Cabral, Pedro da Costa BritoPereira, Miriam Natasha Hadidi2023-11-102023-11-102023-10-24http://hdl.handle.net/10362/159799Dissertation presented as the partial requirement for obtaining a Master's degree in Data Science and Advanced Analytics, specialization in Business AnalyticsThis research was part of the EXPL/CTA-AMB/0165/2021 project, supported by a grant of the Portuguese Foundation for Science and Technology ("Fundação para a Ciência e a Tecnologia"): https://doi.org/10.54499/EXPL/CTA-AMB/0165/2021The research field of machine learning and supervised image classification is quickly developing. There are many studies regarding the different use cases of image classification. However, a comprehensive study on the primary algorithms in ArcGIS Pro has not been assessed for numerous classes. This study attempts to bridge that gap by evaluating the effectiveness of the three primary classification algorithms available in ArcGIS Pro, and to determine an optimal algorithm for the given study area. This scope covers 12 classes of land cover in San Joaquin County, California. Maximum Likelihood, Random Forest, and Support Vector Machine were tested based on their general usability in image classification as well as their proven characteristics through research. The training and ground truth validation data were provided by USGS, in the form of a Landsat 8 image, and crop planning map. The accuracy assessment was performed with a stratified random sampling strategy. Based on the Kappa statistic, this study determines Random Forest (Kappa = 0.68, Accuracy = 0.76) to be the most suitable algorithm for detecting a series of crop types, bodies of water, and urban spaces apart from the rest of the land cover in San Joaquin County, California, USA. In addition to determining a preferred algorithm, it is also apparent that certain parameters when tweaked, produce the optimal classifier for this dataset. In this case, this means most parameters set to default, with an increased spectral detail and a decreased spatial detail. What this indicates for crop planning is that the current algorithms used in California are already quite effective at accurately identifying unique types of land cover. This builds confidence in the field, however parameters could be similarly tweaked to produce an even better classification. This study can be useful for improving crop and water planning.engMachine LearningRemote Image SensingGeographic Information SystemsSupervised Image Classification AlgorithmsSDG 8 - Decent work and economic growthSDG 12 - Responsible production and consumptionSDG 13 - Climate actionSDG 15 - Life on landmaster thesis203384121