View : 109 Download: 0
Automated Acoustic Detection of Calling Activity in the Ueno’s Brown Frog (Rana uenoi) for Breeding Phenology Monitoring
- Title
- Automated Acoustic Detection of Calling Activity in the Ueno’s Brown Frog (Rana uenoi) for Breeding Phenology Monitoring
- Authors
- 손도연
- Issue Date
- 2026
- Department/Major
- 대학원 에코과학부
- Keywords
- Ueno’s brown frog, Rana uenoi, passive acoustic monitoring, PC-NMF, breeding phenology, automated acoustic detection
- Publisher
- 이화여자대학교 대학원
- Degree
- Master
- Advisors
- 장이권
- Abstract
- Amphibian breeding phenology is highly sensitive to environmental changes such as temperature, precipitation, and hydrological conditions. Long-term monitoring of breeding phenology is therefore important for understanding population dynamics and the effects of climate change. However, field surveys alone may not accurately identify the actual first date of breeding activity because of limitations related to survey intervals and weather conditions. This study developed an automated acoustic detection model for the advertisement calls of Ueno’s brown frog (Rana uenoi) and used it to estimate breeding-season calling onset, which was then compared with manual call onset and egg onset.
Automated detection was conducted using the PC-NMF-based soundscape_IR model. Acoustic data collected in 2025 were used for model validation and for establishing post-processing criteria, whereas the 2026 data were used as an independent application dataset. In the 2025 validation, event-level performance was evaluated based on temporal overlap between manual positive selections and model detections. Exclude-only detections were omitted from the evaluation, and precision, recall, and F1 were calculated using a many-to-many overlap coverage approach.
The frog-only model showed relatively high recall and F1 in the 2025 event-level validation. However, in the 2026 onset estimation, it generated false-positive days before the manual call onset at most sites. Therefore, a corrected noise-merged model was constructed by incorporating non-target sound sources, including bird, wind, rain, and dog sounds. Although this model showed lower event-level recall and F1 in the 2025 validation, it was more suitable for reducing pre-onset false positives in the 2026 site-date-level acoustic onset estimation.
An SNR cutoff was initially examined as a post-processing criterion, but it was excluded from the final criteria because it substantially reduced recall and F1. Instead, detection duration and spectral concentration were used. Spectral concentration was calculated as the proportion of energy concentrated within ±150 Hz of the peak frequency relative to the total energy in the 500-2000 Hz band. The final criteria consisted of the corrected noise-merged model, a threshold of 10.0, duration ≤1.0 s, spectral concentration ≥0.30, and site-date-level criteria of n_detections ≥10 and total_detection_duration ≥5.0 s.
Using these criteria, the estimated acoustic onset dates at seven of the eight sites fell between manual call onset and egg onset. At YS, acoustic onset was not detected under the conservative criteria because detection intensity was low. This study demonstrates that passive acoustic monitoring can be used to estimate early breeding-season calling activity of R. uenoi at the site date level. It also shows that the performance of automated detection models should be interpreted separately for individual call-event detection and phenology-level onset estimation.;양서류의 번식 시기는 기온, 강수, 수온 등 환경 변화에 민감하게 반응하므로, 번식 생물계절을 장기적으로 모니터링하는 것은 개체군 변동과 기후변화 영향을 이해하는 데 중요하다. 그러나 현장조사만으로는 조사 간격과 기상 조건의 제한으로 실제 첫 번식 활동일을 정확히 파악하기 어렵다. 본 연구는 큰산개구리(Rana uenoi)의 구애 노래를 대상으로 자동 음향 탐지 모델을 구축하고, 이를 이용해 번식기 calling onset 을 추정한 뒤 manual call onset 및 egg onset 과 비교하였다.
자동 탐지는 PC-NMF 기반 soundscape_IR 모델을 이용하여 수행하였다. 2025 년 음향 자료는 모델 검증과 후처리 기준 설정에 사용하였고, 2026년 자료는 독립 적용 자료로 사용하였다. 2025 년 validation 에서는 manual positive selection 과 model detection 의 시간적 중첩을 기준으로 event-level 성능을 평가하였으며, exclude-only detection 은 평가에서 제외하고 many-to-many overlap coverage 방식으로 precision, recall, F1 을 산출하였다.
Frog-only model 은 2025년 validation에서 상대적으로 높은 recall과 F1을 보였으나, 2026 년 onset 추정에서는 대부분의 지점에서 manual call onset 이전의 false-positive day 를 생성하였다. 이에 따라 bird, wind, rain, dog 등 비표적 음원 source 를 병합한 corrected noise-merged model 을 구성하였다. 이 모델은 2025 년 event-level recall 과 F1 은 낮아졌지만, 2026 년 site-date 단위 acoustic onset 추정에서는 onset 이전 false positive 를 줄이는 데 더 적합하였다.
후처리 기준으로 SNR cutoff 를 검토하였으나, SNR cutoff 는 recall 과 F1 을 크게 감소시켜 최종 기준에서 제외하였다. 대신 detection duration과 spectral concentration을 이용하였다. Spectral concentration 은 500–2000 Hz 대역 전체 에너지 중 peak frequency ±150 Hz 범위에 집중된 에너지 비율로 계산하였다. 최종 기준은 corrected noise-merged model, threshold 10.0, duration ≤1.0 s, spectral concentration ≥0.30, 그리고 site-date 단위 n_detections ≥10 및 total_detection_duration ≥5.0 s 로 설정하였다.
이 기준을 적용한 결과, 8 개 지점 중 7 개 지점에서 추정된 acoustic onset date 가 manual call onset 과 egg onset 사이에 포함되었다. YS 는 detection intensity 가 낮아 보수적 기준에서 미검출로 남았다. 본 연구는 passive acoustic monitoring 을 이용해 큰산개구리의 번식 초기 calling activity 를 site-date 단위로 추정할 수 있음을 보여주었다. 또한 자동 탐지 모델의 성능 평가는 개별 call event 탐지와 phenology-level onset 추정을 구분하여 해석해야 함을 확인하였다.
- Fulltext
- Show the fulltext
- Appears in Collections:
- 일반대학원 > 에코과학부 > Theses_Master
- Files in This Item:
There are no files associated with this item.
- Export
- RIS (EndNote)
- XLS (Excel)
- XML