CHARACTERIZATION OF IMPERVIOUS SURFACES USING HYPERSPECTRAL DATA AND MACHINE LEARNING TECHNIQUES

Abstract

The United Nations Sustainable Development Goals (SDGs) aim to eradicate poverty, increase economic growth, tackle climate change, and conserve natural resources. SDG 11, among the 17 SDGs, focuses on ensuring city sustainability and building resilient communities, which is possible through constantly understanding urban processes. These processes happening at a microscale level profoundly affect the socio-economic trends, weather patterns and biogeochemical cycles. Understanding them requires continually observing the urbanization extent or the ‘imperviousness’ through satellite data and repeatedly characterising impervious surfaces into distinct rooftop, pavement and natural cover information classes, achievable only through high spatial, spectral and temporal resolutions. However, there is no satellite sensor with such capabilities. New spaceborne hyperspectral sensors launched since 2018 offer high spectral and temporal resolution and have ensured hyperspectral data continuity since decommissioning the Hyperion in 2017. Although having wider ground area coverage, their 30 m spatial resolution is insufficient to distinguish the spatially small and spectrally similar urban objects, i.e., built-up structures and open and green spaces, that constitute the impervious surfaces. As a result, they may occur as mixed pixels in such images. Moreover, the sensor optics design makes swapping detailed spatial and spectral information difficult at the hardware stage. Hence, a software-based intervention, resolution enhancement approach and spectral unmixing have been proposed to address these challenges associated with the hyperspectral imaging of impervious surfaces, capturing urban objects distinctly and characterising them precisely. Single-frame super-resolution (SR) takes care of the other resolution enhancement approaches’ shortcomings by spatially improving the resolution of a single low-resolution (LR) hyperspectral input while preserving its rich spectral information. Bearing the user in mind, this study improves the spatial resolution of two hyperspectral spaceborne data sets, Hyperion and PRecursore IperSpettrale della Missione Applicativa (PRISMA), which were obtained over a mixed land use area in downtown Ahmedabad, Gujarat, India, in the years 2002 and 2020, respectively. We generated super-resolved products with spatial resolutions of 15 m, 10 m, and 7.5 m using four different single-frame SR techniques. Open-source panchromatic and multispectral datasets were recommended to study spatial visual analysis, along with a composite comparison evaluation and validation process. The results demonstrated that the original spectral information was well preserved at larger geographic scales. Image learning algorithms, particularly, decreased picture quality as the scale factor (SF) grew. At higher SFs, reconstruction and hybrid-based SR algorithms outperformed learning-based methods. Among the chosen SR techniques for SF = 2, 3, and 4, iterative back projection (IBP) and sparse representations (SIS) created super-resolved products performed the best and worst in the Hyperion example, respectively. IBP was the best-performing SR method in the PRISMA instance for all SFs. In contrast, the worst-performing algorithms were SIS and neighbourhood embedding - locally linear embedding (NELLE) at SF = 2 and 3 and SF = 4, respectively. A robust multi-temporal ground truth in the form of positional information from field surveys and crowd-sourced data has been created for the study area, and pure spectral references of various urban materials and natural cover classes found in the study area have been collected and contained in a spectral library. Resampling the developed spectral library to the Hyperion and PRISMA wavelengths and using the principles of absorption spectroscopy, significant regions and unique spectral bands have been identified for characterizing the urban material and natural cover classes in study area scenes acquired by Hyperion or PRISMA or their derived imageries or any sensor having similar specifications. Most significant wavelengths are present in near-infrared (NIR) regions, followed by visible and shortwave infrared (SWIR). Out of these wavelengths, more than a single wavelength in the same wavelength region is significant for characterising more than one information class. Further, overlay analysis of the PRISMA data over the Hyperion data identifies the change areas experiencing a transition from one land cover (LC) class to another. These are the extreme western portion of the study area and the central region surrounding the Sabarmati River. Semi-automated information extraction pipelines for precise urban information retrieval from the original and the best and worst super-resolved products have been devised by identifying the best performing algorithm at each step of the spectral unmixing process, i.e., dimensionality reduction, end member extraction and abundance estimation, using suitable metrics. Non-linear dimensionality reduction algorithms result in better reduced encoding at higher spatial scales. At the same time, eigenvalue decomposition (EVD)-based principal component analysis (PCA) stands out for the original datasets and super-resolved products at lower spatial scales. Pixel purity index (PPI) tends to outperform the automatic target generation process (ATGP) in the case of PRISMA and its super-resolved products. At the same time, ATGP is a better end member extraction approach in the case of super-resolved products produced from Hyperion at higher SFs, except in the case of the original Hyperion and the worst super-resolved product at SF = 2. The generalized bilinear model (GBM) reports lower reconstruction errors than the linear, fully constrained least squares (FCLS) model, proving a better spectral unmixing approach in retrieving urban material and natural cover classes’ abundance fractions. However, it is more computationally expensive than FCLS. Major information classes’ abundance fractions clearly show the accurate characterization of these information classes in the homogeneous patches of the original datasets and super-resolved products. Even the abundance fractions precisely reflect the transition from one LC to another. In the heterogeneous patches, a good degree of mixing is evident, resulting in the misidentification of the respective urban material or natural cover class, thereby highlighting the spatial resolution limitation of the original dataset. Nevertheless, the proof-of-concept exhibited on airborne hyperspectral datasets has been successfully extended to spaceborne hyperspectral datasets. The results obtained in this study show that the performance can be enhanced with the use of higher spatial resolution input datasets, say 20 m, and the development of proper rules during the utilization of the abundance fractions for retrieving level-3 or 4 urban region maps, which can serve as a key input in studies involving stormwater runoff modelling; modelling the urban greenspace, heat and pollution island nexus; and adopting climate-resilient architectural practices.