Optimization of Sensor Placements Using Machine Learning and LIDAR data: a Case Study for a Snow Monitoring Network in the Sierra Nevada.

Abstract:

We present a methodology for the identification of optimal sensor placements and wireless network structure of remote wireless sensor networks. When applied to an existing snow observation network, our results suggest that greater spatial variability and more optimal networks structures could be achieved compared to existing placements. For sensor networks designed to measure spatially distributed phenomena, it is best to choose sites that capture the full range of variables explaining the underlying spatial distribution. In the context of snow depth estimation, topographical variables affecting the spatial distribution include elevation, slope, aspect, vegetation, and concavity. To extract this set of feature vectors, data is obtained from the NSF Open Topography platform, which uses LIDAR flights with 11.65 points per square meter to produce a one-meter raster for the DEM and surface models. Slope and aspect are calculated with the convolution of the elevation matrix and the Sobel operator and the vegetation layer is estimated from a two-meter height filter on the canopy height model. Two types of terrain concavity are calculated from the DEM raster: profile (parallel to the direction of maximum slope), and planform (perpendicular to the direction of maximum slope). Once this feature space is extracted from the LIDAR data, sensor placements can be found using K-means clustering. We use a normalized feature space (in which all feature vectors are scaled from zero to one, thereby evenly weighting each variable). The number of sensors, K, to be placed is taken as an input to the algorithm, which evenly partitions the data into K Voronoi cells, thereby evenly spreading the sensor locations through the space of observed variables. For regions that do not have LIDAR data, we present a methodology that uses a support vector machine algorithm with user-generated training and cross-validation points to classify vegetation from satellite imagery, and compare its accuracy to the LIDAR product.