Segmented Curve-Fitting Method for Continuum Removal in CRISM MTRDR data

Abstract. A spectrum in a multiband remotely sensed image is generally a mixture of spectra of different materials present in the scene which can be distinguished by distinct absorption signatures. A mixed spectrum possesses a smooth baseline shape, known as a continuum, that masks the individual spectral features. Continuum can also appear due to instrument artifacts and topographic illumination effects. Eliminating the continuum from a spectrum being analyzed and correctly identifying its unique absorption characteristics are crucial for material identification, traditionally achieved by the apparent continuum removal methods like using an Upper Convex Hull (UCH). Nevertheless, most of these methods struggle when baseline curvature exceeds certain limits, often combining distinct absorptions. In this paper, a new apparent continuum removal technique called Segmented Curve-Fitting (SCF) is proposed, which requires no prior information about the spectrum but excels in accurately extracting distinct absorptions, even in the presence of significant curvature. The performance of SCF is compared with UCH and a few other apparent continuum removal methods previously used in literature, using a collection of simulated data of varying complexity as well as a real CRISM TRDR hyperspectral dataset. The identification score is improved by around 8% for the similarity matching method Weighted Sum of Spectrum Correlation and by around 1.5% for a Convolutional Neural Network. Furthermore, an SCF-based mineral identification framework demonstrates its effectiveness in identifying the dominant minerals on CRISM MTRDR hyperspectral data collected from different locations on the Martian surface.