OPTIMIZING MICROCLIMATE CONTROL STRATEGIES IN PRECISION AGRICULTURE USING MACHINE LEARNING TECHNIQUES AND IOT SENSORS DATA

The overarching aim of this PhD thesis is to push the boundaries of microclimate monitoring in precision agriculture through the application of machine learning to IoT sensor data. The sensors provide vast amounts of detailed microclimatic information, which can be effectively processed by machine learning algorithms to optimize control strategies for precision agriculture. The thesis aims to advance the state of the art in microclimate control by integrating machine learning techniques with sensor data, focusing on two key dimensions: spatial and temporal. The scientific work presented in this thesis was developed in collaboration with DNDG srl, a company specialized in the development of fully customized, high performance IoT industrial systems business solutions and digital products, and a partner of the multinational Finder Group. From a spatial perspective, this thesis aims to optimize microclimate monitoring in outdoor agricultural fields with a complex terrain conformation, such as vineyards. As will be discussed in the Introduction, microclimate monitoring can be performed using either remote sensing or networks of IoT sensors. While remote sensing provides highly accurate results at continuous spatial scales, it is expensive and does not allow for continuous, real-time monitoring. IoT sensor networks offer a more affordable solution for real-time monitoring of specific locations within a field. However, these measurements are localized and may miss significant microclimatic patterns elsewhere in the field. An ideal solution would involve deploying a large number of sensors, but this would significantly increase maintenance costs. To address this, zones with similar microclimatic patterns need to be identified, allowing for the strategic placement of sensors within each zone. The literature currently lacks a clear strategy to achieve this. Thus, there is a need for a method that can use IoT data to continuously reconstruct microclimatic variations and identify characteristic patterns. Given the mathematical complexity of modelling the physical behaviour in this task, the approach is based on data-driven methods such as neural networks and machine learning clustering algorithms. This thesis tackles the problem in two steps: (1) using neural networks trained on IoT sensor data to obtain a continuous description of microclimate variations at the meter scale across the entire field; and (2) identifying spatial areas with similar microclimate variations using a machine learning clustering algorithm. From a temporal perspective, this thesis proposes the application of neural networks tailored for time series forecasting to enhance time-based monitoring of microclimate effects on precision agriculture systems. The temporal aspect is addressed through two distinct scenarios. First, a case study is presented regarding the use of neural networks to monitor milk production data in a commercial dairy farm equipped with Automatic Milking Systems, with a focus on the influence of microclimate conditions. Milk production is highly sensitive to the microclimatic conditions experienced by dairy cows. To improve production and enhance animal welfare, precision agriculture should implement strategies that forecast milk production based on the relationship between past microclimate conditions and production data. This thesis applies a modern neural network for time series forecasting (Time-Series Mixer) to predict future milk production, demonstrating how AI can be utilized to optimize production for commercial dairy farms. The second scenario focuses on developing a strategy to improve weather forecasting for local agricultural sites, with a particular emphasis on predicting hazardous events that could impact agricultural production. Accurate knowledge of future weather conditions is crucial for optimizing risk management strategies. Numerical weather predictions generate forecasts at global or regional scales, which lack precision at the local agricultural level. This thesis advances the field by proposing the application of a modern neural network (Time-Series Mixer), specifically designed for time series forecasting, that integrates past local weather data with future numerical weather predictions. The framework has been tested on six agricultural sites in Northern Italy. In addition, precision agriculture is intricately connected to public health, especially when considering the impact of pollutants like heavy metals in water. Indeed, heavy metals absorbed by plants, after irrigation, can enter the food chain, posing significant risks to human health. Metals like chromium and lead are known carcinogens and can affect the kidney, liver, and nervous system even at low concentrations. Therefore, a critical aspect of precision agriculture is the need to address pollution in both water and soil, with a focus on continuous monitoring using big data analysis of sensor networks. These sensors provide valuable information on pollutant levels, capturing both temporal and spatial variations. Effective monitoring of environmental contaminants is vital for safeguarding public health and ensuring the sustainability of agricultural practices. This thesis tackles this challenge by investigating the temporal and spatial trends of heavy metal contamination in the waters of Lombardy, covering the period from 2017 to 2020. This analysis aims to provide a comprehensive overview of contamination patterns and identify potential risks associated with heavy metal pollution, informing strategies for improved environmental management and policy development.