An intelligent system for estimating frog community calling activity and species richness

Over the past decade, dramatic declines in frog populations have been noticed worldwide. To examine this decline, monitoring frogs is becoming increasingly important. Compared to traditional field survey methods, recent advances in acoustic sensor technology have greatly extended spatial and temporal scales for monitoring animal populations. In this paper, we examine the problem of monitoring frog populations by analysing acoustic sensor data, where the population is reflected by community calling activity and species richness. Specifically, a novel acoustic event detection (AED) algorithm is first proposed to filter out those recordings without frog calls. Then, multi-label learning is used to classify each individual recording with six acoustic features: linear predictive coding coefficients, Mel-frequency cepstral coefficients, linear-frequency cepstral coefficients, acoustic complexity index, acoustic diversity index, and acoustic evenness index. Next, frog community calling activity and species richness are estimated by accumulating the results of AED and multi-label learning, respectively. Finally, ordinary least squares regression (OLS) is conducted to reveal the relationship between frog populations (frog calling activity and species richness) and weather variables (maximum temperature and rainfall). Experimental results demonstrate that our proposed intelligent system can significantly facilitate the effort to estimate frog community calling activity and species richness with comparable accuracies. The statistical results of OLS indicate that rainfall pattern has a lagged impact on frog community calling activity (significant in the first day after rainy day) and species richness (significant in the fourth day after rainy day). Temperature is shown to affect species richness but is less likely to change calling activity.