Spectral-based regression model for destriping of abnormal pixel values in PRISMA hyperspectral image

Abstract

Hyperspectral imageries are often degraded by systematic sensor-based errors known as “striping noises”. This study implements a spectral-based regression algorithm from highly correlated consecutive bands, i.e. left band, right band or both, to model and reconstruct the abnormal pixel values, stripe noises, in various bands of PRISMA (PRecursore IperSpettrale della Missione Applicativa) imagery. The modeling performance was evaluated based on the statistical difference between the reconstructed images’ pixel values (reflectance) and their corresponding original pixel values. Results referred to the model’s high accuracy in R2, RMSE, rRMSE and skewness in most bands [Formula: see text]). Furthermore, the results indicated that the combination of both bands had higher accuracy and pixels’ homogeneity preservation compared to single-band modeling. Our findings suggested that the algorithm significantly depends on the spectral similarities between neighboring bands so that the higher spectral similarities lead to the higher model performance and vice versa. Subsequently, the minimum model performance was observed in band 143 due to lower spectral similarity, lower spectral correlation and higher wavelength differences with its adjacent right band. Finally, the study suggests that alongside other methods, our algorithm may be used as a reliable, straightforward and accurate alternative for destriping different Earth observation satellite imageries. Limitations of the proposed approach are also discussed.

Keywords

• Destriping
• hyperspectral image
• PRISMA
• spectral-based regression
• stripe noise