Machine learning based anomaly detection for sedimentological data: Application to a Holocene multi-proxy paleoenvironmental reconstruction from Laguna Boquita, Jalisco, Mexico

Paleoenvironmental reconstructions are critical to determine past climatological and hydrological conditions, such as sea-level rise (SLR) and extreme events including hurricanes and tsunamis. While established quantitative methods, such as principal components analysis and discriminant analysis, have effectively aided reconstructions by demarcating stratigraphic zones, they suffer from limitations due to the underlying assumptions (linearity, normality). Here, we introduce the machine learning technique anomaly detection for sedimentological reconstructions, capable of objectively pinpointing events (anomalies) in sediment cores. We tested this technique on five sediment cores extracted from Laguna Boquita, located along the Pacific coast of Mexico. Each core was subjected to high resolution loss-on-ignition, while the most representative core (core 1) was deemed the training core and scanned with a handheld XRF unit. In general, the sediment cores were dominated by thick units of peat and/or clay, and most cores contained sand layers. This reconstruction represents many environmental settings, ranging from a sandy terrestrial environment (∼6830- ∼ 6370 cal yr BP), organic-rich wetland (∼6370- ∼ 5170 cal yr BP), and clastic-rich backbarrier lagoon (∼5170 cal yr BP to present). The anomaly detection technique proved effective in marking many events, including marine transgression, evidence of SLR, transitions separating dominant depositional environments, and, perhaps most notably, a faded blue clay layer with minimal LOI variability that may represent a shift in backbarrier water level. The anomaly detection also registered anomalies when events were not visually distinct nor present in the LOI datasets (false positives), which represent sediment core sections that require further investigation with multiple proxies. However, anomaly detection failed to register anomalies in certain core sections that should have registered as events due to their distinct nature. Future efforts will look to improve anomaly detection by choosing different train cores and adding additional proxy datasets.

古环境重建对于确定过去的气候和水文条件至关重要，例如海平面上升（SLR）以及包括飓风和海啸在内的极端事件。虽然已有的定量方法，如主成分分析和判别分析，通过划分地层区域有效地辅助了重建工作，但由于其潜在假设（线性、正态性）而存在局限性。在此，我们将机器学习技术中的异常检测引入到沉积学重建中，它能够客观地确定沉积岩芯中的事件（异常）。我们在从墨西哥太平洋沿岸的博基塔泻湖提取的五个沉积岩芯上测试了这项技术。每个岩芯都进行了高分辨率的烧失量测定，而最具代表性的岩芯（岩芯1）被视为训练岩芯，并用手持式XRF设备进行了扫描。总体而言，沉积岩芯以厚层的泥炭和/或黏土为主，大多数岩芯含有砂层。这次重建代表了多种环境背景，从砂质陆地环境（约公元前6830 - 约6370年）、富含有机物的湿地（约公元前6370 - 约5170年）到富含碎屑的障壁后泻湖（约公元前5170年至今）。异常检测技术在标记许多事件方面被证明是有效的，包括海侵、海平面上升的证据、区分主要沉积环境的过渡，以及也许最值得注意的是，一个烧失量变化极小的浅蓝色黏土层，它可能代表障壁后水位的变化。当事件在视觉上不明显且在烧失量数据集中不存在时，异常检测也会记录异常（假阳性），这些代表需要用多种替代指标进一步研究的沉积岩芯部分。然而，由于某些岩芯部分具有独特性质本应被记录为事件，但异常检测却未能在这些部分记录到异常。未来的工作将着眼于通过选择不同的训练岩芯和添加额外的替代指标数据集来改进异常检测。