Implementation of Neural Network Models Using Acoustic and Spectral Features for Rainfall Intensity Classification
June Lorenz E. Capin, Ericson D. Dimaunahan · 2023
Accurate detection and unobstructed dissemination of rainfall data are vital for critical sectors of society, as it ensures people's safety and economic stability in a wide range of scenarios. This is increasingly evident due to its effects on climate change, prompting government units to take proactive measures in mitigating disaster risk and improving their responses. In this study, neural network models are created, which are capable of interpreting environmental acoustic signals caused by rainfall into an estimate of its intensity using a classification scheme. These models are built using Feedforward Neural Network (FNN) and Convolutional Neural Network (CNN) blocks, each taking numerical properties and spectrogram images of the acoustic signal as their respective inputs. Model evaluations done in a test dataset determined that the use of both numerical and image data as inputs to concatenated FNN and CNN blocks made the resultant model gain a high degree of generalization, reaching an accuracy of 97.43%. These models can be used as a basis for the development of a semi-independent computer system for the same purpose.