Generation of Cloud-free Imagery Using Landsat-8

Cloud cover and cloud shadow areas on satellite imagery restrict the practical use of remote sensing data. Thus, cloud screening and filling methods are critical for geospatial users. The recently launched Landsat-8 provides coastal/aerosol and cirrus bands to tackle this problem. In this case, clouds can be accurately detected with Landsat-8 data, and the masked cloud areas could be filled in with image processing methods. This paper presents a novel method for detection and filling of cloud and its shadow areas using the Landsat-8 sensor. First, cloud and its shadow areas are detected from the Landsat-8 bands using Otsu’s thresholding method. The detected cloud and cloud-shadow areas are then replaced using coordinates of a reference image and pixel values from an experimental image corresponding to the coordinates of reference image. Experimental results using the Landsat-8 dataset indicated that the proposed method generates a superior quality of cloud free imagery. Key word: Cloud detection, Cloud removal, Landsat-8, Otsu’s N thresholding INTRODUCTION Landsat is one of the most popular and oldest satellite data sources for observation of Earth’s land surface. A great deal of research using Landsat series data has been conducted for various applications, such as change detection and land-cover classification, because of the medium spatial resolution and spectral variance of these data. However, cloud and its shadows remain serious obstacles when using Landsat data, especially when monitoring land surface. Therefore, a method for replacing the cloud and shadow areas with original land surface is needed to improve image quality and availability when dealing with Landsat data (Xiaolin Zhu, 2012). Many studies have focused on cloud detection, the use of thresholds, and replacement of the cloud and its shadow areas. The study by MA Ying-zhao (2010) on cloud detection used a decision tree based on Landsat-5 thresholds. Min LI (2002) used pixel-ranking, which categorizes the images using thresholds as cloud, shadow, vegetation, and buildings. The results of these studies are good; however, a few disadvantages are noted. For example, the thresholds are chosen by manual inspection, which is time consuming, and the methods are inadequate for applying to any cloud cover images. To tackle this problem, Yi-Shiang (2011) used an automatic thresholds only for thick cloud detection, but did not deal with the shadow areas. Cloud cover areas have been filled using a regression tree and histogram matching (Helmer and Ruefenacht, 2005). Xiaolin Zhu (2012) suggested a modified neighborhood similar pixel interpolator approach, which is generally used to fill gaps in Landsat ETM+ scan line corrector-off images. Suming (2005) devised a concept of a spectral similarity group (SSG) filling method. Recently, Landsat-8 was successfully launched and it has provided greatly enhanced spectral information with the addition of two new spectral bands: a deep blue visible channel (band 1) and a new infrared channel (band 9). These new bands allow the problem caused by cloud and shadow to be resolved more easily compared with previous Landsat-5 and 7 images. The infrared channel (band 9), called the cirrus band, is useful for cloud detection because its wavelength is from 1.360 to 1.390μm. This wavelength of band 9 includes a strong water vapor absorption wavelength area, and water vapor is generally concentrated in the lower atmosphere (Hutchison, 1996), which means that incident solar energy in this channel diminishes prior to reaching the land surface. When this incident solar energy is reflected into space, a similar phenomenon will happen. However, this process does not happen in cloud areas because water and ice in the cloud block incident solar energy and reflect it directly into space. For this reason, DN (Digital Number) is higher in cloud areas than at other surfaces, and cloud areas can be easily detected. This paper proposes a process for generating a cloud-free image based on Landsat-8 images. Two sets of Landsat-8 images were acquired, one of which included cloud areas. Automatic thresholds, using Otsu’s thresholding, were applied to the experimental image in order to create a cloud and cloud shadow masking area. This masking area was ultimately filled in with coordinates from a reference image and pixel values from an experimental image corresponding to the coordinates of the reference image. MATERIALS AND METHODS Materials The proposed method was applied to Landsat-8 satellite images, with 30m spatial resolution in a multispectral image, acquired in April, May, and, June, 2013. Two study sites were chosen, including two images per site: one experiment scene, which includes cloud and its shadow area and one reference scene, which is a clear image taken at a different time (Table 1, Figure 1). Table 1. Landsat data sets for generation of cloud free images Path / Row Experiment image Reference image 38 / 35 24 May 2013 9 June 2013 28 / 45 16 April 2013 2 May 2013 Figure 1. Experimental (left) and Reference (right) images (R, G, B), (a) P38/R35, (b) P28/R45 Landsat-8 has two new bands in addition to those provided by Landsat-5 and 7. These new bands are helpful for detecting clouds, so fewer bands are needed to detect clouds and cloud shadows. The generation of a cloud-free image consists of two major parts (figure 2). The first part produces a cloud and cloud-shadow mask, using Otsu’s thresholding method. The second part fills in the cloud and cloudshadow area, based on an SSG filling method (Suming, 2013). Figure 2. Flow chart Cloud masks The process of generating a cloud mask is divided into two parts: thick cloud masking and thin cloud masking. Both processes used spectral information of Landsat bands 1, 9, and 10 and Otsu’s thresholding method was applied. Otsu’s thresholding is an automatic method for finding an optimum global threshold value in a histogram (Otsu, 1975). Thick cloud was detected by selecting thresholds of bands 9 and 10. Thick cloud usually has high values in band 9, because water and ice in the cloud area usually block penetration of light. Therefore, incident solar energy is strongly reflected in the cloud area at the wavelength of band 9. On the other hand, the temperature of the cloud area is low compared with other areas, so band 10 has low values in cloud areas. Thin cloud was masked using band 1 in addition used to bands 9 and 10. Thin cloud, such as haze, has a small amount of ice in the cloud area, so that the sunlight passes through and reaches the land surface area; therefore, it usually has a low pixel value in band 9. In addition, thin cloud temperature is low, which means lower pixel values in band 10. Clouds are also normally brighter than other features, especially in the blue band, which means higher pixel values in band 1. The conditions for threshholding of each process of cloud masking are shown in table