Informations et ressources scientifiques
sur le développement des zones arides et semi-arides

Accueil du site → Doctorat → Australie → Land Use Mapping Using Constrained Monte Carlo Methods

Australian National University (2016)

Land Use Mapping Using Constrained Monte Carlo Methods

Knapp, Simon Orlando

Titre : Land Use Mapping Using Constrained Monte Carlo Methods

Auteur : Knapp, Simon Orlando

Université de soutenance : Australian National University

Grade : Doctor of Philosophy (PhD) 2016

We present a flexible, automated, Bayesian method designed for broad scale land use mapping. The method is based on a Monte Carlo Markov Chain and integrates a number of sources of ancillary data. It produces a probability density over a finite set of land use classes that can be used directly in further analyses or to classify individual pixels. The method assumes a multi- nomial prior over the possible land use types, and uses agricultural statistics to form stochastic constraints over the total area allocated to each land use within a region. A supervised learner is then used to allocate pixels within the region, while respecting the constraints. We then extend this method in three ways. First, supplementary mapping is used to form further constraints over subsets of the original land use classes. Second, two spatial models are considered ; the first considers the use of partially labelled pixels, where the labels are based on the current state of the Markov Chain, and the second assumes a Markov Random Field. Third, the form of the prior is relaxed, and the method extended to allow the creation of a time series of maps using either cascade or compound classification techniques. The methods are benchmarked against the probabilistic classifier upon which they are built and simple Bayesian modifications to the raw classifier that incorporate the same data. The techniques are demonstrated and assessed using Australian data generated by a national Land Use (LU) mapping program and show promise in many of the test regions we consider.


Version intégrale (10,71 Mb)

Page publiée le 7 février 2018