A Novel Cloud Removal Method Based on IHOT and the Cloud Trajectories for Landsat Imagery

Shuli Chen,Wei Yang,Jin Chen,Xuehong Chen,Miaogen Shen,Xin Cao,Xihong Cui,Xiang Chen

doi:10.3390/rs10071040

Abstract

Cloud removal is a prerequisite for the application of Landsat datasets, as such satellite images are invariably contaminated by clouds. Clouds affect the transmission of radiation signal to different degrees because of their different thicknesses, shapes, heights and distributions. Existing methods utilize pixel replacement to remove thick clouds and pixel correction techniques to rectify thin clouds in order to retain the land surface information in contaminated pixels. However, a major limitation of these methods refers to their deficiency in retrieving land surface reflectance when both thick clouds and thin clouds exist in the images, as the two types of clouds differ in the transmission of radiation signal. As most remotely sensed images show rather complex cloud contamination patterns, an efficient method to alleviate both thin and thick cloud effects is in need of development. To this end, the paper proposes a new method to rectify cloud contamination based on the cloud detection of iterative haze-optimized transformation (IHOT) and the cloud removal of cloud trajectory (IHOT-Trajectory). The cloud trajectory is able to take consideration of signal transmission for different levels of cloud contamination, which characterizes the spectral response of a certain type of land cover under increasing cloud thickness. Specifically, this method consists in four steps. First, the cloud thicknesses of contaminated pixels are estimated by the IHOT. Second, areas affected by cloud shadows are marked. Third, cloud trajectories are fitted with the aid of neighboring similar pixels under different cloud thickness. Last, contaminated areas are rectified according to the relationship between the land surface reflectance and the IHOT. The experimental results indicate that the proposed approach is able to effectively remove both the thin and thick clouds and erase the cloud shadows of Landsat images under different scenarios. In addition, the proposed method was compared with the dark object subtraction (DOS), the modified neighborhood similar pixel interpolator (MNSPI) and the multitemporal dictionary learning (MDL) methods. Quantitative assessments show that the IHOT-Trajectory method is superior to the other cloud removal methods overall. For specific spectral bands, the proposed method performs better than other methods in visible bands, whereas it does not necessarily perform better in infrared bands.

Highlights

Landsat data have been widely employed to investigate the Earth’s surface in applications such as urbanization, hazard monitoring, deforestation, and human-environment interactions as results of land use and land cover change [1,2,3,4,5]
The haze optimized transformation (HOT)-dark object subtraction (DOS) and iterative haze-optimized transformation (IHOT)-DOS were only limited to the removal of thin clouds and are not able to address the issues of thick clouds and shadows (Figures 5 and 6)
As the cloud contamination pattern is very complex with mixed clouds (Figure 5a), both the haze optimized transformation-dark object subtraction (HOT-DOS) and IHOT-DOS methods are ineffective in restoring the correct color of the land cover of green cropland, bare soil, and urban land

Summary

Introduction

Landsat data have been widely employed to investigate the Earth’s surface in applications such as urbanization, hazard monitoring, deforestation, and human-environment interactions as results of land use and land cover change [1,2,3,4,5]. One remaining issue of the Landsat images is that they are invariably contaminated by clouds, which significantly affect the electromagnetic signal transmission in the production of images [3,6,7,8,9]. The effective detection and removal of the cloud contamination are critical steps before the usage of Landsat imagery [3,10,11,12]. Thick clouds block the majority of radiation from the land surface; and pixel substitution is the only way to remedy the information loss [13]. Zhu et al [14] and Cheng [15] made effective attempts at substituting the clouded pixels with the spatiotemporal information derived from their neighboring similar pixels. Li et al [16]

Methods

Results

Conclusion