Distinguishing Cloud and Snow in Satellite Images Via Deep Convolutional Network

Yongjie Zhan,Jian Wang,Jianping Shi,Guangliang Cheng,Lele Yao,Weidong Sun

DOI: https://doi.org/10.1109/lgrs.2017.2735801

IF: 5.343

2017-01-01

IEEE Geoscience and Remote Sensing Letters

Abstract:Cloud and snow detection has significant remote sensing applications, while they share similar low-level features due to their consistent color distributions and similar local texture patterns. Thus, accurately distinguishing cloud from snow in pixel level from satellite images is always a challenging task with traditional approaches. To solve this shortcoming, in this letter, we proposed a deep learning system to classify cloud and snow with fully convolutional neural networks in pixel level. Specifically, a specially designed fully convolutional network was introduced to learn deep patterns for cloud and snow detection from the multispectrum satellite images. Then, a multiscale prediction strategy was introduced to integrate the low-level spatial information and high-level semantic information simultaneously. Finally, a new and challenging cloud and snow data set was labeled manually to train and further evaluate the proposed method. Extensive experiments demonstrate that the proposed deep model outperforms the state-of-the-art methods greatly both in quantitative and qualitative performances.