COS 86-5 - Challenges and Opportunities Scaling Up In-Situ Sensor Networks to Monitor Environments and Biodiversity

Abstract: In-situ remote sensing of biodiversity has emerged as a major activity in ecological research. Increasingly, large-scale applications are emerging that promise minimally-invasive, biologically-detailed monitoring at the landscape-scale and beyond. Here, we discuss a model for scaling up sensing network. We are developing field-deployable, multi-modal sensing networks. Our model is a locally-powered system that sends recorded data over communication networks to a central computing facility where signals are analyzed and reported in near-real time. This design relieves the operator of the regular duty cycle of manually retrieving data and replacing batteries and can provide continuous monitoring. We have developed an open sensor platform design based on the Raspberry-Pi and have made instructions and software publicly available on GitHub. We are collecting base data on power-consumption and reliability of different system components, implementing Wi-Fi and low-power mesh networks, and integrating solar power. In one experiment, we connected 13 hubs that simultaneously record metrics of air quality, weather, soil moisture, acoustical environment, and wildlife-activated cameras. Data are transmitted over wireless networks to off-site repositories where automated analysis pipelines assess a range of relationships. This initial experiment recorded more than 50,000 images from motion-activated cameras combined with physical measurements. In another experiment, we recorded sound from sites distributed in nearby juniper woodlands to profile the acoustic landscape and study bird behavior. Neural-net classification yielded more than one million species-labeled bird vocalizations, the size of which gives some impression of the magnitude of tasks that emerge from continuous sensing. These efforts have resulted in a number of lessons. Hardware reliability and power demands of autonomous sensing remain a challenge, especially in remote areas. The bandwidth demands of streaming continuously recorded data, such as acoustics, consumes considerable energy and can be expensive to transmit across commercial networks. Significant tradeoffs exist between classification on sensor nodes, which reduces bandwidth requirements, and accuracy and reproducibility of results as existing bird-song classifiers are highly sensitive to details of the acoustic landscape. Novel associations emerge from multiple modality sensing, such as between measured animal activity and air quality. Our work, while not yet deployed at scale, can inform other ongoing efforts to create large sensing networks. We believe that many questions remain regarding how sensing efforts across problem domains can coordinate for greater interoperability and capacity for synthesis.