Estimation of Tree Height and Volume Based on the Produced Hammering Sound Using Deep Learning

Abstract

This study aimed to estimate the tree height and volume based on image data, which are obtained by conversion of the sound generated from hammering a stem using deep learning. We hammered 20 trees 100 times, recorded the hammering sound, and generated the spectrogram, which presented the sound pressure at each frequency for 0.6 s. Data comprising 10,000 images were loaded into a deep learning model. We used the Neural Network Console (NNC) as the deep learning system and the LeNet, which forms a programmed regression layer to an output layer, as the deep learning algorithm. We divided 10,000 images into 5 equal sets, and performed three learning patterns (LP-I, LP-II, and LP-III). LP-I used the four sets as training data and the remainder as test data, LP-II used three trees and LP-III used six trees, which select for each tree from three divisions (large, medium, small), as test data. A performance evaluation of the proposed model was performed using three indicators: mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R2). Each learning pattern provided very good estimates (R2 values for each learning pattern for the test data ranged from 0.9192 to 0.9996), except for the height estimate by LP-III (R2＝0.3672). LP-III generated very poor height estimates with a bias and tended to underestimate by＞30 m and overestimate by＜30 m. However, each learning pattern provided a good estimate of the tree volume, generally without any bias. Thus, we found this method to be more effective for estimation of tree volume than tree height.