David De Roure
School of Electronics and Computer Science
University of Southampton, UK
It seems that stories of floods are never far from the news. Early warning is critical to reduce the scale and subsequent cost of flood damage. The latest computing technologies provide an opportunity for closer monitoring and more effective prediction.
We are all aware of the increasing number of computing devices that we encounter on an everyday basis — not just the desktop and laptop personal computers but the mobile phones, PDAs (portable digital assistants), digital cameras and other electronic accessories that we carry on our person. Enabled by technology advances and driven by customer demand for portable devices which are smaller, lighter and which run for longer on batteries, we are seeing a new generation of computing hardware and software which is known as pervasive or ubiquitous computing. These portable devices communicate using a range of wireless technologies including Bluetooth, which was conceived to replace short cables between devices, wireless Ethernet for local area communication and the GSM (Global System for Mobile Communications) infrastructure which provides widely accessible telephony and data services.
Pervasive computing has been the subject of research for well over a decade. There is a clear trend towards increasing numbers of computing devices all around us, not just on our person but embedded in our surroundings and within everyday artifacts — note for example the increasing numbers of processors inside our electrical appliances, and cars now number tens of embedded processors. Many of these devices have a role in sensing the state of the physical world. As the technologies continue to improve we will see smaller, lower power and lower cost devices. The latest research has created novel methods for harvesting energy from the environment to provide self-powered microsystems, giving a glimpse of the degree of pervasiveness the future may hold.
These technologies provide exciting new opportunities for monitoring the natural environment, such as measurement of water levels and air pollution. Traditional solutions involve dataloggers from which data is collected periodically in person or via telemetry. With wireless communications and energy drawn from local sources such as solar cells, devices can be deployed without the constraints of having to wire them up, and data can be conveyed when needed. Significantly, these technologies make it possible to deploy more devices in order to obtain more data more often, and this greater richness of data is set to create a powerful impact on environmental monitoring and decision-making.
Deploying devices in the natural environment brings a number of challenges. Devices need to withstand harsh environmental conditions. It is often also the case that deployment is expensive and it may be difficult or impossible to access the devices again later, in contrast to working with handheld pervasive devices such as mobile phones. Devices need to run for considerable periods of time making use of available power sources, so conserving energy is hugely important. Taking sensor readings consumes some power but transmitting the data typically takes significantly more energy, so the more often data is sent, the more likely it is that the device will then not have sufficient power to continue its function. There may also be large numbers of nodes and they need to have coordinated behaviour, but at the same time we have to assume that node failure, temporary or permanent, is likely to occur — the natural environment is a place of change.
The Centre for Pervasive Computing in the Environment, sponsored by the DTI as part of the Next Wave Technologies and Markets Programme, contains three projects which address these challenges directly. The harsh conditions and inevitable failure are particularly well illustrated by GlacsWeb, a project to monitor glacier movement in which devices are dropped down holes bored deep into the glacier and the eventual destruction of the device is certain. The two other projects adopt contrasting approaches to achieve coordinated behaviour in the face of unreliable devices and communications: Floodnet, described below, uses techniques of artificial intelligence to provide intelligent behaviour which is explicitly encoded in sets of rules, while in SECOAS (Self-Organising Collegiate Sensor Networks) the interactions between devices are simpler and their design draws inspiration from biology. These studies are establishing best practice and identifying research issues for future projects in the environment.
Flood damage represents a major ongoing cost and risk may be increasing due to land-use change, climate change and flood-prone investment. The Environment Agency has undertaken considerable work in responding to the flood risk hazard, for example investing in defences and in research both in predictive monitoring and preventative technologies, disseminating information and providing early warning and advice. However there is still a considerable area of remaining uncertainty, which is complicated by flooding from surface water and groundwater. Both rural and urban flooding and response are complicated by the predominant inability to model the roughness and flow routes to a reasonable level. Within the built environment there are considerable uncertainties based on the flow routing and the interaction between the surface and subsurface drainage. Existing monitoring is sparse, independent and non-adaptive, preventing effective real-time understanding of the evolution of any event and any non-predictable influences.
