Improved cloud screening of OceanSat-3 OCM-3 satellite imagery using machine learning algorithm

Abstract

Cloud masking in satellite imagery is critical for quantitative remote sensing research and its practical applications. However, accurate cloud detection in satellite imagery acquired by the sensors with limited spectral bands remains a challenge. Here, we present a machine learning (ML) approach such as Support Vector Machine (SVM) and Random Forest (RF) for improved cloud screening of satellite imagery acquired by the Ocean Color Monitor-3 (OCM-3) onboard the OceanSat-3 (EOS-06). Adaptive threshold (AT) technique is also used to comprehend efficient cloud screening by ML algorithms. Spectral reflectance and cloud indices derived from OCM-3 measurements in the visible and near-infrared bands are used as ML features. Pixel-level comparisons with visually inspected reference cloud masks over distinct geographic regions of India and adjacent oceanic regions are conducted to evaluate the performance of both ML and AT algorithms. The results reveal that the ML algorithm outperforms the AT algorithm in most metrics for both thick and thin cloud detection. The ML algorithm demonstrates an accuracy of ∼94% for all types of clouds, compared to 84% for the AT algorithm. Overall, this study suggests that underlying surface-specific training samples are crucial for different cloud types to achieve improved cloud screening across diverse geographic regions.