In cooperation with Endemic Species Research Institute (ESRI) and Yushan National Park Headquarters (YSNP Headquarters), a long-term passive acoustic monitoring (PAM) system in YSNP was built to monitor animals and ecosystems from soundscape recordings in year 2020. The data were collected by six PAM stations along the Southern Cross-Island Highway in the Southern Area of YSNP. The stations were maintained by YSNP Headquarters and the data were archived, managed, analyzed and open-accessed by ESRI. The recorders were programmed to record a minute every three minutes. In total, these are 1,776,492 1-minute recordings were collected between 2020-01-20 and 2021-12-31. Animal vocalizations in recordings were detected automatically with the use of the SILIC (Sound Identification and Labeling Intelligence for Creatures) which was developed by ESRI. We used the model exp24 of the SILIC (https://github.com/RedbirdTaiwan/silic) to detect seven sound classes of seven bird species and then 150 samples of each sound class were selected randomly and reviewed manually for evaluating the precision and recall. The detections totalled 6,243,820 for all the seven sound classes, for reducing the data size, the dataset has been summarized to the count of detections per sound class per recording and resulted in 802,670 records in total.
為進行玉山國家公園長期生態監測，至2020年起，行政院農業委員會特有生物研究保育中心（ESRI）與內政部營建署玉山國家公園管理處（YSNP）合作，於園區內架設被動式聲學監測站（PAM站），長期錄製聲景資料，進行聲景生態監測。本資料集收錄設於YSNP南部園區，位於南部橫貫公路沿線的6處PAM站資料，PAM站由YSNP進行定期維運，聲景資料由TESRI典藏、管理、分析及開放。PAM站為定時錄音，設置參數為每隔2分鐘錄製1分鐘，總計有1,776,492筆1分鐘音檔在2020-01-20至2021-12-31間被錄製。錄製之聲景資料利用ESRI開發的生物音智慧辨識與標記系統SILIC exp24模型進行分析 (https://github.com/RedbirdTaiwan/silic) ，目前已完成7種鳥類聲音之資料分析，每種聲音隨機取150筆資料進行人工複核以估算精確率（Precision）與召回率（Recall）。總計本資料集收錄7種鳥類合計叫聲6,243,820聲，為減少資料量，將同錄音檔且同聲音編號之聲音進行數量加總，經加總後之總資料筆數為802,670筆。
The data in this sampling event resource has been published as a Darwin Core Archive (DwC-A), which is a standardized format for sharing biodiversity data as a set of one or more data tables. The core data table contains 439,275 records.
1 extension data tables also exist. An extension record supplies extra information about a core record. The number of records in each extension data table is illustrated below.
This IPT archives the data and thus serves as the data repository. The data and resource metadata are available for download in the downloads section. The versions table lists other versions of the resource that have been made publicly available and allows tracking changes made to the resource over time.
The table below shows only published versions of the resource that are publicly accessible.
How to cite
Researchers should cite this work as follows:
Wu S, Tsai W, Ko J C (2022): Acoustic detections of birds using the SILIC in Yushan National Park, Taiwan. v1.5. Taiwan Endemic Species Research Institute. Dataset/Samplingevent. https://ipt.taibif.tw/resource?r=silic-ysnp&v=1.5
Researchers should respect the following rights statement:
The publisher and rights holder of this work is Taiwan Endemic Species Research Institute. This work is licensed under a Creative Commons Attribution (CC-BY) 4.0 License.
This resource has been registered with GBIF, and assigned the following GBIF UUID: 4410edca-3bdd-4475-98a2-de823b2266bc. Taiwan Endemic Species Research Institute publishes this resource, and is itself registered in GBIF as a data publisher endorsed by Taiwan Biodiversity Information Facility.
