Cloud and Cloud Shadow Detection for Multi-Modal Imagery With Gap-Filling Applications

Cloud and cloud shadow (CCS) detection algorithms play a crucial role in the preprocessing of remote sensing data and directly affect the accuracy of subsequent analyses, making them an essential step in most analytical processes. Recent techniques for detecting CCS often employ deep learning methods, which are effective but typically require extensive training data specific to each type of satellite imagery. This study presents a new methodology that applies a model trained on the preconstructed KOMPSAT-3/3A CCS dataset to Landsat-8 and Sentinel-2 satellite imagery. The experimental results demonstrated that the CCS detection model based on KOMPSAT-3/3A achieved a mean F1 score of 0.846 on the test dataset. When applied to Landsat-8 SPARCS and Sentinel-2 CloudSEN12 test datasets, it also showed high performance, with mean F1 scores of 0.741 and 0.8, respectively, effectively indicating that multi-modal CCS detection can be successfully implemented. Applying this model to different sensor imagery confirmed its effectiveness in gap filling, which can be utilized to enhance time-series analyses where continuous monitoring is required. In conclusion, this approach not only proves beneficial for time-series analysis but also significantly reduces the time and effort required to build datasets in deep learning-based CCS detection.