Abstract
Abstract. The automated analysis of large areas with respect to land-cover and land-use is nowadays typically performed based on the use of hyperspectral or multispectral data acquired from airborne or spaceborne platforms. While hyperspectral data offer a more detailed description of the spectral properties of the Earth’s surface and thus a great potential for a variety of applications, multispectral data are less expensive and available in shorter time intervals which allows for time series analyses. Particularly with the recent availability of multispectral Sentinel-2 data, it seems desirable to have a comparative assessment of the potential of both types of data for land-cover and land-use classification. In this paper, we focus on such a comparison and therefore involve both types of data. On the one hand, we focus on the potential of hyperspectral data and the commonly applied techniques for data-driven dimensionality reduction or feature selection based on these hyperspectral data. On the other hand, we aim to reason about the potential of Sentinel-2 data and therefore transform the acquired hyperspectral data to Sentinel-2-like data. For performance evaluation, we provide classification results achieved with the different types of data for two standard benchmark datasets representing an urban area and an agricultural area, respectively.
Highlights
Hyperspectral imagery acquired from airborne or spaceborne sensor platforms is commonly used for scene interpretation in terms of land-cover and land-use classification (Plaza et al, 2009; Camps-Valls et al, 2014)
The classification results achieved for the Pavia Centre dataset are generally of rather high quality (OA = 95.46 . . . 96.72 %), while the classification results achieved for the Salinas dataset are worse (OA = 84.77 . . . 86.69 %)
These figures reveal that class C04 (“Self-Blocking Bricks”) of the Pavia Centre dataset and classes C08 (“Grapes untrained”) and C15 (“Vinyard untrained”) of the Salinas dataset seem to be problematic for the given classification task
Summary
Hyperspectral imagery acquired from airborne or spaceborne sensor platforms is commonly used for scene interpretation in terms of land-cover and land-use classification (Plaza et al, 2009; Camps-Valls et al, 2014). In this context, the main objective is a classification on a per-pixel basis, whereby the classes of interest are typically defined with respect to a specific application or usecase. A rather intuitive strategy is to consider the reflectance values of each pixel across all spectral bands These reflectance values are concatenated to a highdimensional feature vector which, in turn, is provided as input to a classifier delivering a hypothesis about the respective class label. Thereby, a supervised classification based on a standard classifier such as a Support Vector Machine (Melgani and Bruzzone, 2004; Chi et al, 2008) or a Random Forest (Ham et al, 2005; Joelsson et al, 2005) is often applied, where the involved classifier needs to be trained before on representative training data
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
