Machine Learning Approach For Cloud Detection With Satellite Data

This paper provides a comparative analysis of methods for cloud classification and cloud mask generation using multispectral remote sensing data. It evaluates the performance of traditional threshold-based techniques alongside advanced machine learning methods, focusing on Multilayer Perceptrons (MLP) and Convolutional Neural Networks (CS-CNN). MLP, with its feedforward neural network architecture, is highlighted as an effective approach for pixel-level classifications due to its accuracy and simplicity. However, its inability to directly utilize spatial context limits its performance in complex scenarios. In contrast, CS-CNN stands out as an innovative method that integrates both spatial and spectral information, delivering high-accuracy segmentation without requiring predefined feature extraction. The study also explores the contribution of specific spectral bands in cloud detection and classification using multispectral satellite data. Visible bands, such as 0.6 μm and 0.8 μm, are instrumental in determining aerosol scattering and surface-cloud contrast. The 1.6 μm near-infrared band effectively distinguishes snow, ice, and water clouds, while the 3.9 μm infrared band is crucial for detecting low clouds. Thin cloud temperatures are measured using the 6.2 μm water vapor band, whereas the 10.8 μm infrared band is vital for identifying surface and cloud-top temperatures, volcanic ash, and cirrus clouds. The results demonstrate that CS-CNN outperforms traditional threshold-based methods and the EUMETSAT cloud mask algorithm (CLM). While MLP achieves an accuracy of 88.96% and CLM reaches 86.10%, CS-CNN's ability to incorporate contextual information yields even higher accuracy and flexibility. This makes CS-CNN particularly effective in scenarios where spatial context is critical. The study underscores the potential of machine learning-based approaches in remote sensing applications, positioning CS-CNN as a powerful tool for cloud detection and classification. These findings highlight the significant role of advanced algorithms in enhancing the accuracy and efficiency of remote sensing tasks.