The Endangered Species Act and other environmental regulations, such as the Sikes Act, require Department of Defense (DoD) installations to have Integrated Natural Resource Management Plans and responsible biodiversity stewardship. This project sought to facilitate compliance with these regulatory imperatives by further developing software and hardware to automatically survey and monitor bats, with a demonstration application of this approach for birds. Particularly for rare species, monitoring accrues high costs because of the specialized skills, time, and spatial coverage required. The objective was to develop a system to automatically monitor bats for weeks or months by recording and analyzing the vocalizations they produce to assess species presence/absence, population levels, temporal movements, and acoustically gleaned demographic information.
Although relatively easy to record with ultrasonic-sensitive recording equipment, confident matching of bat sounds to species requires a comprehensive collection of species-known recordings from a variety of conditions to sufficiently cover each species’ call repertoire. This project combined more than 10,000 echolocation call sequences of species-known bats from 37 species in 30 states recorded as high-resolution, full-spectrum data. This format enabled analysis with an intelligent routine to automatically track call trends through noise and echoes to extract and quantify subtle signal parameters and enable the assessment of signal properties for quality control. The compiled known data supported the creation of an expert system to classify similarly parameterized unknown data. This expert classification of calls and sequences of calls uses an ensemble consensus of redundant hierarchical decision algorithms that reports a single species decision only when a result meets or exceeds an acceptance threshold at each decision step and satisfies redundant checks and signal assessments. Because of the greater number of bird species, and the complexity and variety of their calls and songs, this project adopted an alternate approach to recognize target signals for bird signal search and recognition.
Bat species classification using the expert system outperformed tests using other standard machine intelligence systems, e.g., Artificial Neural Networks. Because of signal noise and that many bat species have overlapping call characteristics across some or all parts of their call repertoires, the classifiers cannot discriminate every recording to species, but did achieve correct identification rates of 90-100% from calls and sequences outputted from the classifiers as acceptable following signal assessment and redundant checks. Prototype field recording units enabled testing and assessment of recording under a variety of field conditions, and provided recorded sequences that directed improvements of the automated signal processing routines to reduce misclassifications and provide quality control. Bat classifier systems already developed include Northeastern, Midwestern, Ozark, Pacific Northwest, Great Basin, and Montane North regions of the United States. Some additional fieldwork remains to enable Southeast and Southwest classifiers. The bird signal search algorithm demonstrated robustness in its ability to rapidly find search targets even with low amplitude signals that occur among noise.
The software analysis and hardware approaches developed and demonstrated by this project will enable both short- and long-term non-invasive survey and monitoring of bat activity, bat species occurrence, and occurrence of targeted bird species at reduced cost and increased temporal and spatial coverage. This project will contribute recording data to augment the collection of the Cornell Laboratory of Ornithology’s Macaulay Library.