The Comparison of Machine Learning Techniques for Agricultural Land Use Classifications in the Prairies: A Case Study in Saskatchewan, Canada

Remote sensing (RS) plays a crucial role in land use classification, providing essential information to address various environmental issues. The incorporation of machine learning techniques into remote sensing, including random forests (RFs), support vector machines (SVMs), and artificial neural networks (ANN)s, has garnered significant attention due to its potential for efficient land cover classification in remotely sensed images. However, applying machine learning in the context of agricultural land classification presents challenges, with limited research exploring these techniques for this specific purpose. This study aims to investigate the performance of machine learning techniques in the southern prairie region of Saskatchewan, focusing on agricultural land classifications. Utilizing Sentinel-2 satellite images, publicly available from the European Space Agency, a total of 133,080 samples were analyzed through stratified random sampling, with 70% allocated to training and 30% to testing subsets. Accuracy assessment involved various indicators. Results indicate that random forests exhibit the highest overall accuracy, whereas support vector machines demonstrate the lowest accuracy. Artificial neural networks, on the other hand, display distinct advantages compared to other machine learning techniques. This research contributes valuable insights into the application of machine learning for agricultural land use classifications, emphasizing the need for further exploration and refinement in this challenging domain.