Acoustic rainfall estimation with support vector machines and error correcting output codes

Abstract

Ma and Nystuen (2005) successfully detected and estimated rainfall at sea from passive acoustics. They detected rain from three narrowband frequencies and then estimated log rainfall rate via a regression with energy in the 5 kHz band. Mallary et al. (2022) improved rainfall detection by exploiting broadband spectra while reducing the dimensionality through principal component analysis (PCA). This project builds upon Mallary’s work moving beyond detection to estimate the rainfall by quantization into discrete ranges based on PCA-reduced acoustic power spectra. The classification scheme combines multiple binary support-vector machine (SVM) classifiers (Boser et al. 1992) with Dietterich and Bakiri’s error-correcting output codes (1995) to classify acoustic PSDs into one of 6 rainfall rate classes. Evaluating the PCA/SVM classifier on 4 months of acoustic recordings and meteorological data collected from a shallow water pier in New Bedford, MA found the hourly accumulations from the rain gauge and acoustic estimates had a correlation of 0.97 ± 0.01. Emulating Ma & Nystuen’s estimator on the same data set yields a correlation of 0.76 ± 0.02. [Work supported by ONR.]