Multi-scale adaptive sampling for mapping forest fires

Distributed monitoring applications require wireless sensors that are efficiently deployed using robots. This paper proposes to deploy sensor nodes in order to estimate the time-varying spread of wildfires. We propose a distributed multi-scale adaptive sampling strategy based on neural networks, the Extended Kalman Filter (EKF) and greedy heuristics, named "EKF-NN- GAS". This strategy combines measurements arriving at different times from sensors at different scale lengths, such as ground, air-borne or spaceborne observation platforms. We use the EKF covariance matrix to derive quantitative information measures for sampling locations most likely to yield optimal information about the sampled field distribution. Furthermore, we reconstruct the spatio-temporal forest fire spread, based on parameterized Radial Basis Functions (RBF) neural networks. To replicate the complexity involved in actual fire-spread we simulate it using discrete event cellular automata acting as our "truth model". Finally, we present experimental results with ground vehicles that navigate over a "virtual fire" projected on the lab floor from a ceiling-mounted projector to emulate a sampling mission performed by aerial robots.