In this study, all models were implemented in the PyTorch DL framework and executed on the workstation with a 12th Gen Intel^® Core™ i7-12700F Processor (2.10 GHz), 32.0 GB RAM, and one NVIDIA GeForce RTX 4070 GPU.

During the preprocessing stage, the original sampling rate of the acoustic signal is 16KHz. The cutoff frequency is set to half the sampling rate. Window function selects hamming window. The order of the filter is set to 100. Moreover, the training and testing datasets are split in a ratio of 8:2. Additionally, multiple rounds of 5-fold cross-validation are performed to further assess the effectiveness and robustness of the proposed algorithm across different data partitions. The image size of the input model network is 256*256. The models are trained for 150 epochs on each mini-batch with a batch size of 32. The loss function used in all experiments is cross-entropy. All compared models apply early stopping with a patience of five epochs, whereas our method is trained without early stopping. The hyperparameter settings are as follows: ResNet18 and DenseNet adopt the Adam optimizer, whereas EfficientNet-B0, VGG19, MobileNet, and our proposed model employ SGD. The learning rates are set to 0.001 for ResNet18, DenseNet, and MobileNet; 0.01 for EfficientNet-B0 and VGG19; and 0.1 for our model. Dropout is applied at rates of 0.2 for ResNet18, EfficientNet-B0, and MobileNet, and 0.5 for VGG19. DenseNet and our proposed method do not utilize dropout.

In this study, several evaluation metrics are employed to assess model performance. The confusion matrix analyzes the true versus predicted classifications, providing insight into the types of errors made by the model through the enumeration of true positives, true negatives, false positives and false negatives. Top-1 accuracy (Accuracy@1) measures the proportion of instances where the top predicted class matches the true class, reflecting primary classification accuracy. Macro-Recall quantifies the ratio of true positives to the total number of actual positive instances for each class, then takes the arithmetic mean across all classes. This approach ensures equal evaluation weight for all categories, which is critical when each class holds independent importance. The Macro-F1 score, the harmonic mean of Macro-Precision and Macro-Recall, offers a balanced metric that accounts for the trade-off between these two measures, adopting macro-averaging to ensure uniform assessment of classification consistency. Lastly, the receiver operating characteristic (ROC) curve provides a graphical representation of discriminative power across various classification thresholds, plotting the true positive rate against the false positive rate. The Macro-area under the ROC curve (AUC) serves as an aggregate measure of performance calculated by macro-averaging AUC values across classes, with higher AUC values indicating superior classification ability in maintaining inter-class decision boundary coherence. We deliberately employ macro-averaging to guarantee metric interpretability from a class-agnostic perspective, as this method equally weights the decision patterns of all categories. Furthermore, the number of parameters (in millions) and the computational complexity in GigaFloating Point Operations Per Second (GFLOPS) are used in this study. The number of parameters in the model represents the model’s size and capacity, and GFLOPS is a measure of the computational complexity of the model, with higher values indicating more computation is required.

To ensure rigorous benchmarking, the experimental parameters are standardized across all models: the parameter SegmentFrameNumber Num is fixed at 10 frames per audio segment and the samplingrate is set to 16KHz. As shown in Table 4 , this study demonstrates the superior performance of the proposed PLMS feature and the corresponding model (denoted as Ours in Table 5 ) through extensive comparative experiments. Under the PLMS feature representation, our model achieves the highest performance, attaining an Accuracy@1 of 92.4%. Additionally, it achieves a Macro-AUC of 99.70%, significantly surpassing the performance of other models. This underscores the strong discriminative capability of the PLMS feature, which is further leveraged by our model through tailored network design. The superiority of the PLMS feature becomes even more apparent when compared to other feature representations. For instance, within our model, the Accuracy@1 achieved with the PLMS feature is 92.42%, which is 1.61%, 4.34%, 8.5%, 10.13% and 48.97% higher than those achieved with the S-Transform, STFT, wavelet-based feature, Wigner-Ville and LEAF, respectively, with accuracies of 90.96%, 88.58%, 83.92%, 84.92% and 62.04%. This performance improvement can be attributed to the ability of the PLMS feature to integrate the temporal dynamics of pest acoustic signals with multi-scale frequency domain representations, enabling a more comprehensive characterization of audio semantic information. In contrast, baseline features such as S-Transform, STFT, wavelet, and Wigner-Ville predominantly focus on either time-domain or frequency-domain information in isolation. This singular focus limits their capacity to fully capture the complex and multi-faceted nature of audio signals, thereby constraining their generalization capabilities. For instance, S-Transform and STFT primarily emphasize frequency-domain information, while wavelet and Wigner-Ville focus more on time-frequency representations but may not capture the multi-scale characteristics as effectively as the PLMS feature. Although LEAF employs a learnable frontend to adaptively optimize time-frequency representations and enhance robustness, it may still be less effective than PLMS at fully capturing temporal dynamics across multiple scales.

