運用聲音指數探討長期聲景監測的取樣方法

Abstract

聲景生態學在最近的十年蓬勃發展，關注在地景上聲音的組成，用以討論人為聲音的干擾和生物聲音多樣性隨時間改變等議題，提供另一種在生物群集尺度監測多樣性改變、自然擾動與人為影響的可能。由於近幾年錄音工具的快速發展，促使長期且大量收集聲景錄音的研究漸漸增加，也陸續證明聲景在生物群集的層級上能有效反映生物多樣性的變化。然而，長期監測的聲景研究間，並未有一致的錄音方法，且愈來愈大量的錄音資料在儲存與分析所需的軟、硬體資源上皆造成負擔。因此，本研究希望藉由臺灣北部聲景一整年的監測資料來找出最具有成本效益的錄音取樣方法，並探討錄音頻度與錄音覆蓋率對錄音取樣代表性的影響。我選定關渡濕地、陽明山天然林與東眼山人工林三種棲地樣點，架設SM4自動錄音機，自2018年7月到2019年6月每3分鐘取樣錄音1分鐘收集整年聲景監測資料，同時從2019年4月到8月收集每個月各十天的每日完整的聲景錄音。此研究以六種聲音指數數值在每日尺度下的五個百分位數（第5、25、50、75和95百分位數）量化每日聲景特徵，再分別比較利用19種錄音取樣與利用最密集錄音所量化得到的聲景特徵間的差異，評估各錄音取樣方法的代表性。我先計算每分鐘錄音檔的六種聲音指數，並在不同錄音取樣方法下計算各指數的各百分位數，再藉由bootstrapped resampling的方法，以單一指數的單一百分位數在各取樣與最密集錄音間的差異為重取單位，以一千次重取平均值的九十五信賴區間與0重疊與否，判斷各錄音方法和最密集錄音之間的聲景特徵量化結果是否一致。最後以30個指數百分位數（6個指數x 5個百分位數）中結果一致的數量，作為錄音取樣方法代表性的測量值，再分析取樣代表性隨錄音覆蓋率與錄音頻度的變化情形。除了整體的聲景比較外，我也分別針對單一棲地、季節，以及聲音群集的聲景進行相同的分析。整體聲景的研究結果顯示，隨錄音覆蓋率的降低，錄音取樣的代表性愈低，每小時錄音1次的錄音頻度相對較佳。在特定棲地、季節或聲音群集的聲景分析中，錄音覆蓋率愈高則取樣代表性有愈高的趨勢，而錄音頻度在各特定聲景間沒有一致的影響。各取樣方法的代表性在單一季節中較跨季節要高；在單一聲音群集的聲景，則不比跨群集分析擁有較高的代表性；單一棲地則與跨棲地相似。雖然很多因素可能影響長期聲景監測之錄音取樣方法代表性，本研究建議應避免過低的錄音覆蓋率，愈高的錄音覆蓋率原則上愈具聲景代表性，但為有效利用資源，可考量對監測目標之聲景進行前測，並在短期前測中考量季節的影響，避免單一季節的前測低估長期、跨季節監測下的最佳覆蓋率，在聲景資料收集、儲存、分析、研究或管理目標取捨下，找出符合一個地區的最佳錄音覆蓋率與錄音頻度。本研究透過長期且系統性的資料收集，發展具代表性錄音取樣的測試方法，找出長期聲景監測錄音取樣方法的代表性、提供特定聲景監測下的取樣建議，將有利於未來長期且自動排程錄音的聲景監測工作。Soundscape ecology, which studies the composition of sound at landscape scale, has flourished in the last decade. Soundscape research focuses on potential impacts of anthropogenic noise as well as temporal change of biological sound. At the community scale, soundscape provides another dimension of biodiversity and is useful in tracking biodiversity changes due to natural disturbances and human influences. With the rapid development of recording tools in recent years, it is becoming easier to conduct long-term soundscape projects, which generate big data. Such data has allowed researchers to demonstrate that soundscape metrics can be effective in monitoring biodiversity changes. At the same time, it also poses two challenges: 1) recording methods often vary among projects, making comparasion or synthesis difficult; and 2) the increasing amount of data demands more funds and manpower for storage and analysis. This study aims to find the most cost-effective temporal sampling schemes based on a full-year soundscape data set in northern Taiwan. The cost-effectiveness is evaluated by the ability of a given recording frequency and/or recording coverage (i.e. temporal sampling scheme) to represent the soundscape characterized with the full-year data set. I included three habitat types in this study: a wetland (Guandu), a natural deciduous forest (Yangmingshan National Park) and a tree plantation (Dongyanshan Forest Recreation Area). I collected the full-year data by recording 1 minute for every 3-minute interval from July 2018 to June 2019. In addition, I also collected near-complete coverage data for 10 days each month from April to August 2019. Daily soundscape was characterized using six acoustic indices, each with their five percentiles (5th, 25th, 50th, 75th, and 95th percentiles). A total of 19 temporal sampling schemes were evaluated. Each acoustic index was calculated based on 1-minute recording, which gives many index values for a given day. For each day, the five percentiles of each index were then calculated from those 1-minute values. Therefore, a total of 30 index percentiles (6 indexes x 5 percentiles) can be generated daily for each of the 19 temporal sampling schemes, as well as for the full-year and near-complete data sets. The difference in each index percentile between a given temporal sampling scheme and the full-year (or near-complete coverage) data was calculated for each day. The daily differences for a given index percentile were bootstrapped to estimate its 95% confidence interval. If the 95% confidence interval includes zero, the index percentile is treated as being significantly different between a temporal sampling scheme and the full-year (or near-complete coverage) data. The percentage of the 30 index percentiles that are not significantly different from the full-year (or near-complete coverage) data is used as a measure of the representativeness. In addition to the pooled data (pooled across all habitats, seasons and acoustic communities), I also performed the same analysis on specific soundscapes of a single habitat, season, and acoustic community. The results suggest that soundscape representativeness decreased with reduced recording coverage of a given temporal sampling scheme, and the once per hour recording frequency yielded high representativeness. For the analyses on specific soundscapes, the representativeness also decreased with reduced recording coverage, but recording frequency did not have a consistent effect on the representativeness. The representativeness was generally higher in single-season analyses than in pooled-season analysis across all sampling schemes. On the other hand, for the analyses of a single acoustic community or a single habitat, the representativeness across all sampling schemes was similar to that of pooled-data (pooled across all acoustics communities or all habitats) analyses. The findings of this study suggests that low recording coverage should generally be avoided. Furthermore, it is recommended that a pre-testing protocol be implemented prior to a long-term soundscape monitoring project, particularly for correcting potential underestimation of recording coverage that was determined based on a single-season data. The most cost-effective temporal sampling scheme (i.e. recording coverage and recording frequency) for any given study will ultimately be a balance between research/management goals and logistic constrains. This study demonstrated how a pre-testing can be done to find the most representative sampling scheme for long-term soundscape monitoring, and where the same acoustic indices as I used here are involved, the representativeness of the temporal sampling schemes evaluated in this study might be highly applicable.