Development of geospatial urban growth models supporting urban water strategic planning: the case study of Rethymnon, Crete

The alarming rate of urbanization poses immediate problems to water resources management, mainly, but not limited to water supply, flood risk management and mitigation measures, waste-water treatment and water quality control. Strategic planning of water systems should be fully aware of the prospects of future urban growth in order to maintain high reliability of services provided and satisfy customers in the long term. However, in contemporary studies, this rarely is the norm; commonly urban growth is used as a static input/scenario based on previous studies or general urban planning or the focus is shifted to the related population increase rather than its spatial allocation. Modern urban water strategic thinking needs to incorporate robust tools and methodologies in management practices, able to predict and quantify the outcome possibility of future urban growth. Urban growth is an inherently complex phenomenon. It can be considered as a system of physical expansion and function transitions. It changes trough time by the interaction with many inter-related, mostly unknown, drivers and stimulant factors. These interactions create an open, non-linear, dynamic and emergent system. Thus, it is fundamentally difficult to accurately model and predict urban growth, if not entirely impossible. The focus of modelling urban growth should be in creating plausible results and models with good explanatory power, able to simulate and explore the complex urban dynamics, rather than precisely pinpoint future urban locations. A dominant family of such models in literature is “Cellular Automata” (CA) based: The area of interest is divided in discrete cells that are self-organized. These models operate by applying in each simulation step simple rules defining local interactions among neighboring cells. The vast number of local interactions can result in a complex and dynamic global behaviour, especially when stochastic disturbances and rules are introduced to the model. Their simple in defining, yet complex in outcomes, nature makes CA-based models suitable for urban growth simulation and prediction. Many model sub-types and hybrid interdisciplinary approaches have been proposed and tested extensively in research efforts. A robust, parsimonious in data requirements and flexible CA model is developed in this work, as part of a general methodology framework aimed to assist urban water strategic planning. Briefly, the proposed two-state (urban and non-urban land uses) CA model is conditioned (i.e., constrained) on external drivers and accounts for the allocation dynamics variability via stochastic internal mechanisms. Furthermore, special attention was given in the modularity of the model in order to provide the means for its straightforward extension and reproducibility. Finally, emphasis was given on optimizing the simulation speed (i.e., computational effort), a common weakness of most CA models. The model is comprised of two main modules, an external subsystem that generates the number of new urban cells in the next simulation step, making the model constrained to exogenous drivers, and an internal subsystem that allocates these cells spatially through the use of stochastic mechanisms that have descriptive properties. Each subsystem is calibrated to real data before the urban growth prediction. Specifically, the parameters of the internal CA model are calibrated against the similarity of the output to the real urban changes across a specific timespan, using various (weighted) metrics such as the modified Kappa coefficient of agreement (Ksim), shape parameters and number of clusters. Due to the stochastic nature of the model a Monte-Carlo technique is applied. The model is executed a pre-specified number of times and the median performance index is used as input to a genetic algorithm for optimization. An urban classifier model is developed alongside the CA model, in order to enable it to use open data from remote sensing sources, such us Landsat images. This is a key point of the methodology as in many cases suitable data for applying urban growth models is scarce. The urban classifier model uses an artificial neural network (ANN) structure utilizing not only multispectral bands as inputs, but also the multivariate texture info extracted by novel techniques from the satellite imagery, such as multi-variate variogram with spectral angle distance. Haralick-GLCM texture indices and multi-variogram based metrics are tested alongside multispectral data in a Monte-Carlo calibration scheme. Control samples for calibration and validation are randomly selected from a predefined pool (selected for the most recent image from basemaps, Google Earth etc.) and the classification performance is evaluated a thousand times. The set of parameters with the best distribution of performance is used to select a classifier type, that classifies the most recent satellite image (because control samples are readily applicable only for a recent image in most cases). The output of the classifier is then combined with an object-oriented hierarchical (in a temporal manner) classification methodology to derive the required historical land-use change timeseries from the other preceding images. The latter is required to calibrate the parameters of the CA model. The methodology is applied to the general area around Rethymno city in Crete, so as to include various settlements that differ significantly in size and patch density. The mixing of different land uses greatly increases the difficulty for both the urban classifier model and the CA. Also, historical data are scarce, incomplete and inaccurate. This paradoxically makes it an ideal case to demonstrate the robustness of the new methodology. The results are very promising, indicating that the methodology is capable of both identifying historical urban growth from open remote sensing data using the urban classifier model and simulating the complex nature of urban growth even in such unfavorable conditions via the CA model. Finally, hypothetical future water resources-related examples highlight practical aspects of the proposed methodology, after future urban growth predictions by the CA, in the context of modern urban water strategic planning.