Ripeness classification of cocoa through acoustic sensing and machine learning

In Cocoa harvesting, the perceived hollow sound from tapping the Cocoa pod is the conventional way of determining ripeness. In this paper, acoustic sensing device was used to record noiseless acoustic signals generated from tapping cocoa pods while on tree. Acoustic data were collected from cocoa pods of two classifications, namely, ripe and unripe. Frequency-domain analysis was used using Fast Fourier Transform (FFT) in extracting the spectral characteristics, namely, the first three dominant resonant frequencies, their corresponding amplitudes, and their power spectral densities. Time-domain features particularly the Short-time Energy and Zero-Crossing Rate were also used in this study. The eleven acoustic features of unripe and ripe samples were examined using Scatter plots. From 392 WAV files, 272 were used as training datasets and the remaining were used as testing datasets. The experimental results showed that the combination of the first two dominant resonant frequencies into a feature vector using Support Vector Machine (SVM) classifier tool gave the maximum classification accuracy. The classification model output was tested and found to correctly classify cocoa ripeness with 95.8% overall accuracy.