An Adaptive Kernel Layer Deep Neural Network for Remote Sensing Imageries

Abstract

Convolutional neural networks in recent years have become wildly utilized for the various applications that deal with images, going through different architectures and development to improve their capabilities for features extraction in order to yield better results in deep learning methods. Since images are vital resources that are used to represent various sceneries; multi-temporal images are an important tool used to monitor the changes that happen to those sceneries, which is required for multiple fields such as urban planning and disaster assessment, and with the rapid development of technology, very high-resolution (VHR) images from various sources are now more available, requiring analysis of the convolutional networks for larger images, and further developments to enhance their performance, so that they can process the large amount of data and identify any changes more efficiently. This thesis proposes the utilization of adaptive kernels layer (AKL) in order to extract features from large images, where the layer has been designed to maximize spectral information while retaining the spatial resolution of the information, discussing the benefits gained from utilizing it, and comparing it to other popular feature selection methods in image change detection applications. This thesis also examines and provides models used to generate the change image by utilizing machine learning (ML) models, mainly convolutional neural networks (CNN) and long shortterm memory (LSTM).