Flood monitoring provides an excellent illustration of the benefits and challenges of pervasive computing in the environment. When a flood occurs, the cost of damage has a clear correlation with both the depth of the flooding and the time in advance at which warning is given. By applying pervasive computing technologies on the floodplain we have the potential to obtain better data from which to make predictions, and we can do this in a timely manner in order to improve warning times. Deployment is facilitated by wireless technologies but we have issues of power for the devices and the need for very long unattended periods of operation.
These are the motivations for the Floodnet project, which brings university academics working in pervasive computing and in the management, analysis and processing of environmental data, together with technology providers, system integrators and end users. Significantly, the users have been engaged in all stages of the design, development deployment and operation of the Floodnet system.
The project has deployed a set of intelligent sensor nodes around a stretch of river in Essex, chosen for its tidal behaviour so that, for test purposes, there are regular variations in water level. The nodes are powered by solar cells in conjunction with batteries and each node communicates with its neighbours using wireless Ethernet. A special node, the .gateway., relays the data back to University of Southampton using GPRS. This is an .ad hoc network., with nodes relaying information across the network to ensure data delivery to the gateway. Various parties can subscribe to the incoming data stream. As well as being stored in a Geographical Information System, the data is used to inform flood simulations which are used to make flood predictions. The system is depicted in figure 1. One can envisage a number of such deployments at different locations in the river, each reporting back in this way — the current deployment enables us to explore the issues of working with this spatial density of data.
Figure 1 The Floodnet system architecture
The fundamental tradeoff between the need for timely data and the need to conserve energy is the research focus of the project. The goal is to make the system adaptive so that the sampling and reporting rates of the devices vary according to need, conserving power and minimising the data volume required. Some intelligence is required in this adaptation, because the importance of a device at a given moment will depend upon both its environmental context and its role in relaying data from other devices, each of which will vary dynamically according to circumstances.
The process of adaptive sampling is mediated by the use of a predictive model which allows for the real-time collection of data to update the flood predictions regularly with refreshed data. The predictive model is required to carry out extensive processing in a short period of time (currently 20 minutes to 1 hour). Upon each model iteration the network changes its behaviour, altering the reporting rates of each individual node according to the time variable demand placed upon it by the predictive model. In the broadest sense the nodes that first experience a flood event will have a high initial demand. This demand will ease as the model develops a sufficient level of confidence in the prediction for this particular node and as the flood event itself penetrates into surrounding areas, raising the data demand from these areas. With time the more disparate nodes become more active. As such a wave of activity passes from the sensors in the main channel out to the floodplain areas.
The adaptive sampling is achieved through a series of control loops, as depicted in figure 2. The outermost loop enables the flood predictions to influence the reporting rates of individual nodes, so that closer monitoring can be achieved in anticipation of a possible flooding event. The inner loop is the peer-to-peer behaviour of a set of nodes which can communicate with each other but have no external coordination and are therefore described as self-managing or autonomic, a word borrowed from the notion of the autonomic nervous system. Other possibilities include one node, such as the gateway, taking a coordinating role.
Figure 2 The adaptive sampling control loops
To design the adaptive sampling algorithms, we have built a simulator for a large number of nodes. As well as being an effective means of designing and testing the algorithms before deployment, the use of a simulator is an essential step in addressing the scalability that is required in the environmental context. We have been able to design for a larger scale deployment, with multiple clusters of nodes of varying sizes, and then generate solutions for our particular configuration of deployed nodes. The simulator is shown in operation in figure 3. We use real water level records, simulated food events and adjustable models for available sunlight in order to evaluate the algorithms. Significantly, the simulator has provided a tool to support design sessions with the environmental experts and users. This methodology, in which we focus on providing tools to support the designer, has been established through previous work at the University of Southampton in the Equator Interdisciplinary Research Collaboration, a large UK research project conducting innovative pervasive computing deployments which include environmental monitoring.
