Anomaly detection and compensation for hyperspectral imagery

Abstract

Hyperspectral sensors observe hundreds or thousands of narrow contiguous spectral bands. The use of hyperspectral imagery for remote sensing applications is new and promising, yet the characterization and analysis of such data by exploiting both spectral and spatial information have not been extensively investigated thus far. A generic methodology is presented for detecting and compensating anomalies from hyperspectral imagery, taking advantage of all information available -- spectral and spatial correlation and any a priori knowledge about the anomalies. An anomaly is generally defined as an undesired spatial and spectral feature statistically different from its surrounding background. Principal component analysis (PCA) and the Iterative Order and Noise (ION) estimation algorithm provide valuable tools to characterize signals and reduce noise. Various methodologies are also addressed to cope with nonlinearities in the system without much computational burden. An anomaly compensation technique is applied to specific problems that exhibit different stochastic models for an anomaly and its performance is evaluated.

(cont.) Hyperspectral anomalies dealt with in this thesis are (1) cloud impact in hyperspectral radiance fields, (2) noisy channels and (3) scan-line miscalibration. Estimation of the cloud impact using the proposed algorithm is especially successful and comparable to an alternative physics-based algorithm. Noisy channels and miscalibrated scan-lines are also fairly well compensated or removed using the proposed algorithm.