Date of Award
Doctor of Philosophy (PhD)
Deep learning (Machine learning), Remote sensing--Data processing, Drone aircraft in remote sensing
Land Remote Sensing data analysis presents a distinctive Artificial Intelligence (AI) research paradigm bolstered through its rich collection of data, which are high-dimensional, generally in the form of tens/hundreds of spectral bands, and accommodates significant spatial and spectral/temporal information about the underlying terrain. Traditional remote sensing data analysis methodologies in literature are often observed to be biased towards providing more importance to spectral information, which evidently hurts the efficacy of such approaches. As a result, a considerable amount of effort in terms of research has been invested in effectively building data analysis methodologies which are spatial-, spectral-, and contextual-information inclusive, which improves the overall performance. However, this process of extracting necessary additional information and knowledge discovery requires increased computation and hardware requirements (memory) which are not always readily accessible or available in time-sensitive data analysis applications. Also, such approaches lack the intrinsic ability to prioritize information that is cardinal to boost the performance of such techniques, which impedes efficient automation, performance and the ability to build frameworks for real-time land remote sensing data analysis. Hence, to address the aforementioned challenges, this dissertation presents a series of deep-learning-based attention methodologies that are spatial-spectral information inclusive and inherently have the capability to prioritize information that is pivotal for a robust and light-weighted land remote sensing data analysis and classification framework pertinent to Unmanned Aerial Vehicles (UAVs)-based autonomous systems that have limited resource platforms. This dissertation also presents a new framework for detecting landmarks or key points for UAVs in real-time using image data from an on-board camera sensor. The proposed framework consists of two major phases: the first phase involves training an efficient landmark detection algorithm for detecting the landmarks of interest using both real and synthetic data samples, and the second phase involves detecting the known key points or anchor points from the captured images, which correspond to the key areas of interest from the perspective of UAVs. The framework has the potential to increase the safety and productivity of UAV operations by enabling real-time detection of critical landmarks during disaster relief, unmanned military operations, and other applications. In conclusion, this dissertation provides an efficient means for automation, extraction and prioritization of spatial- and spectral-information present in land remote sensing data compared to the conventional spatial- or spectral information-only based analysis methodologies. Experimental results reveal that the proposed land remote sensing data analysis models outperform the conventional approaches from literature. The research outcomes from this work will have significant implications in advancing the state-of-the-art deep learning attention methodologies for a UAV autonomous system-centric data analysis applications.
Praveen, Bishwas, "A comprehensive study of deep learning frameworks for UAVs-centric land remote sensing data analysis applications" (2023). Dissertations. 330.