A hierarchical partitioning method for optimizing irrigation strategies

Improving water efficiency in irrigated agriculture is a priority for better environmental and economic performance. It can be achieved by reducing the amount of water used and optimizing the timing of application. Among numerous methods, hybrid biophysical/decisional simulation models are effective tools for evaluating and comparing irrigation strategies under different weather conditions. However, optimizing irrigation strategies for some specified agro-environmental expected criteria is a computationally hard problem. In this paper we propose a new approach for optimizing the parameters (quantities, dates, thresholds, etc.) of maize irrigation strategies represented by structured decision rules and simulated with the management-oriented model MODERATO. We introduce a stochastic simulation optimisation method, called Partition-2P (P2P). P2P is designed to completely explore by sampling the domain of the strategy parameters. This exploration is based on a hierarchical decomposition of the domain, heuristically guided toward optimal regions, and allows to consider high-dimensional optimisation problems. We first evaluate P2P by comparing it to a systematic grid search on a simple two-parameter problem, in order to optimise the expectation of the direct margin over 49 years of weather records. We obtain very similar results that confirm the sound behaviour of P2P. We then apply our approach to a more complex optimisation problem that involves an eight-parameters strategy, for which the systematic grid search method is not effective. The best strategy we obtain shows a 28 Euro ha(-1) increase in margin compared to a basic strategy proposed by irrigation advisors. The P2P algorithm is also used in three specific hydraulic contexts concerning the available flow rate. The different optimal parameter obtained for each context agree well with the expert knowledge of irrigation advisors. We finally conclude by discussing some limitations and possible improvements of the presented work. (C) 2003 Elsevier Ltd. All rights reserved.