Investigations Into hyperspectral data analysis techniques for efficient monitoring of vertical hydroponic farms

• Main Author: Srinivasan, Nagarajan
• Other Authors: Murukeshan Vadakke Matham
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Machine learning; Hyperspectral imaging; Artificial intelligence; Data science; HSI data analysis; Computer vision; Robotics; Industry 4.0; Agri 4.0; Real world application
• Online Access: https://hdl.handle.net/10356/180777
• Institution: Nanyang Technological University

Ensuring food security in a country is crucial especially in land-scare nations where traditional agricultural partices face challenges. Vertical hydroponic farms offer smart and sustainable solutions for ensuring year-round crop production in land and water scarce regions. Automated monitoring of the such farms are essential to enable accurate and efficieint strategies. In this context, Hyperspectral imaging (HSI) technology serves as a powerful inspection tool that can provide useful and vast information on hydroponically grown crops. However, handling and analysis of huge HSI data from an unstructured environment such as large farms efficiently is quite challenging and time consuming. In this context, this research primarily focuses on creation of suitable machine learning (ML) models for effective automated prediction of crop stress across its growth period. The main objective of this dissertation is to develop and evaluate data analytics techniques suited for analyzing big data generated by HSI in vertical hydroponic systems. This involves analysis of underlying patterns in the data, based on which automated methods for training datasets were developed for creation of different customized ML models. Various data science methods including machine learning, deep learning, computer vision (CV) and Spectral Angle Mapper (SAM) were explored to identify which technique best suits the requirement of automated plant stress detection. Different ML models and their combinations were evaluated for their performance in terms of accuracy, precision, recall, f1 score, confusion matrix results, time taken, and ability to handle HSI data, binary and multi-class classification capability. Among the ML models explored, XGBoost model exhibited excellent performance achieveing test accuracy of 99.52% and f1 score of 96.71% . The performance of XGB superseded over other algorithms such as Random Forest (RF), SVM, Adaboost and Bagging ensemble models with SVM and RF models as base estimators. An algorithm for automated creation of trainable HSI datasets was developed using SAM and CV techniques as part of this research. These datasets created were used to train and optimize the models for suitable performance. The best ML models developed demonstrated early, stable, and consistent nutrient stress predictions over the plant growth periods, overcoming the challenges of handling complex HSI data and early stress detection. The models developed were tested and integrated in real-world application scenarios and the models were tested in practical settings, with outcomes measured in terms of training times, performance, and scalability. The methods and frameworks developed can be extended to other applications, to handle, achieve, and create custom data science solutions involving proximal HSI data. This research is expected to contribute to increased farm productivity, yield, while supporting the automation of smart plant monitoring systems, offering a sustainable solution for near future.