IDENTIFIKASI AWAN PADA DATA TIME SERIES MULTITEMPORAL MENGGUNAKAN PERBANDINGAN DATA SEKUENSIAL

Abstract

Cloud identification is an important pre-processing step of remote sensing data.
Generally, cloud identifications could be classified into single-date and multi-date methods. Furthermore, the single-date method could be divided into physical-rules-based and machine-learning-based. Physical-rules-based method generally need data with sufficient spectral resolution while machine-learning-based method depend on training dataset. While the multi-date method usually using clear data as a reference. The clear data itself could be a whole scene or built from many scenes. Processing cloud-free data is a challenge in areas with high cloud coverage such as Indonesia. In this paper, a cloud identification method using multi-date time series scenes is proposed. This method only uses RGB channels which are common in remote sensing visual data. In addition, this method does not require or process cloud-free data mosaics in advance. A pixel value from an examined scene is compared to other pixel values from other scenes in the same position. The other scenes are the scenes that were acquired before and after the examined scene. The value differences between the examined pixel and it's before and after then evaluated using some thresholds to determine whether the pixel is a cloud or not. Assessment is done by using L8 Biome as a reference. The result shows that using some thresholds in our proposed method has a Kappa coefficient higher than 0.9.