Overview

In a habitat monitoring deployment it is possible to optimize sensor node placement in order to minimize radio communication and maximize network lifetime. This is a worthy goal because the longer lived the network, the more data it can collect. However, the utility of the data collected is also highly dependent upon where the sensor nodes are placed. A network that lives a long time and collected massive amounts of data that is ultimately useless to application scientists is worthless. Our goal is to explore techniques for optimizing in terms of sensing, irrespective of the communication topology.

Approach

UCLA's Mildred E. Mathias Botanical Gardens collects several different botanical habitats into a relatively small region, combining varying topography with interesting ecology. It sits on over 7 acres of land, and is home to several species of tropical and subtropical plants. Our goal is to try to capture how temperature and humidity vary over the garden. To this end, we've deployed 11 static sensor nodes and two mobile sensor nodes, each equipped with temperature and humidity sensors. The static sensor nodes provide a general overview of the garden. The mobile sensors satisfy two purposes. First, they can be used to zoom in on a particular geographical region in order to capture spatial variability that the static sensors cannot capture. Second, by placing a mobile sensor next to a static sensor it is possible to check how well the one sensors is calibrated with respect to the other, and how much noise the actual sensing hardware is introducing into the sensor readings themselves.

Systems/Experiments

Our sensing system consists of three parts. First are the nodes that are equipped with sensors themselves (both static and mobile). These nodes are connected to temperature and humidity sensors. However, the distance between sensor nodes, combined with the dense foliage of the garden makes communication difficult. The second component of our system is a set of nodes that are devoted to acting as repeaters. They don't do any environmental sensing (although they do report there current battery voltage, which is useful for maintenance planning) but serve only to create a richly connected communication topology. The final piece is a microserver where all the data is collected.

Once we were collecting lots of sensor data reliably, we analyzed the data to try to understand which static sensors were good predictors of other static sensors. We identified two “cliques” of sensor nodes that formed contiguous regions. Using our mobile sensors we were able to collect data from those regions that are near the interface of the two cliques.

Accomplishments

Our initial deployment ran into sever communication problems that were unlike any previous habitat monitoring deployments we had done. Poor connectivity necessitated deploying several repeater nodes, but we had only a rough idea of where to place them. In order to limit the time spent in trial-and-error repeater placement we employed two techniques. First, we used a handheld computer to survey various locations in the garden to get an idea of connectivity. Second, we over-deployed repeaters throughout the garden in order to ensure a connected network. At its peak, the garden had 34 nodes. From this we collected route and link statistics and established which repeaters where helping to form a network backbone, and which were largely superfluous. Once the backbone was identified we pruned the number of repeaters. This process was greatly expedited by using tools that support fast convergence of newly deployed nodes.

We've begun to develop statistical models based upon the data collected that predict what the temperature is at a point in the Botanical Garden, in addition to a confidence in the prediction. Using the mobile sensors we will be able to enhance the model by providing data from a greater set of points than our static sensors would otherwise provide. In addition, the mobile sensors allow us to check our predictions in order to validate our model.

Future Directions

In the immediate future we plan on building our models and validating them as described. In the process, we hope to develop general modeling techniques that will be applicable to other habitat monitoring applications. Using these techniques it should be possible to identify placement for sensors to capture the phenomenon of interest will minimizing the number of sensors required.

People

Tom Schoellhammer, Miodrag Potknojak, Deborah Estrin