The image data collected by the JERS-1 OPS, SPOT and AVHRR systems are affected by a variety of periodic (coherent) noise problems, which can be severe in many cases. This paper examines some of these defects and also describes and evaluates a series of methodologies to recover partially or totally the information contained in these data by means of: (i) filtering in the spatial domain (convolution filters), (ii) Principal Components Analysis (PCA) and (iii) filtering in the frequency domain (using Fast Fourier Transforms). Filtering in the spatial domain using relatively small convolution filters can be successfully applied to tackle elementary periodic noise problems. However, variations of theoretically efficient kernels experimented in this study were only able to minimise the effect of such complex noise structures at the expense of significant modification or complete loss of important raw image data. Principal Component Analysis transforms the data so that the noise component is cast into one or more of the high-order principal components. Our experiments show that this technique is able to confine the noise in the higher order PCs, but a significant amount of residual noise was present in the low-order components. Filtering in the frequency domain proved to be a suitable technique to achieve image restoration whilst preserving most of the raw scene information intact. Conventional Fourier operators like the Notch Filter and the Low-pass Elliptically Symmetric Filter (with a Gaussian-shaped intensity profile) can tackle the problems with some noisy images quite reasonably. However, such filters are unable to recover images on which the amplitude of the noise varies heterogeneously producing a series of impulses from low to high frequencies. Both the Zonal Notch Filter (based on the difference of the Fourier spectra of two channels) and the Synergistic Filter (which relies on the convolution theorem) are designed to contour this problem. These are interactive restoration methods that can successfully eliminate or minimise the effects of multiple two-dimensional periodic structures superimposed on both multi-channel and single channel data, respectively.
