Content Details | NUML Online Research Repository

Title Cloud detection in remote sensing images using deep learning

Author(s) Maria Amjad

Abstract Remote sensing images play a vital role in the analysis of the earth’s surface. The earth analysis is productive if the sky is clear because clouds obscure the earth surface and create problems for remote sensing applications such as change detection, agriculture, surveillance, urban and rural planning. Various methods have been proposed for the detection of clouds, which vary from pixel intensity transformation based methods to deep learning methods. The intensity transformation methods are generally fast but they are susceptible to the variation in the pixel intensities, illumination changes and noise. On the other hand, the deep learning methods are efficient and accurate but require training on dataset(s) prior to cloud detection. In this thesis, You Only Look Once (Yolo) algorithm is investigated for the cloud detection. Yolo has been successfully applied for the detection and recognition of real life objects in indoor and outdoor images. In this thesis, the Yolo algorithm is combined with other three state of the art deep learning algorithms in order to improve its accuracy for cloud detection. These algorithms are Practical Portrait Human Segmentation Lite (PP-HumanSeg_Lite), Deep Dual Resolution Network (DDRNet) and Disentangled Non-Local Network (DNLNet). All these algorithms have been used for the semantic segmentation of the real life objects. The combination of Yolo with PP-HumanSeg_Lite, DDRNet and DNLNet is done through an ensemble learning method where the responses of Yolo and the other algorithms are combined and provided to Random Forest for accurate cloud detection. Experiments are performed on two different cloud datasets which are High Resolution Cloud Detection (HRCD) Dataset and 38-Cloud Segmentation datasets. The experimental result shows that Yolo+PP-HumanSeg_Lite give the best results and achieves accuracy of 96% and 93% on HRCD and 38-Cloud datasets, respectively. Whereas, Yolo achieves 91.2% and 81.5% accuracy, respectively.

Type Thesis/Dissertation MS

Faculty Engineering and Computer Science

Department Computer Science

Language English

Publication Date 2023-01-12

Subject Computer Science

Publisher

Contributor(s)

Format

Identifier

Source

Relation

Coverage

Rights