Hydrocarbon quantification using neural networks and deep learning based hyperspectral unmixing

Ahmed, Asmau Mukhtar (2019) Hydrocarbon quantification using neural networks and deep learning based hyperspectral unmixing. (PhD thesis), Kingston University, .

Full text available as:
[img]
Preview
Text
Ahmed-A-45040.pdf - Submitted Version
Available under License Creative Commons Attribution.

Download (16MB) | Preview

Abstract

Hydrocarbon (HC) spills are a global issue, which can seriously impact human life and the environment, therefore early identification and remedial measures taken at an early stage are important. Thus, current research efforts aim at remotely quantifying incipient quantities of HC mixed with soils. The increased spectral and spatial resolution of hyperspectral sensors has opened ground-breaking perspectives in many industries including remote inspection of large areas and the environment. The use of subpixel detection algorithms, and in particular the use of the mixture models, has been identified as a future advance that needs to be incorporated in remote sensing. However, there are some challenging tasks since the spectral signatures of the targets of interest may not be immediately available. Moreover, real time processing and analysis is required to support fast decision-making. Progressing in this direction, this thesis pioneers and researches novel methodologies for HC quantification capable of exceeding the limitations of existing systems in terms of reduced cost and processing time with improved accuracy. Therefore the goal of this research is to develop, implement and test different methods for improving HC detection and quantification using spectral unmixing and machine learning. An efficient hybrid switch method employing neural networks and hyperspectral is proposed and investigated. This robust method switches between state of the art hyperspectral unmixing linear and nonlinear models, respectively. This procedure is well suited for the quantification of small quantities of substances within a pixel with high accuracy as the most appropriate model is employed. Central to the proposed approach is a novel method for extracting parameters to characterise the non-linearity of the data. These parameters are fed into a feedforward neural network which decides in a pixel by pixel fashion which model is more suitable. The quantification process is fully automated by applying further classification techniques to the acquired hyperspectral images. A deep learning neural network model is designed for the quantification of HC quantities mixed with soils. A three-term backpropagation algorithm with dropout is proposed to avoid overfitting and reduce the computational complexity of the model. The above methods have been evaluated using classical repository datasets from the literature and a laboratory controlled dataset. For that, an experimental procedure has been designed to produce a labelled dataset. The data was obtained by mixing and homogenizing different soil types with HC substances, respectively and measuring the reflectance with a hyperspectral sensor. Findings from the research study reveal that the two proposed models have high performance, they are suitable for the detection and quantification of HC mixed with soils, and surpass existing methods. Improvements in sensitivity, accuracy, computational time are achieved. Thus, the proposed approaches can be used to detect HC spills at an early stage in order to mitigate significant pollution from the spill areas.

Item Type: Thesis (PhD)
Physical Location: Online only.
Uncontrolled Keywords: hydrocarbon; neural networks; deep learning; hyperspectral; unmixing
Research Area: Chemical engineering
Computer science and informatics
Earth systems and environmental sciences
Faculty, School or Research Centre: Faculty of Science, Engineering and Computing
Faculty of Science, Engineering and Computing > School of Engineering
Depositing User: Philip Keates
Date Deposited: 20 Feb 2020 16:08
Last Modified: 29 Jun 2020 15:41
URI: http://eprints.kingston.ac.uk/id/eprint/45040

Actions (Repository Editors)

Item Control Page Item Control Page