A Hierarchical Method for Environmental Sounds Recognition in Complex Acoustic Scenes

• Sound pressure level indicators alone cannot identify noise sources, while source identification can support noise-source analysis, targeted monitoring, and urban noise management.Source: §1 · p. 1
• Compact models that capture discriminative temporal and spectral information remain needed, since deeper convolutional architectures can increase computational costs and restrict deployment.Source: §1 · p. 2
• Single-stage classifiers can confuse acoustically similar classes under imbalance, especially underrepresented classes, while lightweight time-frequency modeling integrated with hierarchy remains underexplored.Source: §1 · p. 2
• FTANet uses asymmetric one-dimensional convolutions to extract temporal dynamics and frequency-band distributions separately, then applies adaptive fusion attention to recalibrate features.Source: §1 · p. 2
• The proposed hierarchy decomposes 25 fine-grained classes into smaller within-category tasks based on first-level classification results to alleviate inter-category confusion under imbalance.Source: §1 · p. 2

2.1. Network Architecture

• The architecture processes preprocessed time-frequency features through FTANet and routes each sample to the fine-grained classifier selected by its first-stage category prediction.Only the first-stage model and one corresponding second-stage model are invoked during inference.Source: §2.1 · p. 2, §2.3 · p. 7

2.2. Frequency-Time Adaptive Network (FTANet)

• FTANet uses parallel asymmetric one-dimensional convolutions to model time and frequency features separately, then applies attention to recalibrate their fused representation.The time dimension represents acoustic-event dynamics, while the frequency dimension represents energy distribution across bands.Source: §2.1 · p. 2, §2.2 · p. 4

2.2.1. Frequency-Time Convolution Module (FTConv)

• FTConv uses separate frequency- and time-branch kernels, concatenates their outputs, and applies 1 × 1 convolution, batch normalization, and SiLU for fusion.The frequency kernel is 3 × 1 and the time kernel is 1 × 3; fused features feed the adaptive fusion attention module.Source: §2.2.1 · p. 5

2.2.2. Adaptive Fusion Attention Mechanism

• AFA pools each fused channel across frequency and time, then uses a bottleneck fully connected structure and Sigmoid weights to recalibrate channel responses.The channel weights are constrained to [0,1], with the bottleneck modeling nonlinear relationships while reducing parameter overhead.Source: §2.2.2 · p. 5

2.2.3. Multi-Stage Feature Modeling and Classification Output

• FTANet repeats FTConv, AFA, residual learning, and max pooling across three stages, progressively reducing time-frequency resolution while building higher-level representations.The third stage maintains 64 channels to avoid additional parameter overhead from further widening.Source: §2.2.3 · p. 6

2.3. Two-Stage Hierarchical Classification Architecture

• The two-stage system assigns sounds to coarse categories before invoking the corresponding within-category classifier for fine-grained recognition.It organizes independent FTANet classification models around the hierarchy without changing FTANet’s internal architecture.Source: §2.3 · pp. 7–8

3.1. Data Preprocessing and Training Strategy

• The self-constructed dataset contains five coarse-grained categories covering 25 fine-grained sound classes.The category taxonomy includes industrial, construction, transportation, social activity, and natural sounds.Source: §3.1 · p. 8, Table 1 · p. 8
• The 9,029 samples are imbalanced and randomly split within each fine-grained class into training, validation, and test sets at 6:2:2.Some common sounds have more samples than occasional sounds or sounds easily affected by background noise.Source: §3.1 · p. 9
• Audio is converted to mono and resampled to 16 kHz before 80-filter Fbank extraction using 25 ms analysis windows.The input representation extends the Fbank features with time- and frequency-direction differences to form a three-channel tensor.Source: §3.1 · p. 9
• The coarse classifier and fine-grained subclassifiers are trained independently, with each model trained for 15 epochs.Training uses Adam, a 5 × 10^-4 initial learning rate, 1 × 10^-4 weight decay, and batches of 32.Source: §3.1 · p. 10

3.2. Performance Comparison with Conventional Models

• Table 2 compares methods and categories on the self-constructed dataset, reporting classification metrics, parameter counts, and FLOPs.Its columns include precision, recall, F1-score, overall accuracy, and Macro-F1.Source: Table 2 · p. 10, §3.2 · p. 10

