Data-Driven Modelling of Soil Moisture Dynamics for Smart Irrigation Scheduling

Abstract

In the face of increasing water scarcity and uncertainties of climate change, improving crop water use efficiency and productivity, while minimizing negative environmental impacts, is becoming crucial to meet the surging global food demand. Smart irrigation has a potential of improving water use efficiency in precision agriculture especially when efficient irrigation control strategies are adopted. Conventionally, irrigation systems rely on heuristic methods to schedule irrigation which either leads to over-irrigation or under-irrigation. This influences the crop physiological characteristics as well as the water use efficiency. To tackle this menace, model-based irrigation management has been overemphasized. A closed-loop irrigation control strategy relies on a mathematical model of the system for irrigation scheduling decisions. In this study, a data-driven approach was used to learn soil moisture dynamics from a drip irrigated tomato in an open field agricultural system. A total number of 9674 data samples were collected using an ATMOS41 weather station, TERROS 12 soil moisture sensor and a YFS-201 flow sensor for crop evapotranspiration and precipitation, soil moisture and irrigation volumes respectively. Data driven modelling was then performed using the system identification toolbox in a MATLAB environment. The model formulation was a multi-input single-output (MISO) system with reference evapotranspiration, irrigation and rainfall as inputs and soil moisture as the output. Different model structures including transfer functions, state space models, polynomial models and ARX models were evaluated. Model performance was evaluated using the mean square error (MSE), final prediction error (FPE) and estimated fit of the model approaches. Simulation results indicate that the soil moisture dynamics model provides a satisfactory approximation of the process dynamics with a state space model giving an estimated fit of 97.04 %, MSE and FPE of 1.74×10−7 and 1.75×10−7 respectively. This model will be used to design a model predictive controller for smart irrigation scheduling in open field environmental conditions.