Acoustic rainfall detection with linear discriminant functions of principal components

Abstract

Ma and Nystuen (2005) pioneered passive acoustic measurement of rainfall rates. This project extends their work with signal processing algorithms exploiting the full frequency band of the acoustic signals. We also extend Schwock and Abadi’s order-statistic power spectral density (PSD) estimation for outlier rejection to reject recreational anthropogenic noise sources and reject diurnal biological sources using two hydrophones spaced by 1 m. Ma and Nystuen reduced the data dimensionality by extracting a few "discriminant frequencies." Our proposed detection algorithm implements principal component analysis (PCA) to reduce the estimated PSD to two principal components. Linear discriminant analysis (LDA) provides a simple detection statistic from the two dimensional principal components. We evaluated our algorithm on four months of acoustic and meteorological data collected from a dock in New Bedford, MA in shallow water (3 m deep). For 1% false alarms, the proposed PCA/LDA algorithm correctly detected 36% (±7%) of rain events exceeding 1 mm/hr, including 64% (±7%) of the rain by volume. Applying Ma and Nystuen’s algorithm to the same data set for the same false alarm rate detected 23% (±11%) of events containing 52% (±26%) of the rainfall volume. [Work supported by ONR.]