Faster-RCNN Based Cloud Region Recognition Algorithm for Single Images

Abstract

Remote sensing imaging is frequently disturbed by clouds which leads to unclear images with low contrast and poor resolution. Cloud obstacles usually arouse the valuable information loss of remote sensing images. Technology of cloud removal from single remote sensing images has attracted worldwide interests since only one image is available. When clouds are distributed unevenly and only a small portion of the images is covered by clouds, it is expected to preserve image information outside of the clouds as much as possible during cloud removal. In this paper, an algorithm based on Faster-RCNN was proposed to detect the cloud regions in the images before cloud removal. The principle of cloud region recognition was analyzed. The structure of Faster-RCNN was introduced. Three convolutional networks i.e. MobilenetV2, Resnet50 and VGG16 were introduced and applied as the backbone of Faster-RCNN respectively. Cloud region recognition algorithm was developed based on Faster-RCNN. Training and testing data sets were established and labeled by applying the proposed remote sensing imaging simulation algorithm to add clouds to the clear remote sensing images. Some experiments were carried out when Faster-RCNN selected MobilenetV2, Resnet50 and VGG16 as its backbone and was trained to optimize the parameters. Their results were compared. It proved the proposed algorithm with MobilenetV2 as the backbone achieved a successful recognition rate of 95% which supports the following cloud removal.