The History of SonoBat

Dr. Joe Szewczak (2004)Dr. Joe Szewczak processing field recordings as
part of the OR Bat Grid project (2004).

The foundational research for SonoBat began in 1991 when Dr. Szewczak and his students began an investigation of bat species and distributions in Eastern California. Recognizing the inherent bias of mist net studies, and frustrated by bats flying past the nets, he attempted to supplement species detection with then routine zero-cross ultrasound recording analysis. Right away the coarse signal resolution inherent with zero-cross data compression left him wishing to see more detail in the bat calls to help discriminate the many acoustically similar species with confidence. Turning to full-spectrum data provided the level of detail he sought, along with additional dimensions of information in the calls, and providing context of the acoustic environment vital for interpreting data integrity. Szewczak gleaned the value of this data using proto-SonoBat signal analysis routines he coded to perform tasks he could not do using existing analytical software. Assuming that advances in hardware and software would eventually make using this data practical, he began rigorous and directed collection of species-known full-spectrum data from his field sites. Szewczak envisioned the benefit of a robust reference library of full-spectrum species-known recordings from different microhabitat elements and regions to support acoustic species recognition and for other scientific research initiatives.

By the late 1990s, Szewczak had thousands of recordings in his library, but still did not have the analysis tool he wanted that would facilitate rapid access to the essential elements of echolocation calls to:

  • visualize and interpret echolocation sequence content,
  • view individual calls in a consistent way to support learning and recognizing species-discriminating characteristics
  • standardize comparison of known recordings with unknown ones
  • quantify call characteristics for statistical analyses.

To fill this need, Szewczak called upon his earlier background in engineering to develop such a tool that met his needs and he shared with the community of bat biologists with the original 1998 release of SonoBat. The value of these core elements have proven essential after two decades of use and research by thousands of users around the world as SonoBat has continued to grow and expand with features and functionality.

Field data interpreted in SonoBat 1 Revealed a wealth of data that allowed for previously undiscernible species to be classified. Full Spectrum’s visualization of the entire ultrasound environment also allowed the researcher to draw stronger conclusions by:

  • Displaying the power distribution of a call, which removed ambiguity between distant partial recordings that mimicked Little Brown Bats and the Little Brown Bats themselves.
  • Cluing the researcher in to distortions in an echolocation pulse by visualizing the whole ultrasound environment before, during, and after an echolocation pulse.
  • Demonstrating the effects of microphone placement on the quality of recordings. Previously untestable assumptions on microphone placement and design were found to alter the conclusions drawn from the data recorded.

The mid-2000s’ development of automated triggered recording overwhelmed users’ capacity for manual processing, resulting in the development of the SonoBat Scrubber. The Scrubber automatically discerned bat calls from other ultrasound signals in order to delete extraneous noise files. Various SonoBat Attributers were also developed that properly formatted the output from various devices and embedding vital metadata notes into batches of recordings efficient data management and archiving.

However, the greatest enabling benefit for managing and processing large data sets came from automating bat call and sequence data extraction and classification decisions. Early versions of SonoBat relied upon users to position onscreen cursors to designate positions of call parameters and quantify this call parameter data. These only captured a limited set of discrete values and would vary by user. Automated extraction of call parameters both sped up this process and provided replication of measurements. More importantly it allowed Szewczak to extract call data from the more than 10,000 species-known recordings he had acquired.

With the benefit of automated trending, the quantitative description of calls could include integrative measures of shape functions and dynamic measures in both the time-frequency and time-amplitude domain. Simply put, this meant that all the information held in a full spectrum recording could be brought to bear on the species identification process. SonoBat assesses more than 90 quantitative measures to describe calls and identify the nuanced differences between different call-types and species. At the time Szewczak ran the first automated data extraction, he’d generated a database in excess of 3 million call parameter data entries that supported the initial classifiers used to build the automated SonoBat 3 released in 2009.

The current form of SonoBat has built upon this. The classification engine is improved to mimic the decision process of an expert user manually vetting files, as well as incorporating advances in the fields of machine learning and signal theory to classify full spectrum data ten times faster than before. In addition, the region-specific library of recordings used to train our classifiers has bloomed to over 4 million discrete bat recordings. The end result is an unparalleled accuracy in species identification with SonoBat 4. The inclusion of SonoBat LIVE for active monitoring extends this capability into real-time classification for field use.

[button open_new_tab=”true” color=”Accent-Color” size=”large” url=”../SonoBat4″ text=”SonoBat 4″]            [button open_new_tab=”true” color=”Accent-Color” size=”large” url=”../sonobat-live” text=”Sonobat LIVE”]