In mountainous areas, rockfalls, rock avalanches, and debris flows constitute a risk to human life and property. Seismology has proven a useful tool to monitor such mass movements, while increasing data volumes and availability of real-time data streams demand new solutions for automatic signal classification. Ideally, seismic monitoring arrays have large apertures and record a significant number of mass movements to train detection algorithms. However, this is rarely the case, as a result of cost and time constraints and the rare occurrence of catastrophic mass movements. Here, we use the supervised random forest algorithm to classify windowed seismic data on a continuous data stream. We investigate algorithm performance for signal classification into noise (NO), slope failure (SF), and earthquake (EQ) classes and explore the influence of non-ideal though commonly encountered conditions: poor network coverage, imbalanced data sets, and low signal-to-noise ratios (SNRs). To this end we use data from two separate locations in the Swiss Alps: data set (i), recorded at Illgraben, contains signals of several dozen slope failures with low SNR; data set (ii), recorded at Pizzo Cengalo, contains only five slope failure events albeit with higher SNR. The low SNR of slope failure events in data set (i) leads to a classification accuracy of 70 % for SF, with the largest confusion between NO and SF. Although data set (ii) is highly imbalanced, lowering the prediction threshold for slope failures leads to a prediction accuracy of 80 % for SF, with the largest confusion between SF and EQ. Standard techniques to mitigate training data imbalance do not increase prediction accuracy. The classifier of data set (ii) is then used to train a model for the classification of 176 d of continuous seismic recordings containing four slope failure events. The model classifies eight events as slope failures, of which two are snow avalanches, and one is a rock-slope failure. The other events are local or regional earthquakes. By including earthquake detection of a permanent seismic station at 131 km distance to the test site into the decision-making process, all earthquakes falsely classified as slope failures can be excluded. Our study shows that, even for limited training data and non-optimal network geometry, machine learning algorithms applied to high-quality seismic records can be used to monitor mass movements automatically.

中文翻译：

---

具有连续随机森林的边坡破坏地震信号的近实时自动分类

在山区，落石，雪崩和泥石流对人类生命和财产构成威胁。事实证明，地震学是监视此类质量运动的有用工具，同时不断增加的数据量和实时数据流的可用性要求自动信号分类的新解决方案。理想地，地震监测阵列具有大的孔径并记录大量的质量运动以训练检测算法。但是，由于成本和时间的限制以及灾难性群众运动的罕见发生，这种情况很少发生。在这里，我们使用监督随机森林算法对连续数据流上的加窗地震数据进行分类。我们研究了将信号分类为噪声（NO），斜坡故障（SF），和地震（EQ）类，并探索非理想的但通常遇到的条件的影响：网络覆盖范围差，数据集不平衡以及信噪比（SNR）低。为此，我们使用来自瑞士阿尔卑斯山两个不同地点的数据：在Illgraben记录的数据集（i）包含数十个低SNR的斜坡故障信号；在Pizzo Cengalo记录的数据集（ii）尽管包含较高的SNR，但仅包含五个边坡破坏事件。数据集（i）中的边坡破坏事件的低SNR导致SF的分类精度为70％，其中NO和SF之间的混淆最大。尽管数据集（ii）高度失衡，但降低边坡破坏的预测阈值会导致SF的预测精度达到80％，而SF和EQ之间的混淆最大。减轻训练数据不平衡的标准技术不会提高预测准确性。然后使用数据集（ii）的分类器来训练一个模型，以对包含四个边坡破坏事件的176d连续地震记录进行分类。该模型将8个事件分类为边坡破坏，其中2个是雪崩，1个是岩石-边坡破坏。其他事件是局部或区域地震。通过将与测试地点相距131 km的永久地震台的地震检测纳入决策过程，可以排除所有错误分类为边坡破坏的地震。我们的研究表明，即使对于有限的训练数据和非最佳的网络几何形状，应用于高质量地震记录的机器学习算法也可以用于自动监测质量运动。