Mobile sensor networks for environmental monitoring

Vulnerability to natural disasters and the human pressure on natural resources have increased the need for environmental monitoring. Proper decisions, based on real-time information gathered from the environment, are critical to protecting human lives and natural resources. To this end, mobile sensor networks, such as wireless sensor networks, are promising sensing systems for flexible and autonomous gathering of such information. Mobile sensor networks consist of geographically deployed sensors very close to a phenomenon of interest. The sensors are autonomous, self-configured, small, lightweight and low powered, and they become mobile when they are attached to mobile objects such as robots, people or bikes.

Research on mobile sensor networks has focused primarily on using sensor mobility to reduce the main sensor network limitations in terms of network topology, connectivity and energy conservation. However, how sensor mobility could improve environmental monitoring still remains largely unexplored. Addressing this requires the consideration of two main mobility aspects: sampling and mobility constraints. Sampling is about where mobile sensors should be moved, while mobility constraints are about how such movements should be handled, considering the context in which the monitoring is carried out. This thesis explores approaches for sensor mobility within a wireless sensor network for use in environmental monitoring. To achieve this goal, four sub-objectives were defined:

Explore the use of metadata to describe the dynamic status of sensor networks. Develop a mobility constraint model to infer mobile sensor behaviour. Develop a method to adapt spatial sampling using mobile, constrained sensors. Extend the developed adaptive sampling method to monitoring highly dynamic environmental phenomena.

Chapter 2 explores the use of metadata to describe the dynamic status of sensor networks. A context model was proposed to describe the general situation in which a sensor network is monitoring. The model consists of four types of contexts: sensor, network, sensing and organisation, where each of the contexts describes the sensor network from a different perspective. Metadata, which are descriptors of observed data, sensor configurations and functionalities, are used as parameters to describe what is happening in the different contexts. The results reveal that metadata are suitable for describing sensor network status within different contexts and reporting the status back to other components, systems or users.

Chapter 3 develops a model which describes mobility constraints for inferring mobile sensor behaviour. The proposed mobility constraint model consists of three components: first, the context typology proposed in Chapter 2 to describe mobility constraints within the different contexts; second, a context graph, modelled as a Bayesian network, to encode dependencies of mobility constraints within the same or different contexts, as well as among mobility constraints and sensor behaviour; and third, contextual rules to encode how dependent mobility constraints are expected to constrain sensor behaviour. Metadata values for the monitored phenomenon and sensor properties are used to feed the context graph. They are propagated through the graph structure, and the contextual rules are used to infer the most suitable behaviour. The model was used to simulate the behaviour of a mobile sensor network to obtain a suitable spatial coverage in low and high fire risk scenarios. It was shown that the mobility constraint model successfully inferred behaviour, such as sleeping sensors, moving sensors and deploying more sensors to enhance spatial coverage.

Chapter 4 develops a spatial sampling strategy for use with mobile, constrained sensors according to the expected value of information (EVoI) and mobility constraints. EVoI allows decisions to be made about the location to observe. It minimises the expected costs of wrong predictions about a phenomenon using a spatially aggregated EVoI criterion. Mobility constraints allow decisions to be made about which sensor to move. A cost-distance criterion is used to minimise unwanted effects of sensor mobility on the sensor network itself, such as energy depletion. The method was assessed by comparing it with a random selection of sample locations combined with sensor selection based on a minimum Euclidian distance criterion. The results demonstrate that EVoI enables selection of the most informative locations, while mobility constraints provide the needed context for sensor selection.

Chapter 5 extends the method developed in Chapter 4 for the case of highly dynamic phenomena. It develops a method for deciding when and where to sample a dynamic phenomenon using mobile sensors. The optimisation criterion is to maximise the EVoI from a new sensor deployment at each time step. The method was demonstrated in a scenario in which a simulated fire in a chemical factory released polluted smoke into the open air. The plume varied in space and time because of variations in atmospheric conditions and could be only partially predicted by a deterministic dispersion model. In-situ observations acquired by mobile sensors were considered to improve predictions. A comparison with random sensor movements and the previous sensor deployment without performing sensor movements shows that the optimised sensor mobility successfully reduced risk caused by poor model predictions.

Chapter 6 synthesises the main findings and presents my reflections on the implications of such findings. Mobile sensors for environmental monitoring are relevant to improving monitoring by selecting sampling locations that deliver the information that most improves the quality of decisions for protecting human lives and natural resources. Mobility constraints are relevant to managing sensor mobility within sampling strategies. The traditional consideration of mobility constraints within the field of computer sciences mainly leads to sensor self-protection rather than to the protection of human beings and natural resources. By contrast, when used for environmental monitoring, mobile sensors should above all improve monitoring performance, even thought this might produce negative effects on coverage, connectivity or energy consumption. Thus, mobility constraints are useful for reducing such negative effects without constraining the sampling strategy. Although sensor networks are now a mature technology, they are not yet widely used by surveyors and environmental scientists. The operational use of sensor networks in geo-information and environmental sciences therefore needs to be further stimulated. Although this thesis focuses on wireless sensor network, other types of informal sensor networks could be also relevant for environmental monitoring, such as smart phones, volunteer citizens and sensor web. Finally, the following recommendations are given for further research: extend the sampling strategy for dynamic phenomena to take account of mobility constraints; develop sampling strategies that take a decentralised approach; focus on mobility constraints related to connectivity and data transmission; elicit expert knowledge to reveal preferences for sensor mobility under mobility constraints within different types of environmental applications; and validate the proposed strategies in operational implementations.