Samplingevent; Taiwan; vertebrates; SILIC; Yushan National Park; Passive Acoustic Monitoring; Automated wildlife sound identification; 脊椎動物; 玉山國家公園; 被動式聲學監測; 野生動物聲音自動辨識
Six PAM stations were deployed between Meishan and Yako along the Southern Cross-Island Highway, with an elevation range from 1,264 m above sea level (MSC01) to 2,739 m (WK01). 6座被動式聲學監測站設置於玉山國家公園南部園區，南部橫貫公路梅山至埡口沿線。6個樣站係沿南部橫貫公路架設，海拔範圍介於1,264至2,739 公尺。
|Bounding Coordinates||South West [23.257, 120.826], North East [23.288, 120.955]|
The taxonomic coverage would increase with the version and precision of the SILIC which is used to detect vocalizations of vertebrates automatically in the soundscape recordings. 本資料集係使用SILIC自動偵測聲景音檔中之脊椎動物鳴叫聲，涵蓋物種數將隨SILIC的可偵測物種數及精確率不定期增加。
|Species||Liocichla steerii (黃胸藪眉), Psilopogon nuchalis (五色鳥), Corvus macrorhynchos (巨嘴鴉), Pericrocotus solaris (灰喉山椒), Heterophasia auricularis (白耳畫眉), Yuhina brunneiceps (冠羽畫眉), Myiomela leucura (白尾鴝)|
|Start Date / End Date||2020-01-20 / 2021-12-31|
The aim of Sound Identification and Labeling Intelligence for Creatures (SILIC) is to integrate online animal sound databases, PAM databases and an object detection-based model, for extracting information on the sounds of multiple species from complex soundscape recordings. SILIC計畫目的為自動化擷取聲景資料中的動物訊息，以作為生態研究及棲地經營管理之用。
|Title||SILIC - Sound Identification and Labeling Intelligence for Creatures|
The personnel involved in the project:
The Song Meter SM4 and Song Meter Mini made by Wildlife Acoustic Inc. were used as the autonomous recording units (ARUs) in this dataset. The ARUs were scheduled to record a 1-min recording every three minutes in stereo, 16-bit WAV format at a sampling rate of 44.1 kHz. The SILIC was used to detect animal vocalizations automatically in the recordings. The threshold of a confidence score was chose once the precision reached to 0.95 for each sound class. All the detections of each sound class with a confidence score no less than the threshold were treated as positives and archived in this dataset. 被動式聲學監測站為定時錄音，使用之錄音機為Wildlife Acoustic公司製造之Song Meter SM4及Song Meter Mini，設置參數為每錄製1分鐘休息2分鐘、雙聲道、取樣頻率44.1 KHz、位元深度16 bit，錄音資料儲存為PCM WAV檔格式。利用SILIC自動偵測音檔中之動物鳴叫聲，取精確率（Precision）為0.95的信賴分數（Confidence Score）作為閾值（Threshold），分數大或等於閾值的偵測結果視為陽性（positive，即認定該筆資料為目標聲音），陽性偵測結果彙整為本資料集。
|Study Extent||The study area was located in the Southern Area of YSNP, a typical montane ecosystem in central Taiwan. Six PAM stations were deployed between Meishan and Yako along the Southern Cross-Island Highway, with an elevation range from 1,264 m above sea level (MSC01) to 2,739 m (WK01). The longest distance between any two stations was around 11.4 km and the shortest distance was 500 m. The habitat types vary from lower (1,264 m) to higher (2,739 m) elevation, including sub-montane evergreen broad-leaved forests, montane evergreen broad-leaved cloud forests, montane mixed cloud forests, and upper-montane coniferous forests. 被動式聲學監測站設置於玉山國家公園南部園區，南部橫貫公路梅山至埡口沿線。玉山國家公園位於臺灣本島之中央地帶，面積廣達103,121公頃，地形以高山及河谷為主，3,000 公尺以上之地區約占全區面積 12.74 %，列名臺灣「百岳」之山峰共有 30 座，其中包括百岳之首的臺灣最高峰—玉山。目前已完成的6個樣站係沿南部橫貫公路架設，分別為梅山（MSC01，海拔1,264公尺）、中之關（ZZG01，海拔2,047公尺）、天池(下)（TT01，海拔2,303公尺）、天池(上)（TT02，海拔2,366公尺）、檜谷（KK01，海拔2,429公尺）及埡口林道（WK01，海拔2,739公尺），設置時間自2020起，預期進行無限期之長期監測。|
