Recipient Organization
SIGMA SPACE CORPORATION
4801 FORBES BLVD
LANHAM,MD 20706
Performing Department
(N/A)
Non Technical Summary
USDA has a goal to enhance detection and response capabilities in controlling destructive pests like the Mediterranean fruit fly. Currently, about a hundred thousand Medfly traps exist in southern California alone that requires USDA personnel to manually check for the existence of the target insect. This labor-intensive activity carries significant recurring costs. Moreover, any delays in the identification of an infestation can quickly produce millions of dollars of damage to crops. We submit that an effective electronic solution to this problem would have many important benefits: 1) reduce costs for monitoring for target insects; 2) dramatically improve response times when a problem occurs; 3) reduce damage to crops; 4) geographically isolate the problem areas for proper response (sterile release, pesticides) trapping, and so on. The purpose of this project is to design and develop a system that can automatically identify targeted classes of invasive flying insects,
such as Medflies, using their audible wing-beat patterns. We propose a novel solution to this problem using electronic sensing (microphone and data processing) in a wireless sensor network, dubbed InsectNet, integrated with modern Medfly traps. The sensor nodes will also be endowed with wake-sleep behavior to minimize communications and conserve energy.
Animal Health Component
100%
Research Effort Categories
Basic
(N/A)
Applied
100%
Developmental
(N/A)
Goals / Objectives
To research, develop, and test the feasibility of new pattern recognition and machine learning techniques for parsimonious electronic implementation in a wide-area sensor network of flying-insect traps. Of particular interest is the automatic identification of fruit flies from their characteristic wingbeat sound patterns in the presence distracters (background noise and other signal sources) and to do so with low rate of false alarms. Requisite hardware and software components will be developed in prototype form and integrated into a small sensor network system for demonstration. Efficacy tests will be conducted in a controlled yet realistic environment using real audio from targeted insects and environmental noise. What we learn in this phase will shape future improvements and engineering of a commercial product for large-scale deployment.
Project Methods
The approach uses a Bayesian framework of graphical models and sparse coding of natural sounds to discover distinguishing patterns useful for signal classification. The acoustic signal of the wingbeat source is sampled from a suitable microphone, filtered, amplified, and processed electronically to extract salient features used to recognize patterns. A graphical model of latent variables of factors that best explain the observed patterns is first learned and then later used to infer the cause of each input signal. This approach provides a principled way to identify the targeted sound (of a fruit fly or other) better than using simple spectral analysis and cross-correlation with saved templates; the latter approach is ill-equipped to deal with natural variation in the signals and environment as well as the significant overlap of spectral components observed in other distracter signals. Rather, our approach trains a model with sufficient statistics to recognize
prototypical patterns over an ensemble of training samples. This model-based approach also takes into consideration the natural variability per sample and, therefore, should be able to classify novel input signals consistently and robustly when deployed in the field.