From the perspective of model comparison, our model demonstrates a comprehensive advantage under the PLMS feature, achieving an Accuracy@1 of 92.42%, a Macro-F1 score of 92.41%, and a Macro-AUC of 99.70%, significantly outperforming other models. In contrast, ResNet18, although yielding the second-best performance under the PLMS feature with an Accuracy@1 of 87.59%, still trails our model by 4.83%. While EfficientNet-b0 benefits from a compound scaling strategy that enhances computational efficiency, its Accuracy@1 is 3.97% lower than that of our model, highlighting the superior capability of our model to capture long-range temporal dependencies and integrate cross-scale features.

It is noteworthy that traditional deep networks exhibit significant sensitivity to feature representations. For example, VGG19 achieves an Accuracy@1 of only 8.33% and a Macro-F1 of 1.28% under the PLMS feature. This poor performance is largely attributed to its fixed receptive field and redundant parameter design, which are ill-suited for capturing dynamic audio features. DenseNet achieves an Accuracy@1 of only 9.22% under the S-Transform feature, reflecting the limitations of its dense connectivity mechanism in capturing frequency-domain discontinuities. In contrast, MobileNet performs reasonably well with the PLMS feature, achieving an Accuracy@1 of 84.26%. However, its lightweight design compromises its ability to extract deep semantic features, as evidenced by its Accuracy@1 of only 60.97% under the S-Transform feature, considerably lower than the 90.96% achieved by our model. This suggests that, although depthwise separable convolutions are employed in MobileNet to reduce computational overhead, the model’s capacity to capture complex acoustic patterns remains limited. Collectively, these findings underscore the importance of the coordinated optimization of the PLMS feature representation and the architectural design of our model as the key determinant of performance improvements.

As presented in Table 6 , the number of parameters (in millions) and the computational complexity in GFLOPS of various models under the PLMS feature are systematically compared. From the perspective of parameter quantity, VGG19 has the highest number of parameters, reaching 139.62 million. This substantial parameter volume is a direct consequence of its deep network architecture, making it well-suited for high-precision applications that are less constrained by computational resources. In contrast, ResNet18 contains 11.23 million parameters, while DenseNet has 6.85 million. ResNet18 balances depth and efficiency through residual connections, whereas DenseNet enhances feature reuse via densely connected layers, though at the expense of a larger number of parameters compared to lightweight models. EfficientNet-b0 further reduces parameters to 4.02 million by optimizing parameter utilization through a compound scaling strategy. MobileNet compresses parameters even further to 2.24 million by employing depthwise separable convolutions, showcasing the advantages of lightweight network design. Notably, our model achieves the most aggressive parameter reduction, with only 1.54 million parameters, even surpassing MobileNet. This result underscores its structural innovations in pre-trained model design and fine-tuning strategies.

Regarding computational complexity, VGG19 incurs the highest computational cost, with a GFLOPS of 1.04, primarily due to its extensive fully connected layers and deep convolutional operations. ResNet18 and DenseNet maintain moderate computational complexity, with GFLOPS values of 0.07 and 0.11, respectively. While their computational demands are consistent with their parameter volumes, DenseNet incurs additional computational overhead due to its dense connectivity pattern. EfficientNet-b0 significantly reduces computational demands to 0.02 through a well-balanced scaling mechanism. MobileNet and our model achieve the lowest computational complexity, with a GFLOPS of 0.01, highlighting their superior efficiency. This characteristic makes them particularly suitable for real-time processing and low-power environments. In summary, our model emerges as the most lightweight architecture in the comparison, with both the lowest parameter count of 1.54 million and the lowest GFLOPS of 0.01. These results emphasize its advantages in model compression and computational optimization, rendering it highly applicable to scenarios with stringent resource constraints.

