Using Artificial Intelligence to Study Natural Environments

By Laurel Symes and Shyam Madhusudhana, October 8, 2020
This katydid is one small component of the Panamanian soundscape. AI can be used to detect and study animals from insects to elephants. (Photo: Christian Ziegler).

In rainforests many animals are nearly invisible, but not all are silent. Rainforests click, whirr, and hum with the sound of life. Like heartbeats offer clues to doctors, these sounds convey endless information about the life of forests, capturing the coming and going of animals, the start and end of seasons, and even the sound of wind and rain. Until recently, making sense of these sounds was slow and laborious. At the Center for Conservation Bioacoustics, we are developing ways to automate and accelerate the process.

The need for rainforest conservation has never been more urgent. In recent times, we have witnessed drastic changes in rainforest ecosystems that are experiencing local and global threats. Preserving these environments means protecting not just the large iconic species, such as jaguars, but also understanding the smaller species that form a vital link in the ecosystem’s food chain.

Insects are fundamental components of terrestrial food webs.

Insects are a ubiquitous element of rainforests. Many tropical insects have not yet been named or described. For most, we know very little about where they occur or how many there are. Recent studies suggest that global insect populations may have declined by half or more. Perhaps even more staggering than the purported magnitude of the decline is the simple fact that we do not know to what extent this is accurate and have very few ways to support or refute these claims.

This picture shows the largest and smallest katydid species found in a rainforest in Panama.

Technological advances are providing new ways of monitoring ecosystems. Advances in machine learning technology have made it possible for cell phones to transcribe human speech. Now, we are applying these same machine learning approaches to transcribe the sounds of the rainforest. In particular, we are focusing on the sounds of katydids — a diverse family of large grasshopper-like insects that are a central element of tropical food webs. Katydids feed on a wide variety of plants and are eaten by many diverse predators.

Tracking the populations of insects like katydids is vital in part because populations can change so quickly. Events in a single season or year can drive massive population declines given that the lifespan of these insects only last weeks or months. At the same time, many insect species have a remarkable potential for population growth, meaning that, if given the opportunity to recover, insect populations may rebound quickly, providing critical support for the species that depend on them.

Calling activity for Pristonotus tuberosus. This is a large katydid that produces approximately 1000 calls/individual each night. Top Panel (in blue): Automated detections of Pristonotus tuberosus over five months from a recorder placed in a rainforest tree at 24 meters (~80 feet). The first ten minutes of each hour are shown. Lower left (red and yellow): The proportion of time clips in which Pristonotus tuberosus is detected by time of night. Lower middle (blue): An alternative data presentation showing quantiles for the proportion of clips in which the species is detected. Lower right (black): A histogram shows the timing of calls in focal caged individuals, allowing us to validate results from forest recordings.
Anaulacomera spatulate is a small katydid, and one of the species that is most commonly captured at light traps. In this species, males produce ~170 calls/night. Top Panel: Anaulacomera spatulate is detected throughout the year, consistent with capture data from light traps. Light and dark circles represent moon phase. Bottom panels: In both focal and forest recordings, Anaulacomera spatulate shows high calling activity shortly after nightfall, low activity in the middle of the night, and a slight increase before dawn.

Our initial efforts have focused on Barro Colorado Island, a Smithsonian-run research station located on an island in the Panama Canal. This site is an excellent starting location because we have a multi-year dataset on katydid abundance and acoustic recordings of many species. The goal, however, is to create tools that translate across sites, allowing researchers to apply these techniques to many different species and habitats. We are now using these approaches in forests of Costa Rica to compare the insect communities in sites subject to different restoration approaches.

Top row L to R: Hannah ter Hofstede, Rachel Page, Christine Palmer, Laurel Symes, Shyam Madhusudhana, Jen Hamel, Sharon Martinson; Bottom row L to R: Tony Robillard, Alvaro Vega, Ciara Kernan, Estefania Velilla, Rick Buesink, Inga Geipel

Top row L to R: Hannah ter Hofstede, Rachel Page, Christine Palmer, Laurel Symes, Shyam Madhusudhana, Jen Hamel, Sharon MartinsonBottom row L to R: Tony Robillard, Alvaro Vega, Ciara Kernan, Estefania Velilla, Rick Buesink, Inga Geipel

Developing and applying analytical approaches is an ongoing endeavor of the Center for Conservation Bioacoustics. The work described here is made possible by support from an AI for Earth grant from National Geographic and Microsoft and a grant from the Arthur Vining Davis Foundations.

In addition to collaborations within Cornell, we are working with researchers at many other organizations, including: Castleton University, Dartmouth College, Elon University, Muséum national d’Histoire naturelle (Paris), Osa Conservation, Smithsonian Tropical Research Institute, Universidad de San Jose, and University of Antwerp.

Preparing to put recording units in trees
Collaborative research broadens our understanding of critical ecosystem scale issues.

Students and technicians including: Aboubacar Cherif, Robin Costello, Abenezer Dara, Nate Gallagher, Angela Golding, McKenna Grey, Lars Hoeger, Alina Iwan, Autumn Jensen, Nicole Kleinas, Sara McElheny, Colleen Miller, Rebecca Novello, Kingsley Osei-Karikari, Caitlyn Lee, Jessica, Ralston Jones, Jean Ross, Matt Sears, Nicole Wershoven, Cassidy Yrsha