Cloud removal in landsat images based on spatial-spectral-temporal random forests

Abstract The Landsat-8 Operational Land Imager (OLI), together with the Landsat-9 OLI2 sensors, provides important optical data for various applications. Due to frequent cloud cover, the obtained Landsat images always contain different degrees of spatial information loss. Although great efforts have been made to remove clouds in remote sensing images in recent decades, there still exists a large challenge, that is, the land cover changes between the target cloudy and the auxiliary cloud-free images, which limits greatly the accuracy of cloud removal. To cope with this issue, a spatial-spectral-temporal random forest (SSTRF) algorithm was proposed in this paper. The SSTRF model aims at removing clouds in Landsat images with the assistance of a remote sensing image with spatially complete cover from another satellite, that is, the 8-day MODIS surface reflectance product (MOD09A1). By compositing the valid data from 8 separate days, the MOD09A1 product has a great possibility to provide cloud-free land cover change information for the cloud removal of Landsat images. Through experiments with simulated and real clouds, SSTRF was demonstrated to produce greater accuracy than two benchmark cloud removal methods and has great potential to be applied to practical use.