FTANet

• 95.45% accuracy and 95.23% Macro-F1 were achieved by FTANet with 0.2004 M parameters and 0.3011 G FLOPs, second only to AST on the self-constructed dataset.AST achieved 96.48% accuracy and 96.34% Macro-F1, but used 85.2588 M parameters and 51.2949 G FLOPs.Source: FTANet · p. 11

3.3. Comparison of Single-Stage and Two-Stage Classification Architectures

• The two-stage model improves fine-grained recognition and reduces some intercategory confusion, but raises inference time from 0.2170 s to 1.2390 s.Confusion-matrix and t-SNE visualizations show more concentrated class responses and clearer clusters for some classes under hierarchical classification.Source: §3.3 · pp. 12–13, Figure 8 · p. 13, Figure 9 · p. 13

3.4. Ablation Study of FTANet

• Adding AFA raises accuracy from 93.00% to 93.10% with standard 3×3 convolutions and from 93.59% to 93.94% with FTConv, adding about 0.005 M parameters and 0.0003 G FLOPs.The reported gains occur with both convolutional structures.Source: §3.4 · p. 14
• Feature fusion alone lowers FTConv accuracy from 93.59% to 92.94%, while combining fusion and AFA yields 95.45%, 2.51% above Method B and 1.51% above Method C.The authors report that AFA recalibrates channel responses after the 1×1 convolution mixes time and frequency branches.Source: §3.4 · p. 14
• 95.45% accuracy, 0.2004 M parameters, and 0.3011 G FLOPs were achieved by complete FTANet, outperforming the standard 3×3 convolutional baseline by 2.45% accuracy.The ablation compares configurations with FTConv, 1×1 feature fusion, and AFA, holding other network structures and training settings identical.Source: §3.4 · pp. 13–14, Table 5 · p. 14

4. Application

• The platform first classifies five coarse categories; in this deployment, only natural sounds receive second-stage fine-grained classification, while the other categories are output directly.Sensors transmit recordings to a cloud server for analysis and recognition, with results returned to display and early-warning modules.Source: §4 · pp. 14–15, Figure 10 · p. 14
• The field-sample distribution reflects site conditions: no valid C1 samples were recorded, and only 18 C4 samples were available.The passage attributes absent sound classes to conditions at the deployment site, including no strong thunderstorms and weak wind sounds that were not captured as valid target events.Source: §4 · p. 16
• 97.21% weighted-average accuracy and 98.12% weighted-average F1-score were achieved by the natural-sound classifier on 215 samples correctly routed to C2.Across both routing and fine-grained recognition, C2 end-to-end accuracy and F1-score were 77.41% and 86.85%, respectively.Source: §4 · p. 16
• 87.18% overall weighted accuracy and 87.40% weighted-average F1-score were achieved across 788 valid field samples.The 788 samples were retained from 1,000 recordings collected over nearly five months after 212 invalid recordings were removed.Source: §4 · pp. 15–17, Table 6 · p. 16
• Predictions at both classification stages are generally concentrated along the confusion-matrix diagonal, with remaining errors in a small number of coarse categories and natural-sound subclasses.The field evaluation treats a prediction as correct when it matches any assigned true label and uses weighted-average metrics for the imbalanced sample distribution.Source: §4 · pp. 16–17, Figure 12 · p. 17

5. Conclusions

• FTANet uses asymmetric one-dimensional convolutions to model temporal dynamics and frequency-band distributions, with adaptive fusion attention recalibrating fused channels.The proposed hierarchy first discriminates coarse categories and then performs within-category fine-grained recognition.Source: §5 · p. 17
• On the 25-class task, hierarchical classification raises overall accuracy from 71.32% to 81.49% and Macro-F1 from 68.51% to 79.58%.On UrbanSound8K, FTANet achieves 79.24% mean accuracy and 80.37% mean F1-score across ten folds; the paper reports that it outperforms convolution-based comparison models.Source: §5 · p. 17
• Across 788 valid field samples, the deployed method achieves 87.18% overall weighted accuracy and 87.40% weighted-average F1-score.The authors conclude that these field results verify practical feasibility for urban acoustic environment monitoring.Source: §5 · pp. 17–18