AI-Based Analysis of Trunk Tapping Sounds for Urban Tree Decay Screening: A Pilot Study on Liquidambar formosana in Taiwan

In tree risk assessment, visual tree assessment (VTA) commonly involves tapping the trunk or branches to detect abnormal sounds as a basis for further assessment. However, interpreting tapping sounds typically requires extensive professional experience and may be subject to subjective judgment. Therefore, research on the classification of tapping sounds may provide valuable supplementary information for assessing local decay in trees. Tapping sounds from a tree trunk can provide cues to the local internal condition, but their classification performance must be evaluated on trees not seen during training. This study examined 30 sweetgum (Liquidambar formosana) trees at a private nursery on the outskirts of Taipei, using mallet taps recorded with a smartphone, with reference labels established by directly sawing open the trunk cross-section after sampling. Each tap point was assigned to one of three original labels—healthy (L0), mildly decayed (L1), and decayed (L2); the binary classification merged L1 and L2 into an “affected” class for comparison with the healthy class (L0), covering all 966 recordings (449 healthy, 517 affected) without excluding any measurement position. The model comparison included a support vector machine, a random forest, and extreme gradient boosting using a 611-dimensional acoustic feature vector, together with a convolutional neural network (CNN) using log-mel spectrograms, with a deeper classification head and feature fusion serving as robustness checks; all per-fold epoch budgets and decision thresholds were chosen by nested cross-validation (tree-grouped inner validation within each outer training fold), and evaluation used five-fold cross-validation grouped by tree. The four main models performed comparably, with five-fold-averaged macro-F1 between 0.6288 and 0.6548 (highest for XGBoost, lowest for the CNN) and ROC-AUC up to about 0.73; no model was decisively superior, and the deep CNN did not outperform the conventional feature-based classifiers, while neither a deeper head nor feature fusion produced a stable improvement. The fold-to-fold variation of every model exceeded its seed-to-seed variation, and all models were lowest on the same fold, indicating cross-tree generalization as the primary source of variation. Although the pooled area under the curve on the 24 mixed-label trees was comparable to that on the full dataset, the median within-tree ROC-AUC (for the RF model) was only 0.511, close to chance, indicating that the pooled metric mainly reflects between-tree differences rather than within-tree discrimination of local decay; the approach therefore cannot reliably localize decay within an individual tree. An exploratory threshold analysis showed that, because the affected class is the majority and the default threshold is already near-optimal, threshold tuning offers limited benefit. Overall, the results provide preliminary support for the feasibility of using tree-trunk tapping sounds as auxiliary information for population-level decay screening; the approach cannot yet identify decayed locations within a single tree, and validation using additional independent trees and external sites is still needed.