Towards sustainable Food-Energy-Water security nexus of farms: Use of agriculture model for precision decision making

Abstract

World population is projected to increase and be around 8.6 Billion by year 2030. In contrast, the agriculture land is limited and significant new cultivation land is not being added to produce more food. To feed the increasing population, fields are now more intensively cultivated. Yield from field has increased by developing and breeding high yielding variety of crops, applying more fertilizer, pesticide, and water. These increased inputs, apart from their input costs, are cause of environmental pollution. Agriculture crop production releases greenhouses gases like Carbon Dioxide during soil cultivation; Methane is released by cattle and livestock; also the soil gets to be more oxygen deficient, and Nitrous Oxide is released from fertilizer and manure. Excess fertilizer pollutes the water table below field and can runoff through rain water to nearby waterbodies and streams. The nitrate form of fertilizer is most widely used by farmers. They also cause the most pollution by mixing with rainfall and irrigation water and subsequently leaching to water table below or flowing to nearby water bodies. Too much nitrate in water breeds more algae which in turn makes water deficit in Oxygen and hampers the marine population. If nitrate contaminated water is consumed then many diseases can occur, a particular disease is the blue baby syndrome where Oxygen carrying capacity of the blood is decreased.

With the above challenges in mind, our goal of this study is to come up with a decision making strategy to prescribe when and how much fertilizer and irrigation ought to be applied so that yield and profit to farmer is maximized as well as environmental pollution caused by agriculture activities is minimized. Towards this goal, we explore the spatial and temporal variable application in the field for efficient farm management. The current trend in farm is to apply fertilizer and water as recommended by general guidelines, e.g apply 150kg N per Hectare of maize field, with most of the fertilizer application occurring during middle and start of the growing season. These recommendations are for large geographic areas and not localized to specific field conditions.

The site-specific and temporal dynamics of the agriculture ecosystem is captured through agriculture model and sensor data. An agriculture system model has many interconnected components or modules. Some of the core modules are crop growth, soil nutrient dynamics, water and heat flow in soil. Each modules’ state, variables and parameters have their own governing dynamics. The agriculture models that represent the actual field have many parameters. Before using the model for decision making, their parameters need to be calibrated. We begin our study by understanding and utilizing an integrated agroecosystem model named RZWQM (Root Zone Water Quality Model). The RZWQM calibration is done with respect to a USDA experimental field in Greeley, Colorado. This study is focused on Maize grown on the experimental field. Measurement of crop and soil variables from sensor were used by us to calibrate the RZWQM model. A new automated calibration method was developed to calibrate the model which showed around five percent improvement over another calibration technique by an agriculture expert. The calibration was validated with a deficit irrigated field also in the Greeley experimental field complex. After calibration, offline optimum fertilization and irrigation recommendation was prescribed, with days of application fixed, that increased the profit for the farm by ~10\%. Three global optimization routines were applied to come up with the recommendation and their recommendation showed ~10 percent increase in profit compared to scenario of the experimental field. The automated calibration has been implemented in R and the recommendation routine implemented in Python.

A model-predictive real-time (in-season) fertilization and irrigation decision-making framework is proposed next, where the optimization steps can be repeated each day, and the recommendations for only the current day's is actually applied. The real-time decisions are dependent on accuracy of weather forecast. A mathematical formulation of the model-predictive decision-making scheme is presented. Unlike the first work above, where days of application are taken to be fixed, this framework also has input days as variables. We compare our model-predictive real-time decision-making strategy with the case of an off-line decision-making that in fact assumes the knowledge of the seasonal weather forecast, which is unrealistic.

Global optimization for decision-making takes time to converge as the algorithm needs to explore a large search space. In this respect, the execution time of the model is critical for real-time farm management. RZWQM though accurate, is slow in calibration and decision-making. This is because the calibration and recommendation routines need to make repeated system calls to RZWQM, which incurs time cost, and further suffers from many file read/write operations. This motivated us implement and use a lean model, one that can be integrated to an optimizer in the same unified framework. In this regard, a lean soil nutrient model was implemented and compared with RZWQM against the same experimental data. The lean model after calibration provided comparable output as RZWQM. The lean model implemented in our software framework is remarkably faster than the complex RZWQM, and also has been made available on the MyGeoHub cloud infrastructure.