|Quality Control||The model version exp24 of the SILIC was used to detect animal vocalizations in this dataset (https://github.com/RedbirdTaiwan/silic). Seven sound classes were selected as follows and 150 random clips of each sound class were sampled to evaluate the performance including the precision and the recall. The threshold of a confidence score was chose once the precision reached to 0.95 for each sound class. All the detections of each sound class with a confidence score no less than the threshold were treated as positives. 本資料集使用SILIC exp24模型進行分析（ https://github.com/RedbirdTaiwan/silic ），目前已完成7種聲音之偵測，每種聲音隨機取150筆偵測結果進行精確率（Precision）與召回率（Recall）估算，取精確率（Precision）為0.95的信賴分數（Confidence Score）作為閾值（Threshold），分數大或等於閾值的偵測結果視為陽性（positive，即認定該筆資料為目標聲音）。 9 - 白耳畫眉 (Heterophasia auricularis) S-01: Threshold=0.54、Precision=0.95、Recall=0.53。 28 - 五色鳥 (Psilopogon nuchalis) S-01: Threshold=0.26、Precision=0.95、Recall=0.80。 122 - 黃胸藪眉 (Liocichla steerii) S-01: Threshold=0.73、Precision=0.95、Recall=0.48。 324 - 冠羽畫眉 (Yuhina brunneiceps) S-01: Threshold=0.71、Precision=0.95、Recall=0.55。 337 - 灰喉山椒 (Pericrocotus solaris) U-01: Threshold=0.57、Precision=0.95、Recall=0.72。 361 - 白尾鴝 (Myiomela leucura) S-01: Threshold=0.51、Precision=0.95、Recall=0.68。 471 - 巨嘴鴉 (Corvus macrorhynchos) C-01: Threshold=0.48、Precision=0.95、Recall=0.64。|
Method step description:
- The Song Meter SM4 and Song Meter Mini made by Wildlife Acoustic Inc. were used as the autonomous recording units (ARUs) in the project. The ARUs were tied on trees around 1.5 m above the grounds and covered with sound-absorbing canopies. All ARUs were scheduled to record a one-min recording every three minutes in stereo, 16-bit WAV format at a sampling rate of 44.1 kHz. 被動式聲學監測站為定時錄音，使用之錄音機為Wildlife Acoustic公司製造之Song Meter SM4及Song Meter Mini，設置參數為每錄製1分鐘休息2分鐘、雙聲道、取樣頻率44.1 KHz、位元深度16 bit，錄音資料儲存為PCM WAV檔格式。
- Memory cards storing acoustic data were replaced monthly and two copies of files were archived separately in local storages and Google Drive for data safety. 每月更換記憶卡及電池，錄音資料存放於電腦硬碟並備份於Google雲端硬碟。
- The “exp24” model in SILIC was used to detect animal vocalizations automatically in the recordings and produced sound labels containing the information of filename, soundclass ID, start/end time and low/high frequency, and confidence score of each detected vocalizations. 錄製之聲景資料利用SILIC exp24版進行自動化分析，SILIC會自動偵測並標記音檔中動物鳴叫聲。
- 150 random labels of each sound class were sampled to evaluate the performance metrics including the scores of precision and recall. 每種聲音隨機取150筆資料進行精確率（Precision）與召回率（Recall）估算。
- The threshold of a confidence score was chosen once the precision reached to 0.95 for each sound class. All labels of each sound class with a confidence score no less than the threshold were treated as positives. 取精確率（Precision）為0.95的信賴分數（Confidence Score）作為閾值（Threshold），分數大或等於閾值的偵測結果視為陽性（positive，即認定該筆資料為目標聲音）。
- To reduce storage requirements, we summarized the positive detections of the same species in the same recordings. For example, a single recording (i.e. an event) may be containing none, one or more species (i.e. occurrences), and each species may have vocalized one or more times (i.e. number of observations). 為減少資料量，將同錄音檔且同聲音編號之聲音進行數量加總，每筆含有動物叫聲的錄音檔為一筆調查活動（event），於錄音檔中的1種聲音視為１筆出現記錄（occurrence），觀測數量（number of observations）則為該錄音檔中該物種的鳴叫總次數。
- Wu, S.-H., Chang, H.-W., Lin, R.-S., & Tuanmu, M.-N. (2022). SILIC: A cross database framework for automatically extracting robust biodiversity information from soundscape recordings based on object detection and a tiny training dataset. Ecological Informatics, 68, 101534. https://doi.org/10.1016/j.ecoinf.2021.101534
- Wu, S.-H. (2022). Soundscapes in Yushan National Park, Taiwan (Version 2022-10-19T07:54:54.865215) [Data set]. https://pid.depositar.io/ark:37281/k5x86156b