A gradient boosted decision tree approach for high-resolution luminescence chronologies

Luminescence ages are powerful agents for tracing past sediment dynamics and deciphering complexities inherent in the evolution of past landscapes. If applied in temporal periods suitable for luminescence-based methods, they provide accurate dating results but with somewhat limited spatial resolution. This is primarily due to the time-consuming nature of luminescence sample preparation and measurement procedures. Luminescence screening methods, for instance, using portable equipment [1] that focuses only on light intensities rather than absorbed dose/dose-rate ratios, provide a convenient shortcut. Assuming a suitable geologically homogeneous environment, they provide initial insights and relative chronologies and can be useful in developing an appropriate sampling strategy for a more detailed study.

However, the hope of delivering, even provisionally, instant chronologies could not yet be satisfied. While our approach similarly cannot offer instant chronologies, we propose here a gradient-boosted decision tree approach [2] to model the complex interactions among physical parameters (e.g., dose rate, water content, sedimentology) and convert luminescence intensity values into the age domain. Our approach uses age-depth relationships of ages and intensities from different profiles, combined with additional features such as geographical information (latitude, longitude, depth below ground surface). We demonstrate that we can satisfactorily and robustly predict pseudo-luminescence ages from signal intensities using only a small training dataset (n = 31). This enables us to considerably enhance the age resolution of luminescence dating chronologies in suitable environments, particularly in those where sedimentary deposits are relatively homogenous.

A limitation of our approach is our reliance on a favourable, homogeneous sampling environment (here, sandy deposits of aeolian origin), which cannot be directly transferred to other geologically more complex settings; however, we are confident that the general approach remains valid and can be adapted on regional scales to increase age resolution.

References

[1] Sanderson, D. C. W. and Murphy, S.: Using simple portable OSL measurements and laboratory characterisation to help understand complex and heterogeneous sediment sequences for luminescence dating, Quaternary Geochronology, 5, 299–305, https://doi.org/10.1016/j.quageo.2009.02.001, 2010.

[2] Chen, T. and Guestrin, C.: XGBoost: A Scalable Tree Boosting System, in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, New York, NY, USA, 785–794, https://doi.org/10.1145/2939672.2939785, 2016.