Megh Sansuchak: A Cloud Mask Algorithm for High Resolution Panchromatic Satellite Imagery

In the era of Remote Sensing application where generation of meaningful hypothesis plays an important role in decision making of certain situation, getting high resolution data with precise features is a necessary and important task. In that context, presence of clouds in satellite imagery sometimes misleads to high reflectance features. This problem becomes more challenging when working with panchromatic (gray band) images as the signature of cloud and high reflectance features like dessert and salt land are somewhere similar in terms of intensity. To address this issue, a cloud mask generation algorithm is proposed in this paper. This algorithm consists of fully connected convolutional neural network (CNN) based on U-net architecture. This algorithm which is named as Megh-Sansuchak requires no specific preprocessing. Experimental results show a significantly accurate cloud mask on a long strip of Cartosat-1 Panchromatic Imagery. Performance matrices for Megh-Sansuchak are estimated with respect to different loss functions.