Hearing the forest for the trees: machine learning and topological acoustics for remote sensing with seismic noise

TL;DR

Together, these results provide the first demonstration that subtle forest-wave interactions manifest in ambient seismic noise and can be harnessed as a scalable tool for continuous vegetation monitoring, offering a robust solution for tracking environmental change challenging regions.

Abstract

Monitoring remote forests is a global challenge central to climate mitigation and biodiversity conservation, yet satellite observations are frequently limited by weather, dense canopies, and solar dependency. Here we show that passive seismic sensing offers a persistent, all-weather alternative for autonomous ecosystem monitoring by capturing characteristic learnable signatures of trees within the ambient wavefield. Using seismic data from Alaska, we demonstrate that cross-correlations between stations provide a physical basis for forest detection by approximating the empirical Green's function of the medium. Supervised machine learning models applied to these data achieve a classification accuracy of 86%, identifying key discriminating frequencies (35 to 60 Hz) consistent with known forest-wave interactions. A topological acoustics analysis of the geometric phase change independently confirms the physical origin of these data-driven classifications. Together, these results provide the first demonstration that subtle forest-wave interactions manifest in ambient seismic noise and can be harnessed as a scalable tool for continuous vegetation monitoring, offering a robust solution for tracking environmental change challenging regions.

Figures (11)

• Figure 1: Graphical overview of the two machine learning workflows employed in this study based on the raw seismic data ("Raw") and cross-correlations (CC). DPx (x$\in\{1, 2, Z\}$) are the signal components considered in the study, multiple machine learning models considered are labeled as algorithm-component-input type (e.g., SVM-DPZ-CC).
• Figure 2: Geographic location of sensors in Alaska near Denali Fault and Richardson Highway intersection. The green dots represent the seismic stations whose signals were used. The blue and orange boxes indicate the forested and non-forested regions, respectively. The dashed lines show the split of the sensors within both regions for intra-class control analysis.
• Figure 3: Cross-correlation record sections of (a) DPZ and (b) DP1 components of the signals from forested and non-forested regions obtained with a bandpass filter from 10 to 100Hz. Red lines indicate the approximate arrivals of primary phases with a velocity of 700ms.
• Figure 4: Confusion matrices of the (a) SVM and (b) RF models based on different input signals. Percentages (and the colors) in the cells represent true positive (TP) rate, false negative rate, true negative (TN) rate, and false positive rate (in clockwise order in every $2\times2$ confusion matrix). F indicates forest, NF non-forest, CC cross-correlation. The percentage values under the input signal labels represents overall classification accuracy defined as $(\text{TP}+\text{TN})/n$, where $n$ is the total number of samples in the given test set (\ref{['tab:dataset']}).
• Figure 5: ROC curves obtained for six input signals (a) using SVM, (b) using RF. The solid lines and dashed lines represent cross-correlation (CC) and raw signals (raw), respectively. Dotted lines represent ROC curves from intra-class analysis within forested (F) and non-forested (NF) regions (obtained with SVM of the DPZ cross-correlations).