Cloud removal analysis of remote sensing satellite imagery

Abstract

Remote sensing satellite images are widely used for many applications, such as land use land cover mapping, agriculture, forestry, marine, disaster monitoring, etc. Unfortunately, there is an obstacle regarding cloud cover. The existence of clouds interferes remote sensing activities. Moreover, about two third of the earthrs surface is covered by cloud at any one time. This issue increases the difficulty to support operational remote sensing monitoring applications.This thesis developed, tested and applied methods for cloud masking and removal that work in tropical regions. Four specific objectives were addressed: (1) developing a cloud and cloud shadow masking method for Landsat 8 images in Indonesia, ensuring accuracy is assessed over heterogeneous land cover; (2) improving the ability of the proposed method for detecting cloud and cloud shadow for Landsat 8 images with a variety of cloud types in Indonesia; (3) developing a method for filling masked pixels in Landsat 8 images in Indonesia; and (4) extension a cloud and cloud shadow masking method for Sentinel-2 images.After introductory, review and research approach chapters one - three, chapter four developed a method for cloud and cloud shadow masking called Multi-temporal Cloud Masking (MCM) for Landsat 8 images in tropical environments. This method used two kinds of data: (1) target image and (2) reference image. A target image is an image may have cloud and cloud shadow contaminated pixels and a reference image is a cloud-free image. This method utilizes the difference of reflectance values between those images. The MCM algorithm was applied and evaluated in a part of Indonesia which has a heterogeneous land cover. This chapter concluded that the MCM algorithm can be used to accurately detect cloud and cloud shadow for Landsat 8 images from areas with a heterogeneous land cover in tropical environments.Chapter five developed and assessed an approach for cloud and cloud shadow removal called Multi-temporal Cloud Removal (MCR) for Landsat 8 images. This approach has three main steps: (1) radiometric correction; (2) cloud and cloud shadow detection; and (3) image reconstruction. In this chapter, the previous MCM was improved for detecting a variety of cloud types. By using images from a sequence acquisition date, the issue of varying land cover changes was minimised. In the image reconstruction step, a method for filling masked pixels was applied. At the end of the process, cloud and cloud shadow free image was generated. To evaluate the results, selected Landsat 8 images over a limited area in Indonesia with heterogeneous land cover and a variety of cloud types were tested. As a result, cloud and cloud shadow contaminated pixels on Landsat 8 images were able to be removed in several test cases. The resultant images had high visual and statistical similarities with the reference image. This evaluation provided a preliminary indication of the utility of MCR in removing cloud and cloud shadow for Landsat 8 images.Chapter six expanded the MCM approach designed for Landsat 8 images for detecting cloud and cloud shadow to Sentinel-2 images. Fortunately, the spectral bands of Landsat 8 have similarity with the spectral bands of Sentinel-2. The results showed that the userrs accuracy and the producerrs accuracy in detecting cloud and cloud shadow were significantly high. Compared to the Fmask algorithm, the MCM has higher accuracies in both detecting cloud and cloud shadow. Our results provide a preliminary demonstration that the MCM algorithm can be used for detecting cloud and cloud shadow for Sentinel-2 images, especially in Indonesia.Chapter four to six achieved the four objectives of this thesis. Chapters seven and eight were added to expand the study area from tropical areas such as Indonesia to global areas, to test the reliability of the method. Chapter seven improved the previous MCM algorithm to be applied in detecting cloud and cloud shadow for Sentinel-2 images in a global area such as Sub-tropical South, Tropical and Sub-tropical North. In statistical evaluations, both userrs and producerrs accuracies were quite high. The comparison between MCM and Fmask showed that MCM has higher accuracy especially in detecting cloud shadow. This chapter concluded that the improved MCM algorithm can be used to detect clouds and cloud shadows for Sentinel-2 across a range of environments globally and the accuracies were significantly high.Chapter eight improved, applied, and evaluated the MCM algorithm for detecting cloud and cloud shadow for Landsat 8 images in a variety of environments. The results showed that the userrs accuracy and the producerrs accuracy of the new-MCM algorithm in detecting cloud and cloud shadow were quite high and higher than L8 CCA results. Therefore, this chapter concluded that the MCM algorithm can be used to detect cloud and cloud shadow for Landsat 8 images in a variety of environments and the accuracies were significantly high.This thesis provided a fundamental framework for cloud and cloud shadow masking and removal from moderate spatial resolution, multi-temporal satellite data over heterogeneous land cover types, a variety of cloud types, and different environments in tropical areas, e.g. Indonesia, and other areas globally. The proposed methods presented the capability to detect cloud and cloud shadow in several kinds of satellite image such as Landsat 8 and Sentinel-2. This thesis has successfully provided an alternative method for cloud and cloud shadow masking and removal with high accuracies.n