Abstract

On going advancements in remote sensing (RS) technology have made a substantial amount of satellite image data available for various applications. Cloud cover, a major issue on many regions on Earth surface, impedes the acquisition of valuable Earth observation data by optical satellite imaging systems and adversely impacts the processing and utilization of optical satellite images. Because of the growing interest in remote sensing and deep learning, cloud detection has become a crucial preprocessing step in preparing optical satellite images for operational systems. Deep learning needs massive data. Transfer learning methods are found as an efficient solution accurate categorization of satellite images with limited dataset. This work presents an integrated strategy comprising the transfer learning based InceptionV3 pre trained model as a feature extractor and Random forest as an ensemble learning based classifier. This study proposes Sentinel-2 optical satellite imagery based new dataset CloudS2 dataset, which has two classes and hundreds of high-resolution images. The proposed model works effectively on limited dataset further improving the network’s performance. The proposed model employs the newly built CloudS2 dataset and classifies images into respective classes. The InceptionV3-RF model outperforms the existing deep learning-based models such as VGG-16, MobileNetV2, Xception, DenseNet201 and InceptionResNetV2 and achieves an overall accuracy of 86.6%. The proposed research work is useful as a preprocessing stage in deep learning applications applied to remote sensing optical satellite imagery.