PRECISION: leveraging deeP REinforCement learnIng algorithm for Sustainable IrrigatiON scheduling

Progress report for GS22-259

Project Type: Graduate Student
Funds awarded in 2022: $16,500.00
Projected End Date: 08/31/2024
Grant Recipient: Clemson University
Region: Southern
State: South Carolina
Graduate Student:
Major Professor:
Dr. Vidya Samadi
Clemson University
Expand All

Project Information


The interaction between human-engineered systems (e.g., irrigation) and natural processes needs to be modelled explicitly with an approach that can intelligently quantify the influence of incomplete/ambiguous information on decision-making processes. New developments in machine learning (ML) and artificial intelligence (AI) techniques hold great potential to unlock the value of big data for precision irrigation technology development. By directing our research on Increasing Sustainability of Existing Farming Practices, Appropriate Technology and Natural Resources/ Conservation/ Water Quality, we seek to develop an intelligent irrigation decision support system (DSS) that can seamlessly learn and optimize crop water use system. This research will leverage our existing data to pilot the development of a novel ML technique to support the next generation of irrigation DSS. Our proposed approach will estimate the daily irrigation needs of a 7-acre crop field irrigated by a center pivot system. This new system will enable a closed-loop control scheme to adapt the irrigation scheduling to local perturbations such as soil moisture variations, rainfall, temperature, and crop evapotranspiration. We propose to use Deep Reinforcement Learning (DRL) as a type of ML technique that enables an intelligent agent to learn an optimal policy to conserve irrigation water in an interactive environment by trial and error using feedback from its past actions and experiences. Our approaches will be validated on a field irrigated by a center pivot located at Clemson University's Edisto Research and Education Centre (REC), near Blackville, South Carolina.

Project Objectives:

Focusing on the interface of three SARE focus areas, i.e., Increasing Sustainability of Existing Farming Practices, Appropriate Technology, and Natural Resources/ Conservation/ Water Quality, the PRECISION project will address two main objectives: (i) Develop an intelligent irrigation DSS that will have the potential to benefit precision irrigation and rule-based irrigation scheduling, and (ii) Provide irrigation decisions that can add value to farm profitability and irrigation management.

The DRL-based irrigation scheduling software will be directly applicable in farms where identical climatic, soil conditions, and crop evapotranspiration are available. However, before its application in irrigated areas with different climate and soil characteristics, the tool's effectiveness will have to be optimized by evaluating its performance with respect to irrigation water use efficiency and yield production. To do so, a 7-acre field irrigated by a center pivot located at the Clemson Edisto REC, near Blackville, SC will be used as the demonstration site for this effort. We will focus on one crop (i.e., cotton) water use simulation. Daily climate data consisting of rainfall, average temperature, average relative humidity and solar radiation, volumetric water content data are monitored daily at 4 locations in the site since 2012. In addition, the depth and timing of applied irrigation water data are also available since 2012 that will be used to calibrate and validate/optimize the DRL models. The proposed DSS will be field tested using Clemson ongoing project at the Edisto REC by automating irrigation based on soil moisture data (monitored by a wireless sensor network). In the ongoing experiment at the Edisto REC, three irrigation treatments are being evaluated in which irrigation is automatically applied to cotton when the weighted-average soil moisture reaches either 30, 40, or 50 kPa using four replications. Soil moisture is measured using Watermark moisture sensors installed at three depths in each plot. These results and data will be used to validate the DSS model. The graduate student will spend fall 2022 and spring 2023 at the Edisto REC to work closely with an irrigation specialist (Dr. Jose Payero), to meet with the farmers and collect the data. 

The developed software will be introduced and freely shared with the Cooperation Extension programs (where climatic and soil moisture data are available) across southeast land-grant universities to test and evaluate its effectiveness and performance in other demonstration sites. The software will benefit both small and large-scale farmers by helping them to increase their seasonal net return through adequate use of irrigation water in a way that does not affect the yield.        


Materials and methods:

This project proposes a program of three major work tasks (WTs), designed to focus on the mission of the SARE program. These WTs are described in detail below.  

WT1: Data Collection and Pre-processing. The goals of this work task are to collect and pre-process the data as inputs for the DRL simulation models. Climatic data as well as soil moisture and water use data are available since 2012 at the Clemson Edisto REC. Additional climatic and water use data as well as soil moisture data will be also collected at the Edisto site during fall 2023 and spring 2024. These data will then be pre-processed to identify outliers and missing values and perform quality control before incorporating them into the models ( see page 1, step 1 in Figure 1 : Conceptual framework and workflow of proposed irrigation decision making system). Depending on the data quality different pre-processing techniques such as filling missing values using forward filling, backward filling, or linear interpolation methods (spline interpolation method) will be used to ensure that the data are of adequate quality to build the models.

Tools: We will use pandas Python package (DataFrame.interpolate() function).

WT2: Pilot the development of DRL Algorithms. Current state-of-the-art irrigation decision support tools combine predictive process-based crop models with soil moisture data (pre-defined triggers) to schedule irrigation during the growing season. We propose to enhance and automate these traditional methods by developing DRL algorithms such as Deep Q-Learning (DQL) and Double Deep Q-Learning (DDQL) as semi-supervised learning models to optimize irrigation efficiency. DRL is a framework where intelligent agents learn to perform actions in an environment to maximize a reward. The two main components of DRL system are the environment (cotton farm), which represents the irrigation scheduling problem to be solved, and the agent, which represents the DRL learning algorithm. A DRL algorithm will be developed as an optimal irrigation control algorithm that is adaptive to the spatial and temporal variability of in-field conditions and farmer objectives. This enables intelligent irrigation recommendations for focused crop to continuously improve over time and maximize irrigation water use productivity. Our proposed DRL model—illustrated in Figure 1—seeks to provide a new synergy among smart technologies, big data (climate data, on-farm sensor-based soil moisture data, etc.), and theory to assess crop water use and optimize irrigation scheduling that will provide scientific and precise information about water-use efficacy and irrigation water budget. The tool will be smartly designed to determine an optimal irrigation schedule for cotton in a field by executing a DRL routine. For a given state of total soil moisture, DRL routine will be configured to: (i) perform an action, the action comprising waiting or irrigating cotton crop, (ii) assign a reward to a state-action pair, the state-action pair comprising the given state of the total soil moisture and the action performed, and (iii) instruct an irrigation system to apply irrigation to cotton crop following the optimal irrigation schedule determined, wherein the optimal irrigation schedule comprises an amount of water and a determined time at which the amount of water should be applied.

