Acoustic indices have been increasingly explored as possible proxies for biodiversity, yet in complex environments non-avian sounds often mask bird vocalizations, compromising assessment accuracy and ecological representativeness. To address this, we developed an integrated framework that combines deep learning-based soundscape classification with a threshold-optimized strategy for removing non-avian recordings. A multi-label model was trained to identify 12 representative soundscape categories. Non-avian recordings were then removed using threshold criteria optimized through systematic evaluation of threshold combinations, maximizing correlations between six acoustic indices [acoustic complexity index (ACI), acoustic diversity index (ADI), acoustic evenness index (AEI), bioacoustic index (BIO), normalized difference soundscape index (NDSI), acoustic entropy index (H)] and bird vocal activity while retaining as much useable recording data as possible to avoid excessive reduction of sample size. To validate the effectiveness of our framework, generalized additive models and random forest regressions were used to compare diel patterns and predict bird vocal activity before and after removal. Using data from 19 passive acoustic recorders in the Central Green Forest Park, Beijing, we found insect sounds exerted the strongest masking effect on relationships between indices and bird vocal activity. Applying optimized removal improved temporal alignment of indices with bird vocal activity and enhanced predictive performance. This study demonstrates that threshold-guided removal of non-avian recordings strengthen the ecological interpretability and predictive utility of acoustic indices in biodiversity monitoring.

Passive acoustic monitoring (PAM) technology is increasingly becoming one of the mainstream methods for bird monitoring. However, detecting bird audio within complex natural acoustic environments using PAM devices remains a significant challenge. To enhance the accuracy (ACC) of bird audio detection (BAD) and reduce both false negatives and false positives, this study proposes a BAD method based on a Dual-Feature Enhancement Fusion Model (DFEFM). This method incorporates per-channel energy normalization (PCEN) to suppress noise in the input audio and utilizes mel-frequency cepstral coefficients (MFCC) and frequency correlation matrices (FCM) as input features. It achieves deep feature-level fusion of MFCC and FCM on the channel dimension through two independent multi-layer convolutional network branches, and further integrates Spatial and Channel Synergistic Attention (SCSA) and Multi-Head Attention (MHA) modules to enhance the fusion effect of the aforementioned two deep features. Experimental results on the DCASE2018 BAD dataset show that our proposed method achieved an ACC of 91.4% and an AUC value of 0.963, with false negative and false positive rates of 11.36% and 7.40%, respectively, surpassing existing methods. The method also demonstrated detection ACC above 92% and AUC values above 0.987 on datasets from three sites of different natural scenes in Beijing. Testing on the NVIDIA Jetson Nano indicated that the method achieved an ACC of 89.48% when processing an average of 10 s of audio, with a response time of only 0.557 s, showing excellent processing efficiency. This study provides an effective method for filtering non-bird vocalization audio in bird vocalization monitoring devices, which helps to save edge storage and information transmission costs, and has significant application value for wild bird monitoring and ecological research.