Abstract
Ma and Nystuen [1] successfully detected and estimated rainfall at sea from passive acoustics. They detected rain from three narrowband frequencies (5.4, 8.3 and 21 kHz), and then estimated log rainfall rate via a regression with energy in the 5 kHz band. Mallary et al. [2] improved rainfall detection by exploiting broadband spectra while reducing the dimensionality through principal component analysis (PCA). This project moves beyond detection to estimate the rainfall using PCA-reduced acoustic power spectral densities (PSD). This thesis compares two estimation methods: regression and quantization. Both estimation techniques start with the binary detection of rainfall with a Support Vector Machine (SVM) decision boundary [3]. After binary detection, the regression estimates the rain rate on the rain detections using a Linear Minimum Mean Square Error regression trained on the PCA-reduced PSDs against the ground truth rain. The quantizer optimizes 5 rainfall quantization levels using the Lloyds Algorithm [4], then trains a multiclass classifier on the rainy PSDs exploiting Dietterich and Bakiri’s error-correcting output codes [5] with multiple binary SVM classifiers. This classifier classifies the binary rain detections into one of 5 rain rate ranges and then estimates the rain rate by quantizing each range, or class, to the levels defined by the Lloyds algorithm. These methods are compared using 43 days of acoustic and meteorological data collected on a mooring in Buzzards Bay, MA.