Can we do better than the grid survey: Optimal synoptic surveys in presence of variable uncertainty and decorrelation scales

Abstract

Regular grid (''lawnmower'') survey is a classical strategy for synoptic sampling of the ocean. Is it possible to achieve a more effective use of available resources if one takes into account a priori knowledge about variability in magnitudes of uncertainty and decorrelation scales? In this article, we develop and compare the performance of several path-planning algorithms: optimized ''lawnmower,'' a graph-search algorithm (A*), and a fully nonlinear genetic algorithm. We use the machinery of the best linear unbiased estimator (BLUE) to quantify the ability of a vehicle fleet to synoptically map distribution of phytoplankton off the central California coast. We used satellite and in situ data to specify covariance information required by the BLUE estimator. Computational experiments showed that two types of sampling strategies are possible: a suboptimal space-filling design (produced by the ''lawnmower'' and the A* algorithms) and an optimal uncertainty-aware design (produced by the genetic algorithm). Unlike the space-filling designs that attempted to cover the entire survey area, the optimal design focused on revisiting areas of high uncertainty. Results of the multivehicle experiments showed that fleet performance predictors, such as cumulative speed or the weight of the fleet, predicted the performance of a homogeneous fleet well; however, these were poor predictors for comparing the performance of different platforms.

A future for intelligent autonomous ocean observing systems

“…In recent years, several approaches have been proposed to extend planning methods for static environments to marine vehicles. Many are based on graphs and dynamic programming such as Dijkstra's method and the A* algorithm (Carroll et al 1992;Garau et al 2009;Rao and Williams 2009;Frolov, Garau, and Bellingham 2014). The key issue in the A* algorithm is the choice of a heuristic that is often difficult to define.…”

“…A similar approach is the error subspace scheme (Lermusiaux 1999a;Wang et al 2009), which keeps track of the time-evolution of a dominant low-rank approximation to the error covariance matrix using nonlinear ensembles and re-runs the ensemble after each candidate assimilation. Other approaches in this category include mixed-integer programming approaches (Yilmaz et al 2008), potential functions (Munafò et al 2011), genetic algorithms (Heaney et al 2007;Frolov, Garau, and Bellingham 2014;Heaney et al 2016), etc. Note that all of the above adopt a Gaussian approximation of the distributions of all the involved state variables, which neglects the non-Gaussian features of the statistics.…”

Marine Rapid Environmental Assessment in the Gulf of Taranto: a multiscale approach

“…MREA has been implemented globally, in several regions, and adaptive sampling has also been developed (Lermusiaux, 2007, Frolov et al, 2014. Intermittency and multiscale processes have led to the concept of nested forecasts, which make use of optimized sampling networks to increase forecast accuracy.…”

An Evolutionary Algorithm for Online, Resource-Constrained, Multivehicle Sensing Mission Planning

“…A branch-and-bound solution to the OP is used in [27] for scalar field estimation using Gaussian processes. In [28] a GA based heuristic is used to solve an OP for marine sampling. In [29] an agricultural monitoring application is shown using multiple robots for sensing.…”