Institutional Repository
Technical University of Crete
EN  |  EL



My Space

Predicting spatial and temporal changes in groundwater levels using artificial Neural networks and geostatistical methods

Tapoglou Evdokia

Full record

Year 2015
Type of Item Doctoral Dissertation
Bibliographic Citation Evdokia Tapoglou, "Predicting spatial and temporal changes in groundwater levels using artificial Neural networks and geostatistical methods", Doctoral Dissertation, School of Environmental Engineering, Technical University of Crete, Chania, Greece, 2015
Appears in Collections


The purpose of this study is to create a spatial and temporal simulation program for the estimation of the hydraulic heads in an aquifer, using computational intelligence and geostatistical methods. Artificial neural networks were chosen for the temporal simulation, due to their proven ability, established in literature. This choice is further reinforced by the fact that their implementation does not require complex and hard to obtain input data. The methodology was completed by using the Kriging method for the spatial interpolation of the artificial neural network’ simulation results. A fuzzy logic system was employed to combine these two methodologies.The first step in the proposed approach is the collection of all available data. In this study, data were collected and the hydraulic head was simulated for two study areas; one in Bavaria, Germany and one in Miami, Florida, USA. These two regions have very different characteristics (geological, climatic and land use) and for this reason the successful simulation of hydraulic head in these regions can confirm the reliability of the model. Next, the simulation using artificial neural networks was performed, while at the same time the use of a fuzzy logic system for the selection of suitable neighbors for Kriging algorithm was tested. In this way, the use the fuzzy logic system was proven to be the more effective. Finally, the interpolation of the point hydraulic head estimation was performed by kriging using three different variogram models in each case. The results are evaluated through a series of cross-validation results error indicators. For the first study area in Bavaria, the power-law variogram was the most appropriate variogram yielding a root mean square error (RMSE) equal to 7.6·10-3 m, while for the second study area in Miami, the most appropriate variogram was the exponential one with error indicator of RMSE = 0.962 m. The key difference between the two case studies, leading to this deviation in RMSE values is that, the first case the hydraulic head change per time step was simulated, while in the second study area the hydraulic head itself was simulated. This difference in output parameters was due to the dataset involved in every case. Uncertainty analysis in different components of the methodology was performed in order to validate the model and verify its accuracy. The results of these processes showed that the methodology used was efficient, accurate and stable with regards to its results and errors.The innovation of this thesis is the development of a data-based model which can be used to predict the hydraulic head in an aquifer both spatially and temporally, without having knowledge about the geological characteristics of the study area. Instead data available for a long time and is easy monitor can be used. Moreover, the use of fuzzy logic to determine the kriging neighbors is an innovation of this study and improves considerably the results. The implementation of the model in two study areas with different geological background, an alluvial and a karstic proves its effectiveness under various conditions.

Available Files