Agricultural harvester sound classification using convolutional neural networks and spectrograms
Esmaeilzadeh Khorasani, Nioosha
This thesis presents deep learning techniques applied to agricultural machines sounds in order to classify them. The objective of the thesis is to determine if it is possible to predict the machines state from the engine sound with high accuracy. The use of deep learning in agricultural tasks has recently become popular. Deep learning networks have been used for analyzing images of crops, identifying paddy areas, distinguishing sick plants from healthy ones, to name a few applications. Besides visual systems, sound analysis of agricultural machinery is a time-sensitive task that can also be incorporated into decision-making and can be done with the help of deep learning models. We propose a method to generate spectrogram images from the sound of a harvester and classify them into three working modes in real-time. We used three convolutional neural networks and use the outputs of these networks as inputs to a stacking ensemble method to improve the accuracy of the system. To achieve 100% classification accuracy, a final decision is made by voting based on several consecutive classifications made by the stacking step. We were able to perform classifications in less than one second which was the standard to be considered a safe time for the harvester.
agricultural machines, deep learning, convolutional neural networks, classification, harvester