Compressive sensor fusion of weather data

Background: Water management is increasingly important. The better you can estimate evaporation losses, the better you can apply irrigation. Empirical formulas for evaporation rely on weather data such as temperature, wind speed, relative humidity, and insolation.

These can be measured by sensor networks. To reduce the cost of operating a sensor network, people use low power devices and seek to reduce radio transmission costs, which also extend service time.

Several compression schemes exist but are mostly application-independent. Using different compression for different measures helps. We are exploring two issues: how much does knowing how the data will be used let you transmit less accurate or less timely data? And can data from one node improve modeling at nearby nodes and overall compression?

Goal: to save power without sacrificing quality.

Methods: We are using a year's NIWA and other weather data from three nearby stations. Given an empirical formula, it is straightforward to derive how timely and accurate the measurements need to be. Given those figures, it is straightforward to run a model-based compressor and determine the transmission costs for a single node.

The next phase will look at sensor fusion algorithms such as the Brooks-lyengar, trading off fault tolerance against reduced transmissions.

Results: Preliminary Results: simple models can yield 10-90% compression depending on consumer's accuracy and timeliness needs. Sometimes nearby stations have similar measurements, sometimes not, so we intend to develop an adaptive fusion algorithm.

Discussions: The methods will need to be tested on other data sets and on other applications, such as frost warning. The search for simple techniques is most important for low-power devices.

Conclusion: This research only began but already shows that transmission costs can be reduced up to 90%. Next task is to maintain these exceptional results in the presence of faults and microclimate differences.