Deconvolving and improving the spatial resolution of satellite data using the Maximum Entropy Method
thesisposted on 2023-05-26, 00:52 authored by Jackett, CJ
Remote sensing satellite imagery provides information about the surface of the Earth at a range of spectral bands and spatial resolutions. This information is a valuable resource for the management of terrestrial and marine environments. During the capturing process, incoming light is reflected or refracted by the instrument optics which causes a small amount of blurring. This effect is described by a mathematical operation called convolution in which the satellite input radiance field is convolved with the instrument Point Spread Function (PSF). This form of instrumental distortion has the largest impact on high-contrast scenes where bright land or clouds are adjacent to dark surfaces such as water. This thesis investigates three mechanisms for improving the quality of recorded satellite data. An efficient convolution method was developed to minimise boundary effects, a deconvolution algorithm was used to remove instrumental distortion, and a resolution enhancement algorithm was developed to improve the spatial resolution of input images. The latter two of these problems are underdetermined and require appropriately selected constraints in order to find unique and stable solutions. An entropy-based method was chosen as the constraint element due to its heavy grounding in statistical mechanics and information theory. MODerate resolution Imaging Spectroradiometer (MODIS) Aqua images were used to quantify the improvement of these algorithms, with a focus on coastal marine and open-ocean environments. Deconvolution is an algorithm-based process designed to reverse convolution effects with a known PSF. Multiscale Entropy deconvolution was applied to MODIS level 1A imagery to remove instrumental distortion from top-of-atmosphere radiance counts. Removing these effects at the beginning of the satellite image processing chain reduces the propagation and amplification of errors in subsequent processing stages. Wavelet transforms were implemented to decompose images into a range of resolution levels that represent different spatial frequencies. This allows both large-scale and small-scale features to be resolved simultaneously. Multiresolution Support images were used to accurately define and target important areas within the imagery. The combination of these techniques includes two-dimensional structural information in the Multiscale Entropy calculation which results in accurate deconvolution. Validation of the Multiscale Entropy deconvolution algorithm was undertaken using in-situ measurements from the Baltic Sea and a QuickBird image of a high-contrast Antarctic ice edge. A novel approach to the spatial resolution enhancement of MODIS imagery uses information about the optical PSF, along with the result of Multiscale Entropy deconvolution. With this information, a system of linear equations was constructed that models how high-resolution PSF convolution redistributes information over a finite area. A new method termed Multiresolution Entropy was developed to constrain the linear system and retrieve an optimal solution. The algorithm successfully improved the spatial resolution of input images and compared favourably to other interpolation-based methods. The key requirement of this technique is to obtain high-resolution PSF measurements at the same sampling frequency as the desired final output resolution. The techniques developed and presented in this thesis contain a range of important research contributions. The combination of Fast Fourier Transform convolution with a boundary renormalisation approach produces an efficient and accurate convolution method with minimal boundary effects. A multi-detector convolution process accurately simulates the MODIS Aqua instrumentation and allows for successful deconvolution. A detector saturated estimation technique for ocean colour bands ensures the correct quantity of instrumental distortion is removed during deconvolution. The formulation of a linear system consisting of highresolution PSF modelling and appropriate physical constraints defines the spatial resolution enhancement problem. The development of Multiresolution Entropy targets high-frequency content, constrains the linear system and results in a unique and stable resolution-enhanced solution. The techniques developed throughout this thesis provide considerable benefit to the quality of remote sensing imagery and can substantially improve the monitoring and management of coastal zones and other marine environments.
Rights statementCopyright the Author