Introduction

 

Open field farming of tomatoes suffers from the presence of harmful moths whose larvas are devastating. Detecting automatically the presence of moths allows regulating the use of pesticides, according to the actual population present in the field. Up to now, sex pheromone traps have been used, the number of captured insects giving some indication about the population. However, proper inspection of the traps is seldom done, thus leading to inaccurate results. In this work, we propose to equip the trap with an electronic system, in order to detect automatically the moths. The trap is equipped with a microphone coupled with a digital sound recorder. The latter is connected to a gateway, which sends the preprocessed signal to a platform, through a Machine to Machine (M2M) network. In this article, we present a novel method for moth detection from digital sound recordings, based on the spectral features of the moth flight sound.

 

Moth detection and identification

 

Overview

 

Moth recordings are characterized by the spectral components of the signal captured by the microphone. Indeed, the periodic wing flutter results in a line spectrum, containing a fundamental between 40Hz and 100Hz and harmonics. Our approach is based on the continuous analysis of the signal on sliding windows. It is made of two main parts: event detection and identification. The event detection aims at selecting the windows that may contain a moth flight recording and at rejecting all the others. Then, the identification algorithm is applied on the selected windows, in order to perform a spectral analysis and decide if the event corresponds actually to a moth or not. Finally, the rate of windows labeled as “moth event” provides a pertinent estimation of the evolution of the moth population.

 

Event detection

 

The input signal, sampled at 2200 Hz, is split into overlapping windows, each corresponding to a 8.13s recording. The temporal shift between two overlapping windows is equal to 749ms, ensuring the continuity of the analysis. A window is accepted as a possible moth event if the two following conditions are satisfied: first, the signal energy is the highest at the central part of the window; second, the power of the signal on the entire window is higher than a given threshold.

 

Moth identification

 

The identification algorithm is based on a linear prediction analysis, which provides 40 LPC coefficients for each selected window. These coefficients model the signal spectral envelope and they are used as feature vector in the identification process.

The identification module is based on a classification algorithm, which classifies the tested event, represented by its vector of LPC coefficients, in one of the two following classes “moth event” or “non-moth event”. Here we propose to compare the LPC coefficients of the tested event with the typical LCP coefficients of a moth event, obtained through a supervised training phase performed on real recordings. The Mahalanobis distance is used for the comparison, since it allows taking into account the statistical properties of the LPC coefficients of moth events. The calculated distance is compared to a threshold, in order to accept or reject the tested event as a “moth event”.

 

Quantitative evaluation

 

We have performed a quantitative evaluation of our approach, the two parts being studied separately and then jointly, with respect to the two parameters of the algorithm: the detection threshold and the classification threshold. ROC curves have been calculated to this aim, representing the true positive rate as a function of the false negative rate for different parameter values. We found that the detection threshold has little influence on the global performances. However, choosing a low value leads to many selected events and consequently more processing. Then, the quantitative analysis enabled us to set an optimal classification threshold, which finally leads to a true positive rate above 80% for a corresponding false positive rate less than 5% on the tested data (non learnt).

 

Conclusion and perspectives

 

This article presents a new system for the automatic detection of moths from sounds captured with a standard microphone, based on two steps, detection and recognition of moth events. The method is based on the recognition of the spectral features of the sound produced by moth flying. The quantitative evaluation on real data has proven the accuracy and the robustness of the proposed method, which will enable farmers to monitor the moth population in real time. Further work will focus on the implementation on the M2M network, for an efficient and low-cost operational system.