A Markovian Approach to Unsupervised Change Detection with Multiresolution and Multimodality SAR Data

Abstract

In the framework of synthetic aperture radar (SAR) systems, current satellite missions make it possible to acquire images at very high and multiple spatial resolutions with short revisit times. This scenario conveys a remarkable potential in applications to, for instance, environmental monitoring and natural disaster recovery. In this context, data fusion and change detection methodologies play major roles. This paper proposes an unsupervised change detection algorithm for the challenging case of multimodal SAR data collected by sensors operating at multiple spatial resolutions. The method is based on Markovian probabilistic graphical models, graph cuts, linear mixtures, generalized Gaussian distributions, Gram–Charlier approximations, maximum likelihood and minimum mean squared error estimation. It benefits from the SAR images acquired at multiple spatial resolutions and with possibly different modalities on the considered acquisition times to generate an output change map at the finest observed resolution. This is accomplished by modeling the statistics of the data at the various spatial scales through appropriate generalized Gaussian distributions and by iteratively estimating a set of virtual images that are defined on the pixel grid at the finest resolution and would be collected if all the sensors could work at that resolution. A Markov random field framework is adopted to address the detection problem by defining an appropriate multimodal energy function that is minimized using graph cuts.

Monitoring the Recovery after 2016 Hurricane Matthew in Haiti via Markovian Multitemporal Region-Based Modeling

“…Differently from MRFs, which are designed to model the prior distribution of the desired output map, CRFs [55] have been introduced to formalize the Markovianity property with regard to the posterior distribution directly, often further enhancing modeling flexibility. The remarkable results obtained by this Markovian approach to multisource fusion in the remote sensing field can be seen in various examples involving multisensor [56], multitemporal [22], multiresolution [57], or multichannel imagery [58], or combinations of the above [59,60].…”

“…The Bayesian MAP rule is equivalent to the minimization of the energy U(L|X ) with respect to L, given the input image X . Multiple information sources can also be fused in this Markovian framework by defining appropriate energy functions as linear combinations of contributions associated with the individual sources [43,59,60].…”

Deep Image Translation With an Affinity-Based Change Prior for Unsupervised Multimodal Change Detection

“…, n). The former is a rather classical conditional independence assumption, which is frequently accepted in change detection studies [43], [26], [72], [73]. The latter is very common, and the reformulation in the case of discrete or mixed variables is straightforward.…”

Multimodal Fusion of Mobility Demand Data and Remote Sensing Imagery for Urban Land-Use and Land-Cover Mapping

“…MRFs are a powerful family of stochastic models for image data in Bayesian image analysis [43,44]. They have been successfully applied for a long time in remote sensing for image classification and segmentation [24,[45][46][47][48][49][50], object-based and regionbased image analysis [10,51,52], cloud detection [53,54], and change detection [55][56][57]. They represent a multi-dimensional generalisation of Markov chains and, in the case of 2D images, are defined in terms of a Markovianity property on the two-dimensional pixel lattice with respect to a given neighbourhood system [11,43,44].…”