논문 상세보기
Perch2 임베딩 기반 개구리 음향 분류와 임베딩 변동성 기반 신뢰도 진단 KCI 등재
Frog Call Classification Using Perch2 Embeddings and Variability-Based Confidence
- 언어KOR
- URLhttps://db.koreascholar.com/Article/Detail/451778
pp.78-88
Frog call classification in field recordings is difficult because vocalizations vary with background noise, recording distance, and calling context. This study presents an embedding-based transfer-learning framework for practical species identification using Perch2 as a pretrained bioacoustic feature extractor. Recordings from seven frog species (69 files) were scanned with overlapping 5-s windows, band-pass filtered between 150 and 6000 Hz, and ranked by RMS energy; up to eight high-energy windows per file were retained, yielding 441 clips. Each clip was converted into a 1536-dimensional Perch2 embedding, standardized, and classified with L2-regularized multinomial logistic regression using five-fold grouped cross-validation in which clips from the same recording were assigned to the same fold. The model achieved an overall out-of-fold accuracy of 0.868 at the clip level and 0.884 at the file level. Fold-level accuracy was 0.873±0.094 for clips and 0.885±0.081 for files. The logistic-regression model consistently outperformed an XGBoost baseline trained on the same embeddings, although paired fold-level tests were interpreted cautiously because only five folds were available. Embedding-space diagnostics showed statistically non-random species structure, but low silhouette values indicated partial rather than complete species separation. Within-file embedding variability showed a weak but statistically supported negative association with file-level confidence (Spearman ρ= - 0.299, p=0.013). These results support Perch2 embeddings with a simple regularized linear classifier as a practical, interpretable approach for amphibian acoustic classification in data-limited settings, while indicating that confidence scores should be used as review-prioritization aids rather than calibrated occurrence probabilities.
서 론
재료 및 방 법
1. 음향 자료 수집과 구성
2. 음향 파일 전처리와 클립 구성
3. Perch2 임베딩 추출
4. 학습 데이터 구성과 가중치 설계
5. 분류기 학습
6. 평가 설계와 지표
결 과
1. 자료 구성과 평가 설정 및 분류성능
2. 종별 혼동과 임베딩 공간 분리도
3. 파일 내 임베딩 변동성과 예측 신뢰도
고 찰
1. 임베딩 기반 전이학습 분류기의 성능과 실용적 의미
2. 오분류 구조와 파일 내 이질성에 기반한 불확실성 해석
3. 한계와 향후 연구 과제
적 요
REFERENCES
