quantifying how different threshold values affect scene uncertainty and downstream segmentation performance. We evaluate two widely used CMs, Cloud Score+ and s2cloudless, across a wide range of thresholds using the CloudSEN12+ benchmark dataset. In addition to traditional accuracy metrics, including Overall Accuracy (OA), F1-score (F1), and Intersection over Union (IoU), we considered the mean relative difference in object counts ($\overline{\Delta NO}$), derived from Segment Anything Model (SAM) outputs, and the mean entropy difference ($\overline{\Delta H}$). Our results reveal a clear trade-off between uncertainty reduction and information loss as masking becomes more aggressive, and show that threshold values that maximize pixel-level accuracy do not necessarily optimize segmentation outcomes. The primary contribution is to show that choosing the right threshold is crucial for balancing the elimination of cloud-related uncertainty with the preservation of valuable information that improves the quality of downstream tasks.