Varying-coefficient models for geospatial transfer learning

Bussas, Matthias; Sawade, Christoph; Kühn, Nicolas; Scheffer, Tobias; Landwehr, Niels

doi:10.1007/s10994-017-5639-3

Cited by 23 publications

(14 citation statements)

References 30 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We leave this extension as a future research task. In any case, in terms of directions for future research, we think that new developments can also be fulfilled in this research topic by capitalizing on the recent achievements of transfer learning [44] in tasks of spatial and spatio-temporal prediction [45], [46]. A transfer learning method, specifically designed for the trend-based temporal clusters, discovered through the multi-stage machine learning methodology of AutoTiC-NN, may also allow us to transfer a cluster model, learned in a geographical area with adequate data, to a new area with few data.…”

Section: Discussionmentioning

confidence: 99%

A Multi-Stage Machine Learning Approach to Predict Dengue Incidence: A Case Study in Mexico

et al. 2020

View full text Add to dashboard Cite

The mosquito-borne dengue fever is a major public health problem in tropical countries, where it is strongly conditioned by climate factors such as temperature. In this paper, we formulate a holistic machine learning strategy to analyze the temporal dynamics of temperature and dengue data and use this knowledge to produce accurate predictions of dengue, based on temperature on an annual scale. The temporal dynamics are extracted from historical data by utilizing a novel multi-stage combination of auto-encoding, window-based data representation and trend-based temporal clustering. The prediction is performed with a trend association-based nearest neighbour predictor. The effectiveness of the proposed strategy is evaluated in a case study that comprises the number of dengue and dengue hemorrhagic fever cases collected over the period 1985-2010 in 32 federal states of Mexico. The empirical study proves the viability of the proposed strategy and confirms that it outperforms various state-of-the-art competitor methods formulated both in regression and in time series forecasting analysis.INDEX TERMS Clustering, machine learning, time-series analysis, predictive analysis.

show abstract

Section: Discussionmentioning

confidence: 99%

A Multi-Stage Machine Learning Approach to Predict Dengue Incidence: A Case Study in Mexico

et al. 2020

View full text Add to dashboard Cite

show abstract

“…For the current study, location-specific nonergodic adjustment terms are calculated for the empirical GMMs of Abrahamson et al (2014); Boore et al (2014); Campbell and Bozorgnia (2014), hereafter called ASK14, BSSA14, and CB14. The nonergodic adjustment terms are based on a varying coefficient model Landwehr et al (2016); Bussas et al (2017), and the cellspecific attenuation model Dawood and Rodriguez-Marek (2013); Kuehn et al (2019). We do not calculate adjustment terms for the model of Chiou and Youngs (2014), because it models the anelastic attenuation term as magnitude dependent, which is difficult to incorporate into the cell-specific attenuation model.…”

Section: Nonergodic Gmms and Pshamentioning

confidence: 99%

“…Dawood and Rodriguez-Marek (2013) proposed to model source-to-site specific path effects as the sum of atteuation over small cells, which was extended by Kuehn et al (2019) using Bayesian inference to account for the uncertainties associated with the cell-specific attenuation coefficients. Landwehr et al (2016) proposed a fully nonergodic GMM for California, based on a varying coefficient model (VCM) (Gelfand et al, 2003;Bussas et al, 2015Bussas et al, , 2017, where the coefficients of the GMM are functions of source and site location. In the VCM of Landwehr et al (2016), however, the anelastic attenuation parameter was modeled as dependent only on the event location, and thus did not fully capture path effects, which should depend on both event and station locations.…”

Section: Introductionmentioning

confidence: 99%

Comparison of Bayesian Varying Coefficient Models for the Development of Nonergodic Ground-Motion Models

Kuehn¹

2021

Preprint

View full text Add to dashboard Cite

A nonrgodic ground-motion model explicitly takes systematic local source, path and site effects on the predicted ground-motion into account. With an increasing number of ground-motion records, it s possible to estimate these effects. Landwehr et al. (2016) proposed a varying coefficient model as a tool to estimate nonergodic ground-motion models, based on Gaussian processes. Gaussian processes are computationally expen- sive, so for large data sets, some approximations have to be made. Here, we compare different Bayesian implementations of varying coefficient models, using the probabilistic programming language Stan (Carpenter et al., 2017), and the integrated nested Laplace approximation (INLA) (Rue et al., 2009). The models are used to fit nonergodic models on the California subset of the NGA-West2 data set (Ancheta et al., 2014). We find that both implementations lead to very similar results, both in the estimated parameters and the predicted ground motions.

show abstract

“…The model presented in this study is a fully non-ergodic GMM that captures the systematic effects of the source, site, and anelastic attenuation from the path. It is developed as spatially varying coefficient model (VCM), following the methodology used in Bussas et al (2017) and Landwehr et al (2016). The non-ergodic anelastic attenuation is modeled with the cell-specific o f E a r t h q u a k e E n g i n e e r i n g anelastic attenuation similar to Dawood and Rodriguez-Marek (2013) and Abrahamson et al (2019).…”

Section: Introductionmentioning

confidence: 99%

A Non-Ergodic Effective Amplitude Ground-Motion Model for California

Lavrentiadis

Abrahamson

Kuehn

2021

Preprint

View full text Add to dashboard Cite

A new non-ergodic ground-motion model (GMM) for effective amplitude spectral (EAS) values for California is presented in this study. EAS, which is defined in Goulet et al. (2018), is a smoothed rotation-independent Fourier amplitude spectrum of the two horizontal components of an acceleration time history. The main motivation for developing a non-ergodic EAS GMM, rather than a spectral acceleration GMM, is that the scaling of EAS does not depend on spectral shape, and therefore, the more frequent small magnitude events can be used in the estimation of the non-ergodic terms. The model is developed using the California subset of the NGAWest2 dataset Ancheta et al. (2013). The Bayless and Abrahamson (2019b) (BA18) ergodic EAS GMM was used as backbone to constrain the average source, path, and site scaling. The non-ergodic GMM is formulated as a Bayesian hierarchical model: the non-ergodic source and site terms are modeled as spatially varying coefficients following the approach of Landwehr et al. (2016), and the non-ergodic path effects are captured by the cell-specific anelastic attenuation attenuation following the approach of Dawood and Rodriguez-Marek (2013). Close to stations and past events, the mean values of the non-ergodic terms deviate from zero to capture the systematic effects and their epistemic uncertainty is small. In areas with sparse data, the epistemic uncertainty of the non-ergodic terms is large, as the systematic effects cannot be determined. The non-ergodic total aleatory standard deviation is approximately 30 to 40% smaller than the total aleatory standard deviation of BA18. This reduction in the aleatory variability has a significant impact on hazard calculations at large return periods. The epistemic uncertainty of the ground motion predictions is small in areas close to stations and past event.

show abstract

Varying-coefficient models for geospatial transfer learning

Cited by 23 publications

References 30 publications

A Multi-Stage Machine Learning Approach to Predict Dengue Incidence: A Case Study in Mexico

A Multi-Stage Machine Learning Approach to Predict Dengue Incidence: A Case Study in Mexico

Comparison of Bayesian Varying Coefficient Models for the Development of Nonergodic Ground-Motion Models

A Non-Ergodic Effective Amplitude Ground-Motion Model for California

Contact Info

Product

Resources

About