الفهرس 
IntroductionOnly 14 pages are availabe for public view
 |  | from | 136 |  |  |
| | | from | 136 | | |
Abstract
In recent years, the use of satellite imagery has grown significantly, creating a need for automated data processing and analysis. Data clustering algorithms are a popular choice for segmentation because they group similar data points together, allowing for the identification of features and patterns within the imagery. Image segmentation is used to segment an image into coherent parts using two types of segmentation techniques: semantic segmentation and instance segmentation. Semantic segmentation segments an image into meaningful parts or predefined class labels, while instance segmentation classifies each instance of an object as itself. Autoencoders, a type of neural network, unsupervised feature learning , are useful for dimensionality reduction and feature extraction as they learn to compress and reconstruct data. The principal objective of this dissertation is to improve and develop data clustering techniques and artificial intelligence on remote sensing data. Toward this objective, this dissertation develops two intelligent automated systems that may assist personnel related to remote sensing and GIS service.
First one: land cover classification involves using autoencoders for satellite imagery segmentation, assessing various segmentation architectures and encoders. Our analysis revealed that the combination of the DeepLabv3+ decoder architecture and ResNet-152 encoder architecture provided the highest performance, with an Intersection over Union (IoU) score of 90.34%, F-1 score of 95.07%, Precision of 95.02%, and Recall of 95.15%. This approach can aid in developing automated tools for processing satellite imagery in various fields, such as agriculture, land use monitoring, and disaster response.
Second one: building footprint extraction involves extracting building footprints from low-resolution satellite imagery. Our proposed model outperformed the most recent approaches it achieved an Average Precision (AP) of 93.074%. Extracting building footprint from aerial photos and satellite imagery is crucial in detecting changes, urban development, and agricultural land encroachments. Deep neural networks with feature extraction capabilities provide high accuracy in detecting and extracting building footprints. However, most approaches for extracting building footprints require high-resolution imagery in the sampling training data and inference phases, which is not always freely or readily available.
This Thesis Contains 6 Chapters and an Abstract (English and Arabic):
Chapter 1: The present work discusses the concept of Satellite Imagery Segmentation, along with the various data characteristics that influence it. The challenges faced by Remote Sensing experts in selecting the optimal approach are also highlighted. Artificial intelligence techniques and data clustering are explored as potential solutions to address these challenges. This study summarizes the key contents of subsequent chapters, which will delve deeper into these topics.
Chapter 2: Discusses the remote sensing knowledge about land cover classification
and building footprint detection and extraction.
Chapter 3: We present a preliminary introduction to the data clustering techniques, machine learning approaches, deep learning methods, transfer learning, segmentation, as well as autoencouders that were utilized in our study.
Chapter 4: Entitled “Land Cover Classification using Satellite Imagery Segmentation
based on Data Clustering and Deep learning”.
Chapter 5: Entitled “An artificial system for Building Footprint Extraction from Low-Resolution Satellite Imagery”.
Chapter 7: Entitled “Conclusions and Future Works” This chapter explains the results and contributions that have been reached in this thesis. It also presents some suggestions and future work to develop work in the remote sensing fields.