A Multi-Agent Model to Help Managing Rainfall Variability in the Rainfed Lowland Rice Ecosystem of Northeast Thailand

Rainfed lowland rice (RLR) production is the main activity in northeast Thailand. Unpredictable droughts and coarse-textured soil are the main constraints usually cited to explain the low yields and economic poverty of this region. Past studies tried to improve the drought tolerance of rice varieties and hydrological functioning at the field level. How water is used at the farm level remains largely unknown. Consequently, it is relevant to understand the dynamic interactions between water availability and water-use in the RLR ecosystem. This article describes the development of an agent-based simulation tool based on multi-agent systems to explore adaptations of RLR cropping systems to rainfall variability. An environment representing the main biophysical entities involved in decision-making regarding water use is modeled and its hydrological functioning is verified. Preliminary simulations are presented to illustrate the model capacities. These preliminary simulations aim at evaluating the efficiency of numerous on-farm reservoirs to alleviate early drought at the vegetative stage. Simulations comparing scenarios with and without ponds show that ponds are less efficient at the beginning of the RLR cycle, when rains are still light. Pond efficiency is stable when the duration of the period separating the two peaks of RLR nursery sowings is more than 2 months. Below this threshold, ponds could not be completely refilled. The next step in the model development will consist in adding autonomous agents to simulate scenarios in which farmer agents cooperate to use water and learn collectively about its dynamics. Introduction The northeastern region of Thailand is a large plateau on sandstone, which is usually characterized by poor soils and erratic rainfall. It covers one-third of the Kingdom and is home for a third of its total population. This region is the poorest of the country and a key RLR growing area. Farmers practice RLR monocropping mostly in the wet season. A common feature across the rainfed production environment is uncertainty of water supply. Past efforts to alleviate water stress in RLR (crop improvement, irrigation, soil compaction, etc.) had limited effects. For the last 15 years, the construction of small on-farm reservoirs seems to be more successful in mitigating drought at the farm level and promoting diversification of agricultural production (Hungspreug 2001). Although water availability has been improved, farmers do not seem to fully use this opportunity to intensify agricultural production. Research is needed to identify actual water needs in local farming systems based on the use of traditional RLR cultivars and to determine appropriate improvement pathways as farmers will probably continue to grow this type of RLR in the near future. An understanding of existing patterns of water use and water users’ needs is required to improve the current situation. This article describes the development of an agent-based simulation tool relying on multi-agent systems (MAS) to explore adaptations of RLR cropping systems to rainfall variability. To do this, three complementary steps were followed: (1) identification of hydrological dynamics at the catchment level, (2) understanding farmers’ decision-making rules regarding water use in RLR, and (3) integrating both components into a multi-agent model to be used for simulating various future scenarios with stakeholders. Later on, this model will be used to promote collective learning about water management across boundaries. The article presents the successive stages of the modeling process, field studies, up to preliminary computer simulations to assess scenarios with farmers. Water resources and RLR in northeast Thailand More than 80% of the farmed area in northeast Thailand (NET) is used to grow RLR (OAE 2001). The cropping cycle starts with the beginning of the rainy season, in late April, when fields are still not flooded. Rice is first seeded in nurseries near water sources so that complementary irrigation can be applied during dry spells. Approximately one month later, rice seedlings are transplanted in flooded fields after the water table has moved up and the rivers spread out in flooded plains. Paddy fields are usually harvested after rains have stopped and the land has drained. Therefore, this agroecosystem is characterized by a low water control and farmers have to adapt the crop calendar to unpredictable rainfall distributions. Their room for maneuver for water management is very limited. This situation is made worse by disadvantageous natural conditions. The high rainfall variability causes successions of dry spells even during the rainy season. Water stress is aggravated by very coarse-textured soils with low water retention. This unfavorable natural environment is usually cited to explain the low paddy rice yields (ranging usually from 1.6 to 2.0 t ha) and the relative economic poverty of this region (Somrith 1997). Past agronomic research focused on improving rice varieties to increase their tolerance of drought (Singh et al 1996). At the same time, government agencies implemented water development projects to increase irrigated area and the availability of water resources at the farm and community levels. The results are not very satisfactory as most irrigation schemes have been underused and improperly maintained. An important reason is that farmers did not actually participate in all stages of project development (Patamatamkul 2001). Several models have also already been built to represent the biophysical environment of the RLR ecosystem. The RLRice, Rainfed Lowland Rice model (Fukai et al 2000), simulates the growth of rice varieties according to the amount of water available. The ORYZA model (Bouman et al 2001) is made up of different modules that calculate water deficiency according to soil, climate, and plant physiology. Other models propose spatial representations (Suzuki et al 2001, Kam et al 2001) that take into account hydrological conditions according to field position along the toposequence, or depend on the soil type and climatic conditions at the regional level. All these models have the same aim: to calculate the terms of a water balance to predict a level of water stress and related RLR yield loss. They were helpful in conceiving the structure of our own model. Nevertheless, the “water use” component was rarely considered in these past studies. The formalization of interactions between the water-resource and water-use dynamics requires a model allowing the representation of the diversity of water uses and water access, as well as their determining parameters. Multi-agent systems (MAS) have proved to be