Design and implementation of a satellite and aerial image processing platform to update spatial objects of geographic information system (GIS) for monitoring urban changes

Object detection and identification from remotely sensed data, especially Buildings, can be considered the foundation of updating Geographic Information system data for improving and monitoring the infrastructure of cities. Small-scale objects like buildings may now be recognized because of the development of extremely high-resolution remote-sensing images. However, manually separating the buildings from the images requires substantial processing time. Comprehensive research has been conducted on various approaches, including traditional image processing techniques, supervised and unsupervised machine learning methods, and deep learning architectures. The literature survey explores the evolution of feature extraction algorithms, classification models, and their applications in urban environments. Notable studies on semantic segmentation, object-based image analysis, and multi-sensor data integration are discussed. Additionally, insights are provided into the challenges and limitations faced by current techniques, paving the way for the development of novel strategies proposed in this thesis. The synthesis of this extensive review establishes a foundation for the research, highlighting gaps in the current state-of-the-art and setting the context for the proposed advancements in automated building detection. Therefore, a robust building detection methodology is necessary. We present two approaches for extracting buildings from high-resolution images. The first approach is based on a supervised machine learning technique, and for the second, we use deep learning methods. In the supervised approach, the image is firstly divided into superpixel patches, from which the colors and texture features are retrieved. Buildings, roads, trees, and shadows are then separated into four groups using the Support Vector Machines technique (SVM). The approximate location of the building has been determined using a seed point start and an adaptive regional growth approach based on the previously known position of the shadows. A contouring procedure involving an open morphological operation was applied to extract the final shape of buildings. The second method involves four main steps: homogeneous superpixel image segmentation through an altered Simple Linear Iterative Clustering (SLIC), extensive feature extraction via a variational auto-encoder (VAE) adjust on the superpixels for training and testing data collection, classification of four classes (buildings, roads, trees, and shadows) utilizing extracted feature data as feedback to a Convolutional Neural Network (CNN), and extraction of building forms by morphological processes and regional growth. Our innovative methods have excellent accuracy rates for identifying building units. Accuracy assessment over different study areas shows the advantage of our novel approach, making it a robust, realistic, and accurate tool.