Design of a Neural Network development framework for plant monitoring applications

This thesis is part of the promising and fast-growing technology aspect of the "agri-food" sector. With the world's population growing rapidly and climate change posing always increasing risks to farming, food production must be made more efficient and secure. Much work has already been performed in this research field, especially by the MINES Research Group of the Politecnico di Torino under "the Plant Project". The project goal is to develop a self-sustaining monitoring system to be deployed on the field so that the plants that are being cultivated can be supervised autonomously, and the farmers can be remotely provided with the health status of their crops. This approach allows for a more efficient and timely response to negative factors affecting the plants. The scope of this thesis, specifically, is to study and develop machine-learning-based algorithms to understand plant status. As part of the thesis' efforts, a framework was developed in the Python language to test different architectures of machine learning and data processing algorithms to discover the best solutions to link environmental and impedance data about plants to their health status. In particular, data about tobacco plants grown and monitored at the Politecnico di Torino laboratories has been employed. The developed software takes advantage of the data gathered by the MINES Research Group of the Politecnico di Torino, with a large and continuously growing dataset of environmental and impedance data of the tobacco plants. Software developed by the MINES Research Group can generate easy to handle CSV files with all the data needed to explore the various solutions and train and validate the neural networks. This data has been dispensed to the tool developed under this thesis work and, along with different configuration files describing how the data should be handled and processed, the topology of the Neural Network as well as how it should be trained, the program can train a Neural Network as specified and save the statistics associated with it. The included statistics consist of Mean Square Loss and Root Mean Square Error over the test subset or the whole dataset, as well as visual representations of the training process such as the evolution of the errors over the iterations and graphs of the predictions versus the actual data that can be used to check how the model behaves over all the available datasets. To facilitate the exploration of the various architectures so that the code does not have to be changed for every training run, the software can be executed from a Command Line Interface by providing the various settings files as parameters. It is also possible to specify if the training should be a one-off effort or if multiple training attempts should be performed by sweeping a specific setting to find their optimal values without running multiple training series one at a time. This paradigm allows the user to search for the best parameters, leading to better predictions. The results of the training efforts with the available data are presented, along with findings of the best Machine Learning and data processing architectures.