This thesis investigates the application of artificial intelligence to smart farming. The Amaranthus Viridis crop has been grown within London South Bank University, and
different machine learning models have been used to evaluate the crop dataset.
A comparative analysis of the performance of the machine learning models for the datasets from the Nutrient Film Technique (NFT), Aeroponic (AER), Aggregate (AG), and Floating Hydroponic systems from the Department of Agriculture, University of Peloponnese (UP), Kalamata, Greece to predict the Onion Bulb Diameter (OBD). The dataset has been from four different hydroponics systems, namely Aggregate (AG), Aeroponics (AER), Floating and Nutrient Film Technic (NFT). The onion crop has been grown for 92 days after transplant from the nursery.
Artificial intelligence subsets, such as machine learning and deep neural networks, have been used to evaluate the Amaranthus Viridis crop. The centralised smart farm network models evaluate the provided dataset while sharing the data with the server during training. The decentralised network for a smart farm has been considered, a scenario where the raw dataset has yet to be shared with the server during the dataset evaluation. Federated learning models allowed the researcher to train the models and make predictions of the dependent variables.
Smart farming involves applying information and communication technology to the traditional farm system. It implies the use of the Internet of Things (IoT), edge and cloud computing, and centralised and decentralised machine learning models for predictions of the farm produce. A survey of the existing smart farms identifies the various challenges experienced by farmers, which include low technological know-how of smart farm techniques, inadequate infrastructure provision, computational power issues with technological devices, poor internet facilities, high latency within the existing internet connectivity in the farms, insufficient human, technical skills capacity to manage smart farm operations, security of smart farm data during transmission, data reliability, the communication cost of smart farm network.
The floating hydroponic system involves planting the crop on the water surface, and wool rock holds the crop within the pot, allowing its roots to touch the water beneath as it
The Onion Bulb Diameter (OBD) dataset for four different hydroponic systems, namely Floating, AER, AG, and NFT hydroponic systems, has been analysed using the XGBoost, Linear regression, Deep Neural Network (DNN), and Federated Split Learning (FSL) models for their predictive and interpretive abilities. To automate the predictions of the OBD, machine learning models have been used to predict the OBD for days not considered manually within the crop life cycle. The model helps the farmers determine the harvest even before the commencement of the new planting season based on the previous planting season dataset. The developed models have been used to analyse the Amaranthus Viridis Leaves image
dataset comprehensively. The Convolutional Neural Network (CNN) model has been used to determine the percentage of the predicted Amaranthus leaves that match the original images from a hydroponic smart farm. The CNN forecasted a higher accuracy than the KNearest Neighbour, Support Vector classifier and Decision Tree model.
The crop growth rate was analysed using machine learning algorithms. The XGBoost models produced higher prediction values for the crop growth rate than the Theil-Sen, Decision
Tree, Support Vector, Quantile and K-Neighbours regressors.
The dataset is shared with the server during training in a centralised network. The federated learning models train the dataset from the smart farm network without having access to the raw dataset with the server. The federated averaging model has been used to investigate the prediction of crop types using climatic parameters as the independent variables. The hyper-tuned federated Learning model predicted the crop chickpea and achieved high accuracy.
The federated Learning smart farm network emulation was used to compare the client, server and centralised node performance. The Server model converged faster than the edge node models, classical centralised models,since it uses all the combined weights from the edge nodes to aggregate the combined models before sending them to the edge node for further training of the raw dataset. The results show that the combined decentralised network model produced a higher accuracy than the classical centralised model.