CSDUNet: Automatic Cloud and Shadow Detection from Satellite Images Based on Modified U-Net

Detection of clouds and shadows in remote sensing imagery is important due to its wide range of applications. There are a lot of applications in remote sensing images such as monitoring of the environment, change detection etc. It is an important and booming research area. Ineffective and inaccurate cloud and cloud shadow masking will cause undesirable effects on different task that can be performed by using remote sensing images. Because of high spectral conglomeration and the spectral and temperature discrepancy of the underlying surface the detection of clouds and associated shadows is not candid. In this paper, we propose CSDUNet a modified U-Net network for precise pixel-wise semantic segmentation of cloud and its associated shadow from optical remote sensing images. It uses an encoder network and a decoder network. This method concatenated feature maps at different scales. We have proposed a novel network for cloud detection, which extract features corresponding cloud and shadow at different scales from multilevel layers to generate sharp boundaries. Which will help to detect clouds in heterogeneous landscape, under complex underlying surfaces with varying geometry. Experimental analysis on the Landsat satellite dataset proves that the proposed CSDUNet achieves a dice coefficient of 95.05%. Our method got 95.93% precision, recall of 94.71% and Jaccard index of 97.29%. CSDUNet achieves accurate detection accuracy and surpass several traditional methods.