To ensure rigorous benchmarking, the experimental parameters SegmentFrameNumber Num is fixed at 10 frames per audio segment. The original sampling rate is 16KHz. As shown in Table 7 , we compare classification performance across different sampling rates and model combinations based on the PLMS feature. On one hand, a comprehensive analysis of the dynamic correlation between model performance and sampling rate reveals significant variations among different architectures. Under the PLMS feature, our model exhibits optimal performance at 2,500 Hz, achieving an Accuracy@1 of 96.49%, a Macro-F1 score of 96.49%, and a Macro-AUC of 99.93%. In contrast, VGG19 demonstrates extreme instability at higher sampling rates, particularly at 16 kHz, where its Accuracy@1 plunges to 8.33%, and Macro-F1 drops to 1.28%. This can be attributed to the densely connected fully connected layers in VGG19, which amplify sensitivity to high-resolution noise artifacts. Consequently, the model tends to overfit specific frequency bands, leading to a pronounced degradation in generalization performance. While ResNet18 and EfficientNet-b0 exhibit relatively stable performance across different sampling rates, their performance still shows a slight degradation as the sampling rate decreases. ResNet18 achieves peak performance at 1,500 Hz, attaining an Accuracy@1 of 93.78%, while EfficientNet-b0 performs optimally at 3,000 Hz and 4,000 Hz, reaching an Accuracy@1 of 93.32%. These differences arise from the architectural characteristics of ResNet18 and EfficientNet-b0. ResNet18 utilizes residual connections to effectively mitigate gradient vanishing, enhancing its adaptability to multi-sampling-rate features. In contrast, EfficientNet-b0 employs neural architecture search to optimize the balance between depth, width, and resolution, thereby improving computational efficiency across various sampling rates.

In contrast, lightweight models such as MobileNet and DenseNet perform reasonably well at mid-to-low sampling rates but degrade significantly at the original high sampling rate. MobileNet achieves its best Accuracy@1 (91.67%) at 2,000 Hz, while DenseNet peaks at 2,500 Hz with an Accuracy@1 of 92.13%. However, at 16 kHz, the Accuracy@1 of MobileNet drops to 84.26%, while DenseNet declines sharply to 77.34%. This performance gap stems from their distinct architectural constraints. MobileNet employs depthwise separable convolutions, which effectively reduce the number of parameters but also weaken cross-channel dependencies, resulting in suboptimal performance on high-resolution data. In contrast, DenseNet faces increased computational complexity at high sampling rates, where excessive redundant features introduce gradient noise accumulation during backpropagation, ultimately impairing model convergence stability. Notably, despite being designed as a lightweight model, our model demonstrates remarkable cross-sampling-rate stability. By integrating structural innovations, it effectively compensates for the information loss induced by downsampling, thereby enhancing classification robustness under low-sampling-rate conditions. This suggests that lightweight architectures, when appropriately designed, can mitigate performance degradation across varying sampling rates.

One the other hand, examining metric consistency and model reliability further highlights the impact of sampling rates. All models exhibit Macro-AUC values consistently exceeding 99%, indicating strong class discrimination capability. However, discrepancies between Macro-F1 and Accuracy@1 reveal subtle classification biases. For instance, VGG19 achieves a Macro-AUC of 99.29% at 2,000 Hz, yet its Macro-F1 remains lower at 91.00%, suggesting potential class imbalance issues or suboptimal recognition of minority classes. In contrast, our model maintains exceptional consistency across all sampling rates, with Accuracy@1, Macro-F1, and Macro-AUC remaining highly aligned. At 2,500 Hz, it achieves an Accuracy@1 of 96.49%, a Macro-F1 of 96.49%, and a Macro-AUC of 99.93%, confirming its balanced classification capability and low-variance characteristics. These findings underscore that sampling rate has a profound impact on model performance, necessitating a delicate balance between lightweight design and data adaptability. The results suggest that while certain architectures, such as VGG19, struggle with high-resolution noise sensitivity, others, like ResNet18 and EfficientNet-b0, exhibit better adaptability to varying sampling rates. More importantly, our model, through architectural innovation, achieves robust performance at lower sampling rates, making it a highly efficient and practical solution for real-world deployment, particularly in resource-constrained environments where computational efficiency and robustness are crucial.