The simulator is based on software agent technology and this reflects our interest in adopting an agent-based approach in deployments of this type, since the agent-based computing paradigm pays considerable attention to issues of autonomy and to dynamic adaptation. We also see a role for agents in handling the evolving behaviours of a set of nodes deployed for a considerable period of time, as their purpose changes and multiple .applications. may run over them. Moreover, so-called mobile agents provide a mechanism for an activity to move from one node to another, perhaps in response to imminent node failure or environmental events. In the software running on the current Floodnet nodes we are borrowing ideas from agent-based computing without adopting an existing software agent framework, as currently these are relatively heavyweight solutions to run on low power devices.
Figure 3 Investigating a deployment of 12 sensor nodes using the Floodnet simulator
One side effect of the higher spatial and temporal density of the incoming data, and the move towards a more rapid turnaround in the processing of the data, is that there is a significantly greater amount of computation in performing the flood predictions. For this we have turned to Grid computing, the technology that harnesses the power of multiple networked computers in order to achieve computations which would not otherwise be possible. By running multiple flood simulations simultaneously we can automatically try out different .what-if. scenarios and these are informed in near real-time by the incoming stream of data. In fact this poses some challenging requirements for Grid computing, which is typically used in a more batch-oriented fashion rather than within a real-time process. This work also raises an open research question: to what extent can the computations occurring on the Grid be distributed to occur on the sensor node network instead, increasing the autonomy of the deployed network?
The goal of projects like Floodnet is to produce information about the environment, and it is important that the information can be combined easily with other sources, since it is the fusion of multiple information sources which leads to valuable knowledge to support environmental decision-making. Accordingly we have adopted an information-systems focus throughout the work. The requirement for interoperability and automation applies not only to the environmental data but also to the operational data, as this will enable the sensor nodes to interoperate with other systems in the future. In Floodnet this machine-processable operational knowledge includes the descriptions of the adaptive sampling behaviours. To represent this knowledge we have adopted the latest technologies to emerge for the next generation of the World Wide Web. Known as the Semantic Web, the focus is on information and knowledge being processable by machine in order to facilitate automation and interoperability. In combination with Grid computing, this approach is known as Semantic Grid and it has the potential to be particularly effective in the environmental context.
Floodnet is an effective demonstrator of the potential of pervasive computing in the environment, and highlights solutions to some important issues, such as adaptive sampling for prioritised data gathering and energy conservation. It has enabled innovation within the environmental application and has also provided an important case study to explore the relevant research and development issues in the computer science field, bringing together three major trends in computing — pervasive computing, grid computing and the Semantic Web — and the emerging field of autonomic computing. We plan to take this project forward, together with the other projects in the Centre, in order to provide cost-effective wireless alternatives which will enhance existing monitoring solutions and to innovate advanced monitoring solutions for future deployment.
The partners in Floodnet are ABP Marine Environmental Research Ltd, Halcrow Ltd, IBM, Multiple Access Communications Ltd and University of Southampton. The author wishes to thank Dr Craig Hutton, Floodnet Project Manager, for his assistance in the preparation of this article.
The Centre for Pervasive Computing in the Environment — see
The DTI Next Wave Technologies and Markets programme — see www.nextwave.org.uk
Self-powered Microsystems — see for example www.perpetuum.co.uk
Equator Interdisciplinary Research Collaboration — see www.equator.ac.uk
Semantic Grid — see www.semanticgrid.org
Professor David De Roure is Director of Envisense, the DTI Next Wave Centre for Pervasive Computing in the Environment, hosted at the University of Southampton. A member of academic staff at Southampton since 1987, he is currently head of Grid and Pervasive Computing in the School of Electronics and Computer Science. He is a member of the steering group of the Global Grid Forum and advises national programmes on the application of Grid and Pervasive computing. He is a Fellow of the British Computer Society.