Machine Learning Techniques for Smart Agriculture

We present the research results that are a basis for integration of hardware devices (sensors and actuators with communication devices, computers, and web servers), data-bases, machine learning (ML) techniques, and related application software. We also describe innovative machine learning approaches and their potential impact on plant treatment and protection. The scientific and technological challenges are discussed as well. The developed machine learning techniques assist farmers in making data-driven and rule-based decisions to reduce the pesticide use and environmental harm. We merge information coming from the features, variables, and correlations extracted by the machine learning algorithms together with information generated by the previous knowledge in the existing databases. The final goal is to learn actionable rules that assist the farmer and help the decision-making process. Using a ML framework, we map scenarios and situations (meteorological data, soil characteristics, nutrients, aggregated daily temperature or features extracted by the knowledge module) to actions (for example, irrigation, dose of fertilizers or pesticides to apply). The developed models can be used to simulate the interaction with the environment and estimate rewards and the expected value function. The framework integrates the objective data, ML algorithms and application software, that can generate and provide end-user relevant information (in the case of traditional crop protection) or control information for actuators (in the case of automated crop protection).