ITR: Pattern Recognition for Ecological Science and Environmental Monitoring

点击翻译按钮获取中文摘要

Ecological science and environmental monitoring need inexpensive methods for obtaining population counts of insects. Existing methods require identification by human experts, which is too costly to support large-scale scientific and environmental monitoring studies. This project is developing a mechanical/computer device that can automatically separate, photograph, and classify insects to produce insect population counts. This device is being developed and tested on two important environmental problems: (a) water quality monitoring in streams (by recognizing and counting stonefly larvae in stream substrates) and (b) measurement and characterization of forest beetle biodiversity (by recognizing and counting soil mesofauna).Developing such a device raises two challenges for pattern recognition: (a) develop methods that can recognize flexible 3-dimensional objects, and (b) develop general-purpose methods that can be applied to new sets of insects simply by retraining. To meet these challenges, the project is exploring two new approaches to pattern recognition: (a) feature dictionaries (algorithms that find and extract a large number of visual features from images), and (b) relational appearance-based methods (algorithms that decompose the object into its parts and learn what each part looks like under a wide variety of viewing angles and lighting conditions).This project has many potential impacts. A low-cost, efficient method for monitoring insect populations in streams would revolutionize water quality monitoring. An automated device for classifying and counting soil mesofauna would reduce the amount of time required to obtain population counts 1000-fold. This would allow a broad spectrum of biologists to employ biodiversity indicator analysis on a wide scale for basic research and environmental monitoring.