A wireless sensor network is a collection of tiny, autonomously powered 
devices - commonly called sensor nodes - endowed with sensing, communication, 
and processing capabilities. Once deployed over a region of interest, sensor 
nodes can collect fine-grained measurements of physical variables, like the 
temperature of a glacier or the concentration of a pollutant. To report their 
readings to one or more data sinks, sensor nodes communicate using their 
integrated radio-transceivers and build an ad-hoc - possibly multi-hop - relay 
network. Thanks to the potentially large number of nodes they are composed of 
and their ability to operate unattended for long periods of time, wireless 
sensor networks allow monitoring the environment at an unprecedented spatial 
and temporal scale. However, enabling a wireless sensor network to reliably 
report large quantities of data over long periods of time is still a 
challenging goal. In particular, since the operation of the radio is known to 
be the major factor of energy consumption on sensor nodes, limiting 
communication is crucial for increasing the lifetime of the network. On the 
other hand, meeting the requirements of wireless sensor network applications 
may require sensor nodes to collect and report large amounts of sensor 
readings. The efficient operation of a wireless sensor network thus requires 
careful scheduling of node participation in sensing and communication. Beyond 
the role that medium access control and routing protocols may play in this 
context, so-called sensor selection algorithms can provide for significant 
communication savings by identifying subsets of the deployed nodes that are 
still able to comply with the application requirements. This thesis argues for 
endowing sensor selection algorithms with the ability to dynamically adapt to 
the observed data and the local topology of the network. It offers novel 
sensor selection strategies that can continuously tune their parameters in a 
distributed fashion, thereby relying on no or only little a priori knowledge 
about the phenomena of interest. In particular, the thesis first addresses the 
sensor selection problem in the time domain by considering scenarios in which 
sensor nodes autonomously adapt their data reporting rate. To this end, nodes 
report their sensor readings along with a forecasting model, which the sink 
can in turn use to estimate future readings. Nodes can thus suppress data 
communication as long as the estimation error at the sink does not grow beyond 
a pre-specified, application-dependent threshold. Since the choice of a proper 
forecasting model is instrumental in allowing for high communication savings, 
this thesis proposes a generic and lightweight procedure to perform adaptive 
model selection on sensor nodes. The proposed algorithm concurrently maintains 
and evaluates the performance of a set of models on the nodes and lets them 
report to the sink the model with the lowest expected communication overhead. 
Second, the thesis addresses the sensor selection problem in the spatial 
domain, in particular for applications requiring spatial coverage of a region 
of interest. It proposes a novel sensor ranking strategy to efficiently 
schedule the activation of sensor nodes across the deployment region. 
Information about the local network topology is used to determine the actual 
relevance of a node for the current sensing task. The resulting ranking of the 
nodes, properly combined with randomization techniques, is then used to select 
subsets of nodes whose activation can guarantee compliance with specific 
application requirements. The sensor selection strategies proposed in this 
thesis require little memory and computational resources, and are thus 
implementable on resource-constrained wireless sensor network platforms. Their 
practical feasibility is evaluated by means of simulations and experiments on 
a small-scale deployment in the context of a concrete application scenario, 
namely the monitoring of noise pollution levels in urban environments. The 
contribution of this thesis is therefore threefold: it describes the design of 
novel temporal and spatial sensor selection algorithms; it provides an 
implementation of these algorithms on simulators and state-of-the-art wireless 
sensor network platforms; and it describes a thorough performance evaluation 
based on simulations, real-world experiments, and a specific application 
scenario. Taken together, these contributions constitute an additional step 
towards enabling long-term environmental monitoring applications using 
wireless sensor networks.