To develop DRL algorithms, we will first use existing outreach programs (farmers and growers focus groups) at the Clemson Edisto REC to facilitate a focus group discussion with farmers and professionals to determine barriers to changes in irrigation practice that limit grower adoption of intelligent DSS tools. Farmers feedback on irrigation regime according to their experience, the criteria of irrigation water depth in each growth stage, etc. will be coded in the model as farmers objective functions. DRL algorithm acts as an irrigation agent using cotton-water use and other input data. DRL modelling will be defined and the characteristics of a farm system, as well as crop-water use that necessitate its use, will be coded in the model. In this procedure, the DRL agent controls its own strategy regarding whether to conserve or consume more water to achieve a better economic return based on an initial allocation scheme (step 2 in Figure 1). We will feed crop water use data into the DRL algorithm by defining irrigation decision criteria based on in-field conditions and farmer objectives. The result will be a computational environment that enables simulating and comparing various scenarios and demands that can be used to optimize cotton water use and irrigation scheduling. 

Figure 1 illustrates the conceptual framework of the proposed DRL system; firstly, historical data such as soil moisture and climatic data will be collected, analyzed and pre-processed as needed. The data will then be split into training and testing data sets and used as inputs into DRL. The models will be calibrated, and water use will be optimized based on farmers objectives in step 2. We will use temporal-difference learning, which is utilized in the highly successful Q-learning approach (see Watkins and Dayan, 1992), to optimize irrigation water use.

Tools: We will use open-source algorithms such as Google TensorFlow and/or PyChrono pipelines to develop DRL algorithms. Crop evapotranspiration (the amount of water lost from plants and soil through evaporation) will be then calculated from weather data (Edisto REC climate data) and predicted crop growth stage.

WT3: Extension and Outreach, and Results Dissemination.

After DRL policy optimization, the results will be disseminated as factsheets and Extension workshops to farmers and stakeholders (step 3 in Figure 1). The results of this project will be broadly disseminated by:

  • Conducting workshops to transfer the results and knowledge of this project to diverse farmers through the Clemson Cooperative Extension Water Resources program.
  • Conducting a facilitated focus group discussion with farmers and professionals and mentor farmers to practice the outcomes and recommendations of the DRL tool and incorporate their feedback to refine the tool  
  • Writing factsheets, newspaper/media article to summarize the research goals and key findings
  • Conducting in-service training for County Extension Agents to share this knowledge and information with farmers with whom they work and interact.
Research results and discussion:

WT1: Data Collection and Pre-processing are almost completed. we are now focused on WT2: Pilot the development of DRL Algorithms. We used multiple deep-learning algorithms to simulate cotton yield values predicted by Aquacrop using climate, irrigation, and crop growth stage data. Multiple deep learning algorithms such as Long-Short Term Memory (LSTM), Gated Recurrent Unit (GRU) and LSTM-Convolutional Neural Network (LSTM-CNN), Random Forest (RF), and Multi Linear Perceptron (MLP) were used. Deep learning approaches were implemented and validated on a 7-acre cotton crop field irrigated by a center pivot located at Clemson University's Edisto Research and Education Centre (REC), near Blackville, South Carolina, USA. Analysis suggested that LSTM and RF performed equally with very good to excellent performances while GRU and LSTM-CNN were the second-best performers. We conclude that deep learning algorithms are useful tools that can provide insights into how much yield to expect in an upcoming season and help farmers optimize energy, water, and fertilizers applications accordingly. 




Participation Summary

Educational & Outreach Activities

1 Curricula, factsheets or educational tools
3 Webinars / talks / presentations

Participation Summary:

Education/outreach description:

This project is developing machine learning tools to predict the cotton yield. The machine learning tools were trained with the Edisto data to predict the cotton yields. multiple algorithms were tested to understand the soil and climate influences on yield amount. we continue working and enhancing the tools. Three oral presentations were presented at the American Geophysical Union, the European Geophysical Union, and SC Water Resources conference. One manuscript is currently under review while two are in preparation. 

Project Outcomes

3 New working collaborations
Project outcomes:

The outcomes of this project will help farmers in two ways:

  1. Estimating crop yield helps farmers plan ahead for equipment needs, labor, and fuel requirements, cash-flow budgeting, etc.
  2. Forecasting crop yield is also useful for analyzing weather-related variability to guide decisions such as irrigation scheduling and crop management.
Knowledge Gained:

Equipping farm advisors with the necessary tools and knowledge to address complex issues on farms is often viewed as the most effective to support changes in attitude. we are currently working on developing a survey to understand the farmers' needs and then incorporate their needs into the algorithm. The results of our simulation will be presented to the farmers and provide them with the knowledge of algorithmic-guided irrigation amount optimization that can enhance their knowledge of the sustainable use of water resources.  



Any opinions, findings, conclusions, or recommendations expressed in this publication are those of the author(s) and do not necessarily reflect the view of the U.S. Department of Agriculture or SARE.