Lead Investigators:

Diane Budzik, Maxim Batalin, William Kaiser

Overview

Perhaps the most challenging scaling problem for distributed sensing is that of large region characterization by sensor networks.  Many applications require both microscopic scale detection of phenomena and, at the same time, wide area measurement to enable regional characterization.  The problem of ecosystem monitoring and the optimization of land use, the proper distribution of water resources, or the characterization of environmental contaminants all are examples where the scales of centimeters to kilometers may contribute to investigations.  The conventional approach of dense sensor deployment is not applicable due to the impractically large deployment numbers that apply, in particular to three dimensional volumes. Thus, there is an urgent need for new networked sensor systems that combine sensing on multiple scales for both multiscale and multimode sensor data fusion and multiscale sensing for guidance of actuated sensing.

Approach

NIMS multiscale sensing is based on a hierarchical system that enables autonomous arrangement of sensors with the objective of optimizing sensing fidelity, spatial coverage, and mobility characteristics. This system of sensors then can be used for efficient high fidelity sampling of high frequency spatiotemporal phenomena.

Figure 1. A schematic view of a NIMS two-tier multiscale architecture.

A schematic view of the two tier architecture is shown in Figure 1. In this architecture, dynamic phenomena characterized by high spatiotemporal frequency associated with its field variables is captured by a first tier sensor. The first tier is represented by a static low-fidelity high spatial coverage sensor providing “global” information about the environment. Note that this first tier sensor could instead of being a single sensor be comprised of several static sensors in which their data is interpolated to provide a “global” view of the environment. This information is then used to extract the regions of interest (regions of high phenomenon variability). These regions form a set of sampling tasks for the second-tier sensors to pursue. The second tier is represented by the mobile robots equipped with high-fidelity low spatial coverage (spot measurement) sensors. A set of new tasks is given as an input to the Task Allocation module. The task allocation module prioritizes tasks based on the selected utility and assigns the task with highest utility to the available mobile robot for high-fidelity sampling.

An output of the system is a set of high-fidelity phenomenon measurements in a given region, which then can be used by scientists. In the future, we plan to augment the described architecture with two new modules. The first module is Task Characterization. Task Modeling will collect the complete task information from the Task Allocation module and build a model of future task arrivals and distributions. Task modeling would further improve task allocation and at the limit yield an optimal solution. Another future module is phenomenon characterization based on the received high-fidelity sensed values. This information can be used to improve first-tier sensor data processing and segmentation, as well as calibration. In theory, if the first-tier sensor is calibrated in accordance with the spatiotemporal nonlinearities in the environment, the high-fidelity phenomenon information can be extracted (or much closer approximated) directly from the first-tier sensor.

Systems/Experiments

The performance of the multiscale paradigm for sampling light intensity was tested and analyzed through simulations and on a physical system.  The first tier sensor data (image sensor data sources) were captured from a study area located at the James San Jacinto Mountain Reserve during daylight periods. Images were acquired every 15 seconds - note that the physical delay in image acquisition dictates the length of the decision epoch to also equal 15 seconds. A down-looking imager captured snapshots (768x480 pixels) of the understory of a forest canopy covering an area approximately 6 meters in length by 4 meters in width. Images captured between 10:00 a.m. and 11:00 a.m. were analyzed in simulation and also were sampled in the laboratory using the NIMS 3D system.

These images were experimentally verified (by analyzing images during other times of the day) to be representative of the spatial and temporal variations occurring in the transect throughout the day. The images captured constitute the information sensed using a high spatial coverage, low-fidelity sensor (imager) and were processed to extract a set of tasks that represent possible regions that could be sampled using the low spatial coverage, high-fidelity optical sensor carried by the mobile robot.

Figure 2 displays results comparing actual and reconstructed scenes.  Here, the dark and light areas of this image display the presence and absence of solar radiation.  The form of this distribution is of interest and is the measurement objective of the user.  This result shows an example where the multiscale system performs task allocation with priority assignments based upon the measured area of an environmental region.  Figure 3 displays the complete results for varying actuated sensor motion speed capability (for actuated sensors moving at speeds of 40 to 500 cm/s). Here, this compares normalized sample area (ratio of total sampled and reconstructed area to actual area) to sampling period (density), s, where s varies from one sample per 4 centimeters to one sample per 20 centimeters.  An experiment was also performed to compare the performance of a multiscale approach to a traditional full raster scan of the environment. The raster scan samples the complete environment with a desired density. The raster scan was implemented with an average speed of 40 cm/s for all sampling densities  

Figure 2. Comparison of light intensity distribution captured in the field (at left) and reconstructed data (at right). This example displays the results for an algorithm that assigns sampling task priority to the field regions of largest continuous area.  This image shows the results in the instant of time when the actuated sensor system has just successfully captured part of the scene at lower right and has accurately reconstructed this data as seen in the right hand panel.

