Towards a coherent theory of statistical land use change models

Changes in land use and land cover are one of the main drivers of global change. As such, they are frequently modelled in environmental sciences with the aim, on the one hand, of assessing their impacts on ecosystems or their contribution to climate change and, on the other hand, of informing public decision-makers in territorial planning on the potential impacts of land use policies. These models are mobilized on territories ranging from landscape entities to continental dimensions, at varying degrees of spatial resolution, and over a time horizon of a few decades at most.

Several modeling strategies are available for this purpose. One of the most common is based on a statistical approach to the problem. Past changes are described in a probabilistic way, by correlating the observed changes with different geographical and/or socio-economic characteristics of the territories considered, in a spatially explicit way, and the changes are allocated in the future in a statistical way on the basis of these correlations according to global scenarios of evolution of the territory. In practice, several softwares implement such strategies.

It is well known in the specialized literature that these softwares give different and sometimes incoherent results on the same problem using the same set of data, whereas the reliability of the results obtained is critical both from a scientific point of view and for public decision support. However, for the moment, the origin of these differences in behaviour is not identified, for two main reasons:

1. the modeling choices implemented in these softwares are different even at the level of probability theory, and are not always explained in the literature or the documentation of these softwares;
2. in all likelihood, the designers of these softwares do not have sufficient competence in probability theory to define conceptually satisfactory modeling strategies while maintaining a high level of simplicity of implementation and flexibility of use.

The objective of the thesis is multiple:

1. develop a generic framework in probability theory that defines, in principle, the most consistent and satisfactory formulation of the problem ;
2. to define different levels of approximation allowing more or less complex implementations of this framework, in a controlled and testable way;
3. to identify theoretical and methodological errors present in existing software, by illustrating them on real case studies;
4. propose alternatives and/or corrections, by evaluating their relevance and effectiveness on various examples.

In practice, results have already been obtained on these four points. The doctoral student's work will consist of systematizing them and developing suitable digital implementation strategies, as well as demonstrations of principle in concrete cases. The aim is not to develop a new land use change software, but to highlight theoretically and numerically where the main existing software introduces errors either conceptually or in implementation.

This thesis work thus directly targets one of the main theoretical and practical obstacles to land use change modeling. The potential interest of this work for the communities concerned is considerable (mainly ecology and quantitative geography).