The development of superspectral approaches for the improvement of land cover classification

This article develops a critical review of the hyperspectral splitting of the solar reflected radiation acquired by hyperspectral imaging sensors. The bandwidth used in the range from 2.0 micrometer to 2.5 micrometer by many hyperspectral sensors sometimes is too narrow for land cover classification. In fact, hyperspectral imagers often suffer from low signal-to-noise in the shortwave infrared region of the e.m. spectrum, resulting in noisy image collection. This paper presents a new methodological approach to the splitting of the solar reflected radiation, called “Superspectral approach”. It is based on the principle of increasing the channel bandwidth by increasing the number of wavelengths, to build synthetic spectral bands with higher SNR. The methodology has been applied to the MIVIS (Multispectral Infrared and Visible Imaging Spectrometer) sensor, an airborne hyperspectral scanner used for environmental remote sensing applications in Italy. Interesting results have been achieved in crop classification, processing the Cordenons survey carried out in August 2001 in the north-eastern part of Italy. The Spectral Angle Mapper (SAM) algorithm was used for classification because insensitive to shadows. For accuracy assessment, the overall accuracy (OA) and kappa coefficient (k) were calculated and used in the comparison. Using the Superspectral approach, an increment in the overall accuracy of about 42% and an increment in the kappa coefficient of about 51% were obtained in comparison to the classification accuracy of unprocessed original MIVIS data (OA=41.21, k=0.35). A second case study is presented using the NASA’s experimental hyperspectral imager HYPERION. Data acquired over the lake of Garda (Italy) in October 2002 was processed with the Superspectral approach. Comparing the simulated HYPERION superspectral bands with the original data, SNR improvements are achieved in the SWIR region (from 0.7 to 54.2 for 2.012 micrometer wavelength and from 0.7 to 64.5 for 2.365 micrometer wavelength). The methodology proposed is sensor independent and can be applied to any of the hyperspectral sensors currently available.