To rigorously analyze the impact of varying patch sizes on the classification performance of different models based on the PLMS, the downsampling rate is set to 2500Hz, as indicated by the results in Table 7 . Subsequently, as delineated in Table 8 , we conducted a comparative analysis across four patch sizes. Experimental results demonstrate that patch size has a significant impact on model performance. The proposed model achieves optimal performance when the patch size is configured to a temporal window encompassing 10 successive frames, attaining an Accuracy@1 of 96.49%, a Macro-F1 and Macro-Recall of 96.49%, and a Macro-AUC as high as 99.93%, significantly outperforming the other four models. This finding suggests that the patch size of 10 frames effectively balances local detail and global contextual information. In contrast, a patch size of 5 frames may lack sufficient temporal correlation, leading to incomplete feature representations, while patch sizes of 15 or 20 frames could introduce redundant noise, diminishing the model’s sensitivity to informative signals.

From the perspective of model-specific characteristics, ResNet18 attains an Accuracy@1 of 92.86% when the patch size is 10 but experiences a decline to 91.34% at a patch size of 20. This decrease arises from interference due to redundant information, underscoring the residual network’s proclivity for localized temporal contexts. EfficientNet-b0 maintains relatively stable Accuracy@1 across different patch sizes (ranging from 91.01% to 92.72%). However, when the patch size is 15, it exhibits an anomalous drop in Macro-AUC to 94.48%, which results in spectral confusion among certain categories within the PLMS features. VGG19 exhibits minimal sensitivity to patch size variations, maintaining an Accuracy@1 between 90.34% and 90.61%, as its fixed receptive field and extensive parameterization limit its capacity for capturing dynamic temporal features. MobileNet performs the worst when a patch size is 10 (an Accuracy@1 reaching only 86.51%) but recovers to 89.04% at a patch size of 15. This suggests that its lightweight architecture struggles with modeling medium-length sequences, while longer sequences partially mitigate this limitation through increased information density. DenseNet achieves a peak Accuracy@1 of 92.26% at a patch size of 15 but drops to 89.35% when the patch size reaches 20, likely due to gradient redundancy or noise propagation within its densely connected structure under long-sequence conditions.

In terms of classification metric consistency, the Macro-Recall and Macro-F1 values of all models are highly similar, indicating good inter-class balance in classification results. However, MobileNet exhibits a slight discrepancy at a patch size of 20, with a Macro-Recall of 89.15%, marginally lower than its Macro-F1 score of 89.36%. This suggests reduced sensitivity to certain low-frequency or low-amplitude categories. Notably, EfficientNet-b0 exhibits an anomalous performance at a patch size of 15, with its Macro-AUC decreasing sharply to 94.48%, significantly deviating from the consistently higher Macro-AUC values observed at other patch sizes. This decline suggests that, at this specific patch size, spectral feature confusion occurs among certain categories within longer sequences, thereby impairing the model’s discriminative capabilities. Such observations highlight the necessity of jointly modeling local and global features in temporal signal processing to effectively capture both detailed and contextual information.

Figure 8 shows the confusion matrices of 12 insect species for the six models based on the PLMS feature, with a patch size of 20 and a downsampling rate of 2,500 Hz. The horizontal axis represents the predicted labels, while the vertical axis represents the true labels, with 0–11 corresponding to the insect species listed in Table 3 above. It is worth mentioning that our model exhibits remarkable superiority, as evidenced by its confusion matrix. Except for the 4^th and 8^th classes, where the diagonal elements are relatively low, all other classes have diagonal elements approaching 1.0, indicating exceptionally high classification accuracy. Moreover, the off-diagonal elements are minimal, signifying an exceedingly low misclassification rate. In contrast, models such as ResNet18, EfficientNet-B0, VGG19, MobileNet, and DenseNet, while achieving high accuracy in certain categories, exhibit significant misclassifications in specific classes. For instance, ResNet18 shows higher off-diagonal values for the 6^th, 8^th, 9^th, and 11^th classes; EfficientNet-B0 for the 4^th, 5^th, 8^th, and 11^th classes; VGG19 for the 4^th, 5^th, 8^th, 10^th, and 11^th classes; and DenseNet for the 5^th, 8^th, and 11^th classes. Notably, MobileNet exhibits the diagonal values of approximately 0.6 for the 4^th and 11^th classes.

