Fusion of Remote Sensing Imagery: Modeling and Application

Current satellite remote sensing systems compromise between spatial resolution and spectral and/or temporal resolution, which significantly limits their utility in various applications. Image fusion, including spatial and spectral fusion (SSF) and spatial and temporal fusion (STF), provides powerful tools for addressing this limitation. A multitude of methods have been proposed in the last three decades for SSF and STF, involving technologies in the fields of signal processing, statistics, computer vision and machine learning. The variability of the methods makes it not straightforward to understand the fusion process and to develop new fusion methods.;This thesis considers these two branches of image fusion as a common inverse problem, which is attempting to predict the unobserved high spatial resolution (HR) images from its low spatial resolution (LR) version with the help of another HR images acquired under different conditions. This thesis then formulates a general framework of modeling and inversion for the prediction. The modeling is to link the observed and prediction images or to describe prior information of the prediction images. The inversion is to estimate the prediction images. Such a framework gives the possibility to physically and intuitively interpret the classical fusion methods and to design novel models.;Based on such framework, two novel methods are designed for fusion. The first one utilized the spectral/temporal pattern information of the satellite images instead of utilizing detail injection methods. The spectral/temporal pattern information is derived from the observed LR image pair. The unknown HR images are then estimated from the learned spectral/temporal patterns and the observed HR images. The experiments on pan-sharpening of the QuickBird and WorldView-2 images and STF of the MODIS and TM images showed improvement of the proposed method over the state-of-the-art fusion methods. The improvement is particularly found where the observed and predicted HR images show much difference in terms of spatial details.;The second novel method is specifically designed for vegetated land surfaces, which utilizes the prior information of the seasonal pattern of variation by using a double logistic function. And the HR images are estimated via a multi-objective optimization technique. The fusion results were shown by reconstructing the seasonal variation of vegetation index at the Landsat resolution by fusing frequent MODIS data. The reconstructed vegetation index is validated to benefit for leaf area index retrieval at the both high temporal and spatial resolutions.