Investigating urban processes requires detailed built-up surface material composition information, which is possible through constant monitoring by recently launched spaceborne hyperspectral sensors. However, they are plagued by medium spatial resolutions and mixed pixels. Super-resolution (SR) and spectral unmixing can address the former and latter, respectively. Machine learning, due to its predictive capability, has become indispensable for studying these datasets. Hence, remote sensing researchers, need to understand machine learning algorithms thoroughly. Herein, works on machine learning application for urban hyperspectral sensing have been thoroughly analyzed. A functional SR classification scheme has also been introduced. Super-resolved product quality metric evaluation, open-source urban spectral libraries, benchmark urban scenes meant for training and testing new SR, or unmixing algorithms have been briefly reviewed. Finally, difficulties with hyperspectral image processing based on machine learning have been raised, along with future research directions.
Top1. Introduction
The global urban area is expected to increase from 0.7-0.9 million km2 in 2010 to 1.2-3.1 million km2 in 2050 (Angel et al., 2011), and the number of cities with more than 10 million inhabitants is projected to rise from 33 in 2018 to 43 in 2030 (United Nations, 2018). Therefore, a continuous conversion of land cover (LC) happens due to modifications in the biophysical properties of every constituent of the heterogeneous urban area, impacting the global economy and society as well as biogeochemical cycles and climatic patterns (Bren d’Amour et al., 2017; Oke et al., 2017; Seto et al., 2012; van Vliet et al., 2017). Evaluating these urban expansion effects necessitates quantifying urban area components and mapping them constantly in a standardised manner (Ridd, 1995).
'Imperviousness' or 'impervious area' defines the urbanisation extent since the early 1990s, serving as the environmental quality indicator (Arnold & Gibbons, 1996) and uniting stakeholders at all watershed scales like planners, engineers, landscape architects, scientists, social scientists, local officials, and the civil society (Schueler, 1994). Increased impervious cover has led to an increased urban runoff (Weng, 2001) and declined groundwater recharge and baseflow (Brun & Band, 2000). Increased entrapment of radiation within the urban structures also leads to significant warming of the urban environment (Grimmond, 2007; Mishra & Garg, 2023a; Ren et al., 2019; Verma & Garg, 2021). Conventional ways of delineating impervious surfaces are time-consuming, expensive and not viable for mapping large areas. Remote sensing systems, whether airborne or spaceborne, overcome these difficulties to facilitate the monitoring of urban areas (Roberts & Herold, 2004; Roy et al., 2017) and have different spatial, spectral, radiometric and temporal resolutions (Jensen & Cowen, 1999).
Optical or passive remote sensors only provide useful information from the Earth's surface in cloud-free conditions. In contrast, active or radar remote sensors record weather-independent data day and night (Tripathi et al., 2021) and can be used to study object texture and structure (Soergel, 2010). Depending on the spectral detail observed, optical sensors can be multispectral or hyperspectral. Hyperspectral systems hold a distinct edge over multispectral systems by enabling a thorough differentiation regarding the presence of surface materials in built-up areas due to over 200 very narrow, contiguous spectral bands (Gamba & Dell’Acqua, 2007; Masudul Islam et al., 2022). However, spaceborne hyperspectral systems are few (Duca & Frate, 2008; Müller et al., 2016; Pearlman et al., 2003), with medium spatial resolutions of 20 m to 30 m, which restricts capturing thematically explicit spectral information of urban areas. Such an image contains predominantly mixed pixels and has fewer spectrally pure pixels, indicative of only a single surface material.