ASIMUD - Remote Sensing Data Assimilation in Modelling of Urban Dynamics: Final report

Land-use models are increasingly used by planners and policy makers to assess the impacts of their decisions on spatial systems. This stresses the importance of a correct calibration of these models. Existing calibration methods of cellular automata (CA) based land-use models, such as historic calibration, however, do not sufficiently take into account uncertainties in the reference data and the model input parameters. As a result, uncertainties in the model simulations are mostly unknown.
The objective of this study is to implement a spatial metric-based calibration method in a probabilistic framework in order to quantify and reduce the uncertainty in simulating future land use. This is done through the use of data-assimilation techniques and more in particular a sequential importance resampling (SIR) particle filter.

This probabilistic framework applies Monte Carlo analysis and the particle filter data-assimilation algorithm, provided in the freely available PCRaster Python framework, to calibrate the land-use model parameters by taking into account (a) uncertainties in the reference data derived from remote sensing images and (b) uncertainties in the model input parameters. The approach is tested for both the urban scale (Dublin) and the regional scale (Flanders and Brussels-Capital Region).
Reference land-use maps were obtained from Landsat and SPOT images through (i) sub-pixel estimation of sealed surface cover for each urban pixel and (ii) application of a multiple layer perceptron (MLP) approach to infer urban land use from urban form, based on the spatial arrangement of sealed surface cover fractions at street block level. To model uncertainty in the fraction of sealed surface cover, use is made of a first-order autoregressive model, incorporating spatial correlation observed in the fractional errors. To deal with uncertainty associated with the land-use classification, a Bayesian approach is proposed, combining information on the confusion between land-use classes, obtained from the error matrix, with local uncertainty information produced by the MLP classifier.

Regarding the land-use change model, the MOLAND CA-based land-use model of Dublin and Flanders has been integrated into the Python-based framework for spatio-temporal modelling. In a first stage the MOLAND model was run as a stochastic model using Monte Carlo techniques in order to propagate uncertainties in the model input parameters through the model. The model output uncertainty is analysed in terms of spatial metrics. Next, the particle filter data-assimilation algorithm was used to calibrate the model, taking into account the uncertain input parameters. Hereto, the probability density functions of spatial metrics derived from medium resolution remote sensing images are assimilated in the model at each time step for which they are available. The particle filter algorithm only continues successful realizations of the land-use model, thus reducing its uncertainty by removing unsuccessful realisations from the original Monte Carlo ensemble. In this way, only those parameter sets that result in patterns similar to the patterns observed in the remote sensing data, are used in the simulation of land-use changes, resulting in a reduced uncertainty in the spatial metrics at each time step for which a remote sensing image was available.

This automatic procedure results in an improved parameter set of the land-use model. However, the best results are obtained for rather simplified calibration problems in which only a limited number of model parameters are assumed uncertain. Increasing the dimension of the problem, by increasing the number of parameters to be calibrated, leads to particle collapse, meaning that the posterior probability density is represented by only one unique particle.