Figure 3 demonstrates that for corresponding speeds, a multiscale approach performs better than a simple raster scan in terms of the amount of information extracted from the environment. Additionally, the multiscale paradigm yields greater fidelity as well. In the raster scan, the information extracted by sampling the complete transect area initially results in greater error because of the phenomenon dynamics.  This is evident, for example, from the number of images processed in the raster scan. They varied from 1 image for s = 4 to 9 images for s = 20. The total number of images processed (using the same average speed as for the raster scan) using a multiscale approach varied from 13 images for s = 4 to 121 images for s = 20. Thus, a multiscale approach captures more up-to-date information from the environment resulting in greater fidelity.

Figure 3.  Simulation results for Multiscale and Raster scan sampling comparing normalized sampled area for different densities and speeds.

Accomplishments

We were able to show the potential for multiscale sensing through simulations that used images from the environment.  We demonstrated that trends shown in simulation are representative of performance on a physical system by running the multiscale algorithm on NIMS-3D.  We are currently extending this work which we will submit as a journal paper to TOSN in April.  Current work focuses on the comparison of static sensor sampling, deterministic actuated sensor sampling (raster scan), adaptive actuated sensor sampling, and multiscale sampling.  These methods are the different classes of methods that are currently available for sampling environmental phenomena.  When choosing the method to use, the user must consider the dynamic nature of phenomena, the measurement objectives, and available resources. This paper will provide the first investigation of relative performance for static sensor network sampling and actuated sampling approaches (raster scan, adaptive sampling and multiscale sampling) applied to environmental phenomena displaying variable degree of spatiotemporal dynamics. This study is being conducted on a case study of actual environmental phenomena - solar light distribution, which is critical to environmental ecosystem investigations. We will perform an in-depth analysis of representative approaches in simulation, using the solar light patterns captured in the real environment, as well as in an actual environment using a cabled 3D actuated system, NIMS-3D.

We will examine phenomenon with roughly four spatiotemporal dynamics corresponding the four combinations of predominantly low or high spatial frequency and low or high temporal frequency characteristics  For each dynamics characterization of the phenomenon we will compare the performance of static sensor network sampling, raster scan, adaptive sampling and multiscale sampling. We plan to use the following performance metrics: the fidelity of light intensity field reconstruction and the accuracy of the derived average light intensity.

Experimental evaluation will occur with two methods designed to permit a wider scope of characterization.  First, a field variable library was acquired from the environment for the evaluation of static and actuated sensor systems where each system is be modeled in emulation, based on a duplication of software systems used in the field.  Then, in addition, the NIMS 3D system will be used to validate this method through operations in the field.

Future Directions

In the coming year, multiscale sensing will be applied to water quality experiments.  Rhodamine dye will be released in the San Joaquin-Merced confluence zone.  The flow of the dye will be tracked using the multiscale sensing algorithm so that data about river flow patterns and dissolved oxygen can be recorded.  Two NIMS transects will span across the width of the river.  The first NIMS transect will scan rapidly and coarsely giving a “global” view of the dye flow patterns.  The second, downstream NIMS transect will more densely scan a section of the river based on the output of the multiscale task allocation module, which takes as input data from the upstream NIMS transect.  This will demonstrate multiscale’s wide applicability and the flexibility of the two-tier architecture.

People

Faculty:

• William Kaiser, UCLA, EE, Professor
• Gaurav Sukhatme, USC, CS, joint appointment in EE, Associate Professor
• Mark Hansen, UCLA, Statistics, Associate Professor

Staff:

• Maxim Batalin, UCLA, EE, PhD

Graduate Students:

• Diane Budzik, UCLA, EE, PhD student
• Amarjeet Singh, UCLA, EE, PhD student