Focusing on our experimental results, the confusion matrix reveals significant misclassification issues between the 4^th and 8^th classes. Samples of the 4^th class are primarily misclassified as the 5^th and 11^th classes, each accounting for 3%, while samples of the 8^th class are misclassified as the 4^th and 5^th classes, each accounting for 2%, and as the 11^th class, accounting for 3%. Since the data are collected under controlled conditions with minimal background noise, and based on our careful inspection of the spectrograms, the spectral features of the 4^th, 5^th, 8^th, and 11^th classes exhibit minimal differences. Their frequency distributions and energy concentration regions highly overlap, making traditional spectrogram-based features insufficient to effectively distinguish these classes. Additionally, although the training data quality is high, the diversity of sound samples may be inadequate to capture subtle acoustic variations present in real-world conditions, which limits the model’s discriminative power.

To address these challenges, we propose several improvements. First, incorporate richer and more discriminative acoustic features such as transient signal characteristics or nonlinear dynamic features. Second, exploring additional data augmentation techniques, such as pitch shifting and the introduction of simulated environmental effects, may further enhance the diversity of insect acoustic samples and improve the model’s generalization ability. Finally, consider multimodal fusion approaches by integrating additional sensory data such as insect vibration signals and behavioral patterns to better differentiate similar classes and reduce misclassification rates between the 4^th and 8^th classes.

Figure 9 shows the validation accuracy and loss progress of all the used six models during the training; each line consists of 150 points, one for each epoch. As illustrated in Figure 9 , our model rapidly reduces the loss value during the early training stages and achieves high validation accuracy within a relatively small number of iterations, maintaining stability thereafter. This demonstrates fast convergence and strong generalization capability. In comparison, although VGG19 and EfficientNetB0 demonstrate excellent convergence properties, with the loss value of VGG19 approaching zero, their final accuracy remains lower than that achieved by our model. ResNet18 also converges relatively quickly in terms of both loss and accuracy during training, yet its final accuracy still falls short. DenseNet and MobileNet, however, exhibit a slower loss decline and a more gradual accuracy increase during the initial stages of training, followed by noticeable fluctuations in the later phases. Such fluctuations may arise from their network architectures and parameter complexities, which could impede stable training. In contrast, our model excels in training efficiency, accuracy, and stability, clearly demonstrating its advantages.

As presented in Sections 6.1 to 6.5, a fixed 8:2 dataset split was employed to conduct a baseline performance comparison between our method and representative CNN architectures. While this setting offers a consistent benchmark, it may be influenced by the specific train–test partition. To further evaluate model robustness and generalization across diverse data splits, we subsequently performed a five-fold cross-validation experiment, as shown in Table 9 . In this setting, ablation studies were performed on the proposed method, including its original version, a variant without (w/o) spectral augmentation (SpecAug), and a variant employing patchout, a regularization method that randomly discards a portion of input patches to enhance generalization. These ablation experiments were designed to isolate and quantify the contribution of each component within our framework. In parallel, we carried out comparative experiments with other algorithms, specifically EfficientNet Lite (Sangar and Rajasekar, 2025) and MobileViT (Gu et al., 2024), in order to benchmark our method against lightweight architectures proposed in recent literature. Together, these evaluations provide a more comprehensive and rigorous assessment of both accuracy and stability under varying training conditions.

The results shown in Figure 9 indicate that, regarding the fixed 8:2 dataset partition, our model attains the highest performance across all evaluation metrics. In contrast, the variants without SpecAug or incorporating Patchout show a marginal performance decline but still substantially outperform the baseline methods reported in references (Sangar and Rajasekar, 2025) and (Gu et al., 2024). Notably, the method in (Gu et al., 2024) exhibits significantly inferior results, with all metrics falling well below those of our model. From the perspective of five-fold cross-validation, the robustness of our model is further confirmed, consistently leading with an accuracy of 95.23% and a minimal standard deviation of 0.17, underscoring its strong generalization capability. The variant without SpecAug and the one incorporating Patchout maintain stable performance, although with slightly reduced scores. Conversely, the methods from (Sangar and Rajasekar, 2025) and (Gu et al., 2024) demonstrate comparatively poorer and more variable outcomes, particularly (Gu et al., 2024), which shows a markedly high standard deviation, indicating less stable performance. In summary, the proposed approach not only achieves superior accuracy on the fixed dataset partition, but also demonstrates exceptional stability and generalization under the more rigorous five-fold cross-validation, thereby validating the effectiveness of the model architecture and data augmentation strategies employed.