Impact of the Degree of Incorporation of Domain Knowledge into the Automatic Calibration of Land Use (Natural Hazard Exposure) Models

Natural hazard risk is generally quantified using a combination of hazard, exposure and vulnerability. Consequently, quantifying future risk requires integrated modelling of these three components over time in response to drivers of change, such as climate change and population growth. We have developed such an integrated model to better understand future natural hazard risk, UNHARMED, in which land use and building stock models simulate changes in exposure. Application of the model to different regions requires calibration of the land use model to the local conditions. Calibration is critical given the sensitivity of the model to changes in land use, however it remains challenging and hence is the focus of this presentation. Automated methods for calibrating land use models are being increasingly used for this purpose to increase the efficiency and repeatability of the calibration process. However, there are multiple challenges with implementing automatic calibration. First, there is a need to appropriately balance individual cell-by-cell agreement of modelled outputs with historical change whilst capturing realistic processes by which land uses change, generally quantified by metrics measuring landscape pattern agreement. Second, there is a need for automatic calibration to explore all parameter combinations that optimise these two objectives while exploiting domain knowledge to ensure parameters are consistent with expected values achieved through process-specific approaches. In this study, the performance of three calibration approaches with different degrees of reliance on domain knowledge were compared for a case study of Madrid, Spain. Results indicate that a hybrid approach, combining optimisation with process-specific output, provided a good balance between competing needs and hence good model performance.