A Deep Learning Method for Near-Real-Time Cloud and Cloud Shadow Segmentation from Gaofen-1 Images.

Mehdi Khoshboresh-Masouleh,Reza Shah-Hosseini,Rodolfo E Haber

doi:10.1155/2020/8811630

Mehdi Khoshboresh-Masouleh, Reza Shah-Hosseini + Show 1 more

Open Access

https://doi.org/10.1155/2020/8811630

Copy DOI

Abstract

In this study, an essential application of remote sensing using deep learning functionality is presented. Gaofen-1 satellite mission, developed by the China National Space Administration (CNSA) for the civilian high-definition Earth observation satellite program, provides near-real-time observations for geographical mapping, environment surveying, and climate change monitoring. Cloud and cloud shadow segmentation are a crucial element to enable automatic near-real-time processing of Gaofen-1 images, and therefore, their performances must be accurately validated. In this paper, a robust multiscale segmentation method based on deep learning is proposed to improve the efficiency and effectiveness of cloud and cloud shadow segmentation from Gaofen-1 images. The proposed method first implements feature map based on the spectral-spatial features from residual convolutional layers and the cloud/cloud shadow footprints extraction based on a novel loss function to generate the final footprints. The experimental results using Gaofen-1 images demonstrate the more reasonable accuracy and efficient computational cost achievement of the proposed method compared to the cloud and cloud shadow segmentation performance of two existing state-of-the-art methods.

Highlights

Most remote sensing satellites with high spatial resolution imaging have limited spectral channels due to device considerations [9]. e remote sensing images with limited spectral channels, such as Gaofen-1 images, often lack complete radiometric calibration parameters due to the absence of the thermal and the water vapor absorption channels [10]. e process of identifying clouds accurately and separating them from some features, i.e., coastlines or buildings, is highly complicated [11]. In this context, providing a solution to detect and eliminate clouds and cloud shadows from images with high spatial resolution in different scenes is of great importance. e process of eliminating the cloud and cloud shadow from images depends on the accuracy of the cloud and cloud shadow detection [12]
Two state-of-the-art methods, including Fast Multifeature Combined (Fast-MFC) [13], and Multiscale Convolutional Feature Fusion (MSCF) [54], were used for comparisons. Because of their known efficiency in cloud and cloud shadow segmentation from Gaofen-1 satellite images, these methods were selected. e Fast-MFC and MSCF were tested in this study from the beginning, using the same testing set that was applied for the testing of the proposed method
Fast-MFC is a statistical method of pattern recognition for cloud and cloud shadow detection from Gaofen-1 satellite images based on the Mean Absolute Error (MAE) and the Mean Relative Error (MRE). e first step of this method is implementing a threshold for segmentation based on spectral features and segmentation refinement based on a guided filter in order to generate the cloud initial range. en, the geometric features are combined with the texture features for improving the results of cloud detection and the final production of cloud maps

Summary

Introduction

Cloud and cloud shadow are among the causes of disruption in processing passive sensors’ images in remote sensing [1].e presence of cloud and cloud shadow in remote sensing images disrupts the processes that involve segmentation, classification, matching, and the production of 3D models [2,3,4]. e accurate detection of cloud and cloud shadow is a significant step in the multispectral image preprocessing [5].e most important studies on the cloud detection, such as the Global Cloud Monitoring Project [6], have used an Advanced Very High-Resolution Radiometer (AVHRR), AVHRR processing program for ice, snow, and cloud monitoring [7], and the International Satellite Cloud Climatology Project [8] has used the thermal channel data with low spatial resolution.Monitoring Earth using the high-spatial resolution remote sensing images has been of great interest during the recent years. Most remote sensing satellites with high spatial resolution imaging have limited spectral channels (e.g., red, green, blue, and near-infrared) due to device considerations [9]. E process of identifying clouds accurately and separating them from some features, i.e., coastlines or buildings, is highly complicated [11] In this context, providing a solution to detect and eliminate clouds and cloud shadows from images with high spatial resolution in different scenes is of great importance. Computational Intelligence and Neuroscience of recent studies using deep learning methods on visible and near-infrared channels images of Zi-Yuan 3 satellites with a spatial resolution of 5.8 meters, Gaofen-1 with a spatial resolution of 16/8 meters, and Gaofen-2 with a resolution of 4 meters indicate an improved cloud detection accuracy with a mean accuracy of 92%, while revealing the margin details of clouds and cloud shadows due to various complications at this level of image resolution is still a significant challenge [19, 20].

Methods

Results

Conclusion