Gargiulo, Massimiliano (2021) A Deep Learning Framework in Selected Remote Sensing Applications. [Tesi di dottorato]

[img]
Preview
Text
Gargiulo_Massimiliano_33.pdf

Download (46MB) | Preview
[error in script] [error in script]
Item Type: Tesi di dottorato
Resource language: English
Title: A Deep Learning Framework in Selected Remote Sensing Applications
Creators:
CreatorsEmail
Gargiulo, Massimilianomassimiliano.gargiulo@unina.it
Date: 15 June 2021
Number of Pages: 170
Institution: Università degli Studi di Napoli Federico II
Department: Ingegneria Elettrica e delle Tecnologie dell'Informazione
Dottorato: Information technology and electrical engineering
Ciclo di dottorato: 33
Coordinatore del Corso di dottorato:
nomeemail
Riccio, Danieledaniele.riccio@unina.it
Tutor:
nomeemail
Ruello, GiuseppeUNSPECIFIED
Date: 15 June 2021
Number of Pages: 170
Keywords: remote sensing; sentinel 1; sentinel 2; super resolution; segmentation; data fusion; sensors integration
Settori scientifico-disciplinari del MIUR: Area 09 - Ingegneria industriale e dell'informazione > ING-INF/02 - Campi elettromagnetici
Date Deposited: 04 Jul 2021 23:31
Last Modified: 07 Jun 2023 11:09
URI: http://www.fedoa.unina.it/id/eprint/13771

Collection description

The main research topic is designing and implementing a deep learning framework applied to remote sensing. Remote sensing techniques and applications play a crucial role in observing the Earth evolution, especially nowadays, where the effects of climate change on our life is more and more evident. A considerable amount of data are daily acquired all over the Earth. Effective exploitation of this information requires the robustness, velocity and accuracy of deep learning. This emerging need inspired the choice of this topic. The conducted studies mainly focus on two European Space Agency (ESA) missions: Sentinel 1 and Sentinel 2. Images provided by the ESA Sentinel-2 mission are rapidly becoming the main source of information for the entire remote sensing community, thanks to their unprecedented combination of spatial, spectral and temporal resolution, as well as their open access policy. The increasing interest gained by these satellites in the research laboratory and applicative scenarios pushed us to utilize them in the considered framework. The combined use of Sentinel 1 and Sentinel 2 is crucial and very prominent in different contexts and different kinds of monitoring when the growing (or changing) dynamics are very rapid. Starting from this general framework, two specific research activities were identified and investigated, leading to the results presented in this dissertation. Both these studies can be placed in the context of data fusion. The first activity deals with a super-resolution framework to improve Sentinel 2 bands supplied at 20 meters up to 10 meters. Increasing the spatial resolution of these bands is of great interest in many remote sensing applications, particularly in monitoring vegetation, rivers, forests, and so on. The second topic of the deep learning framework has been applied to the multispectral Normalized Difference Vegetation Index (NDVI) extraction, and the semantic segmentation obtained fusing Sentinel 1 and S2 data. The S1 SAR data is of great importance for the quantity of information extracted in the context of monitoring wetlands, rivers and forests, and many other contexts. In both cases, the problem was addressed with deep learning techniques, and in both cases, very lean architectures were used, demonstrating that even without the availability of computing power, it is possible to obtain high-level results. The core of this framework is a Convolutional Neural Network (CNN). {CNNs have been successfully applied to many image processing problems, like super-resolution, pansharpening, classification, and others, because of several advantages such as (i) the capability to approximate complex non-linear functions, (ii) the ease of training that allows to avoid time-consuming handcraft filter design, (iii) the parallel computational architecture. Even if a large amount of "labelled" data is required for training, the CNN performances pushed me to this architectural choice.} In our S1 and S2 integration task, we have faced and overcome the problem of manually labelled data with an approach based on integrating these two different sensors. Therefore, apart from the investigation in Sentinel-1 and Sentinel-2 integration, the main contribution in both cases of these works is, in particular, the possibility of designing a CNN-based solution that can be distinguished by its lightness from a computational point of view and consequent substantial saving of time compared to more complex deep learning state-of-the-art solutions.

Downloads

Downloads per month over past year

Actions (login required)

View Item View Item