A Supervised Learning Approach for Rainfall Detection From Underwater Noise Analysis

Underwater noise analysis allows estimation of parameters of meteorological interest, difficult to monitor with in situ devices, especially in very harsh environments such as polar waters. Rainfall detection is a fundamental step of acoustical meteorology toward quantifying precipitation and, indirectly, wind. To date, this task has been conducted with some success by using a few frequency bins of the noise spectrum and combining their absolute values and slopes into some inequalities. Unfortunately, these algorithms do not perform well when applied to spectra obtained by averaging multiple noise recordings made over the course of an hour. Supervised, machine learning models allow the use of all the frequency bins in the spectrum, exploiting relationships that are difficult for a human observer to identify. Among the different models tested, a binary classifier based on random forest performed well with moderate computational load. Using a dataset consisting of over 18 000 hourly averaged spectra (approximately 25 months of in situ recordings) and comparing the results with measurements from a surface-mounted rain gauge, the proposed system detects precipitations greater than 1 mm/h with 90% probability, keeping the false alarm probability below 0.5%. This system has demonstrated remarkable robustness as performance is achieved without intentionally excluding any spectra corrupted by sounds produced by other sources, such as naval traffic and wind blowing over the sea surface.