Rastin SJ, Rhoades DA, Rollins C, Gerstenberger MC, Christophersen A, Thingbaijam KKS. 2022. Spatial distribution of earthquake occurrence for the New Zealand National Seismic Hazard Model revision. Lower Hutt (NZ): GNS Science. 65 p. (GNS Science report; 2021/51). doi:10.21420/YKQ8-1C41.
Abstract
Our goal is to estimate the decadal-to-centurial (~100-year) spatial distribution of earthquake occurrence for the distributed seismicity model of the New Zealand National Seismic Hazard Model (NSHM) update. This report describes the development of candidate models for the spatial seismicity distribution that will need to be assigned to branches of a logic tree to contribute to the distributed seismicity model. Multiplicative hybrid models help us to assess the relative value of different data inputs, including information from fault studies, tectonics, the earthquake catalogue and strain rate models for forecasting the long-term spatial distribution of earthquake rates. We have used catalogue-independent and catalogue-dependent inputs (covariates) to fit the hybrid models to 70 years of New Zealand earthquakes with magnitudes M >4.95 from 1951 to 2020 (350 target earthquakes) using a catalogue with revised magnitudes. The inputs and hybrid models are defined on a spatial grid with 0.1 degree spacing. Catalogue-independent covariates include Proximity to the Plate Interface (PPI), Proximity to Mapped Faults weighted by slip rate (PMF), the Haines and Wallace (2020) maximum shear strain rate (HWS) and the presence or absence of a mapped fault in each cell (FLT). Catalogue-dependent inputs are smoothed seismicity covariates estimated using a variety of spatial kernels and declustering methods. The smoothing methods are proximity to past earthquakes (PPE), Gaussian 50 km (G50) and adaptive spatio-temporal (HELM). These are fitted to seven windows of 10 years with strict separation of contributing and fitting data. We aim to optimise the information gain of hybrid models with respect to a spatially uniform baseline model (SUP). The most informative catalogue-independent covariate over the 70 years is PMF, followed by HWS and PPI. This contrasts with previous findings that strain rates are more informative than all other covariates for forecasting ahead over one or two decades. The information value of PPE and G50 depends on details of how they are defined. When extracted from declustered learning catalogues and combined additively and multiplicatively, they are more informative than all the catalogue-independent covariates. However, the most informative covariate of all is HELM. The best performing multiplicative hybrid has covariates HELM, PMF and HWS. We also consider additive combinations of likelihood models derived from the covariates. The best performing additive hybrid is an optimised linear combination of HELM, PPE and PMF. The final selection of the distributed seismicity model(s) for the NSHM will depend on a range of considerations, only one of which is the information gain. Two models with identical information gains can present very different spatial distributions. Therefore, careful investigations of spatial distributions are vital. We present spatial distributions and information gains for a range of multiplicative and additive model candidates. A selection of these can be advanced for further analysis. Selection criteria include respecting all available information and recognising the limitations of the data and modelling techniques. Previous studies show that the medium-term forecasting model EEPAS (‘Every Earthquake a Precursor According to Scale’) can provide information on future earthquakes expected to occur over the next decade or two. When combined with a multiplicative or additive hybrid, EEPAS should serve to improve estimation of the spatial distribution of expected earthquakes over the next 100 years. We present 20-year forecasts based on the time-lag-compensated EEPAS model (LEEPAS). LEEPAS produces magnitude-dependent spatial distributions. Hybrid model candidates can then be combined in 80:20 ratio with the LEEPAS forecasts. Sensitivity analysis can be applied to such combinations in aid of assigning logic-tree branches. (The authors)
