Intelligent irrigation management using machine learning, sensors, and crop models (NERC EAO Doctoral Training Partnership)

Irrigation underpins agricultural productivity worldwide, enabling farmers to buffer crops against risks posed by rainfall variability and drought. However, the availability of freshwater for irrigated agriculture is increasingly constrained by rising demands for water from other sectors, shifts in water availability due to climate change, and growing awareness about the environmental impacts of irrigation.
In response to these challenges, precision irrigation tools and technologies have been developed to enable farmers to increase the productivity and efficiency of water use in their operations. Advances in on-farm sensor technologies, satellite vegetation imagery, weather forecasting, and smart irrigation machinery provide vast sources of big data that could help farmers to optimize use of limited water resources, improve crop yields, and enhance profitability (Evans & Sadler, 2008). However, the value of these datasets remains limited without associated decision analytic tools to enable farmers to analyse and operationalize data for real-time irrigation scheduling.

New developments in machine-learning and artificial intelligence techniques hold great potential to unlock the value of big data for irrigated agriculture (Adeyemi et al., 2017). Current state-of-the-art irrigation decision support tools combine predictive process-based crop models with pre-defined triggers (e.g. soil moisture targets) to schedule irrigation during the growing season. Machine-learning methods (e.g. reinforcement learning) have the potential to enhance the efficiency of these irrigation control approaches due to their capacity to learn from past experiences. In particular, machine-learning could be used to design optimal real-time control algorithms that are adaptive to spatial and temporal variability of in-field conditions and farmer objectives (McCarthy et al., 2010), enabling irrigation recommendations for individual fields to be improved continuously over time to maximise irrigation water use productivity.

Deep machine-learning techniques (e.g. artificial neural networks, support vector machines) trained on historical datasets also offer currently unexploited opportunities for real-time prediction of optimal irrigation decisions based solely on observation data from sensors. Sensor-based control could reduce reliance of existing decision support tools on process-based crop models, which require costly expert calibration and are affected by uncertainties in the representation of soil-plant-atmosphere processes (Roberts et al., 2017). Indeed, deep machine-learning approaches have been applied successfully for prediction of relevant hydrological processes such as soil moisture and groundwater levels (Karandish & Šimůnek, 2016; Sahoo et al., 2017). However, to date, there has been little research to understand the potential benefits of these methods in the context of irrigation optimization.

The overall aim of this project is to evaluate and develop novel machine-learning and artificial intelligence techniques to support the next-generation of real-time irrigation decision support tools. This is unique opportunity to conduct research that is at the forefront of the UK’s Industrial Strategy theme on AI & Data-Driven Economy, and develop skills that are in significant demand from employers in research and private industry. Key research questions and topics that will be explored are expected to include:

(1) Benefits of precision irrigation: What are the potential gains (water use reductions, improved crop yields, higher profits) from adoption of adaptive real-time irrigation control systems relative to existing rule-based irrigation scheduling?
(2) Value of data: Which types of observation or forecast data are most useful for efficient real-time optimization of irrigation decisions, and how does the added value they provide to farm profitability compare with current and projected future costs of these technologies?
(3) Comparison of methods: Can deep machine-learning methods outperform model-based irrigation control systems, and to what extent can deep learning also yield new insights to improve performance of model-based irrigation control?

The project would be ideally suited to a student with a strong quantitative background in computer science, engineering, or the physical sciences. Prior knowledge of machine-learning techniques is not a pre-requisite. Under the guidance of the expert supervisory team, the student working on this cross-disciplinary project will gain a wide breath of cross-disciplinary training in crop-water modelling, optimization and machine learning, meteorology/forecasting, and resource economics. You will also work closely with industrial partners at Netafim – a world-leading provider of drip irrigation equipment and intelligent irrigation management software. Netafim will support the project through provision of experimental and sensor datasets required to train and test new approaches for real-time irrigation management for case studies in Europe/North America.