Abstract Sedimentation rate data are applied in a range of studies such as understanding the sediment carbon budget and accumulation of pollutants in marine sediments. Over many years, sedimentation rate samples have been collected within the Baltic Sea. However, to understand patterns in sedimentation rates more broadly across this sea basin requires converting these point data into continuous spatial predictions. The generation of continuous maps remains problematic and under-studied. This study explores the feasibility of machine learning to estimate sedimentation rates, using 137Cs measurements from the Baltic Sea. A random forest model was applied to predict sedimentation rates based on a range of predictor variables that related to the hydrodynamic regime, bathymetric complexity of the seabed, substrate type and proximity to sediment sources. The accuracy of this prediction was tested against an independent set of sedimentation rate samples that had been withheld from model training. The model was also compared against a simple spatial interpolation to assess whether machine learning produces an improvement on previously applied methods. Overall, the modelling approach explained 41.9% of the variance, far surpassing the spatial interpolation method (4.2% variance explained). This study is a first step towards the spatial prediction of sediment accumulation rates in a repeatable and validated way, but further refinement is desirable to improve the accuracy of the predictions. We discuss the potential sources of model error that could have limited the success of this approach and suggest how some of these could be addressed. Our results indicate that short-term sedimentation rates are highest in small coastal basins, while rates in the deep basins of the Baltic Sea are generally low, thereby seemingly contradicting long held views of the deep-basins as the major depocentres in the Baltic Sea. This apparent contradiction might be attributed to the higher spatial detail of our analysis and differences in the time scales of analysis, indicating that sedimentation patterns in the Baltic Sea might be complex in space and time.

中文翻译：

---

波罗的海的沉积率：一种机器学习方法

摘要 沉积速率数据被应用于一系列研究，例如了解沉积物碳收支和海洋沉积物中污染物的积累。多年来，在波罗的海收集了沉积速率样本。然而，要更广泛地了解整个海盆的沉积速率模式，需要将这些点数据转换为连续的空间预测。连续地图的生成仍然存在问题且研究不足。本研究使用来自波罗的海的 137Cs 测量值，探讨了机器学习估算沉积速率的可行性。应用随机森林模型根据与水动力状况、海床测深复杂性、底物类型和与沉积物源的接近程度相关的一系列预测变量来预测沉积速率。该预测的准确性针对一组独立的沉降速率样本进行了测试，这些样本已从模型训练中保留下来。该模型还与简单的空间插值进行了比较，以评估机器学习是否对以前应用的方法产生了改进。总体而言，建模方法解释了 41.9% 的方差，远远超过空间插值方法（解释了 4.2% 的方差）。这项研究是以可重复和经过验证的方式对沉积物积累率进行空间预测的第一步，但需要进一步改进以提高预测的准确性。我们讨论了可能限制这种方法成功的模型错误的潜在来源，并建议如何解决其中的一些问题。我们的研究结果表明，小型沿海盆地的短期沉积速率最高，而波罗的海深盆的速率普遍较低，因此似乎与长期以来将深盆视为波罗的海主要沉积中心的观点相矛盾。这种明显的矛盾可能归因于我们分析的更高空间细节和分析时间尺度的差异，表明波罗的海的沉积模式可能在空间和时间上是复杂的。