Information-based sampling for spatiotemporal field estimation and reconstruction in environmental monitoring

This dissertation addresses the near-optimal deployment problem of robot-sensory nodes in a spatiotemporal field. With limited resources, monitoring of a complex environment may face serious challenges in providing sufficient information for spatiotemporal signal estimation and reconstruction. It is therefore essential to retrieve most useful information from sampling locations while using a small number of sensor nodes. In this dissertation, three aspects are investigated to overcome the shortcomings of the existing information-based sampling methods. First, a sensor node deployment method is designed to find the minimum number of sensor deployment locations while achieving near-optimal field estimation error. To this end, a sampling-based field exploration method is used to find near-optimal sampling locations over an infinite horizon. Moreover, spatiotemporal correlations of the sampling data are studied to find redundant signals. The corresponding sampling locations of the redundant signals are eliminated concerning the network connectivity. Second, a deep reinforcement learning approach is proposed to accelerate the field exploration. Typically, field exploration methods are heavily dependent on random sampling, which has low efficiency. To avoid unnecessary or redundant sampling locations, observations from the sampling locations are utilized. Then a model-based information gain determination of the sampling locations is developed to evaluate the effectiveness of the approach. The proposed method can determine the informativeness of the spatiotemporal field by learning the information gain from the sampled area. The mobile sensory agents are then encouraged to take more samples in the area of higher information gain. Consequently, the spatiotemporal field can be efficiently explored. Moreover, the selected sampling locations can near-optimally reconstruct the spatiotemporal field using statistical methods. Third, a deep learning framework is designed to provide accurate reconstruction and prediction of the spatiotemporal field, using a limited number of observations. Nonlinear mapping from limited observations to the entire spatiotemporal field is needed in a sufficiently large spatiotemporal field. Hence, a deep learning method is proposed to extract sparse representations of the field and their nonlinear mappings. It is also proven that the proposed framework obeys Lipschitz continuity and that the observations collected by sparse representations are sufficient for spatiotemporal field reconstruction.Applied Science, Faculty ofMechanical Engineering, Department ofGraduat