Machine Learning Applied to Reach Classification in a Northern Sweden Catchment

• Main Author: dos Santos Toledo Busarello, Mariana
• Format: Bachelor Thesis
• Language: English
• Published: Umeå universitet, Institutionen för ekologi, miljö och geovetenskap 2021
• Subjects: machine learning; geomorphology; random forest; channel type; Computer Sciences; Datavetenskap (datalogi); Natural Sciences; Naturvetenskap; Oceanography; Hydrology and Water Resources; Oceanografi; hydrologi och vattenresurser; Computer and Information Sciences; Data- och informationsvetenskap; Northern Sweden
• Online Access: http://urn.kb.se/resolve?urn=urn:nbn:se:umu:diva-184140

An accurate fine resolution classification of river systems positively impacts the process of assessment and monitoring of water courses, as stressed by the European Commission’s Water Framework Directive. Being able to attribute classes using remotely obtained data can be advantageous to perform extensive classification of reaches without the use of field work, with some methods also allowing to identify which features best described each of the process domains. In this work, the data from two Swedish sub-catchments above the highest coastline was used to train a Random Forest Classifier, a Machine Learning algorithm. The obtained model provided predictions of classifications and analyses of the most important features. Each study area was studied separately, then combined. In the combined case, the analysis was made with and without lakes in the data, to verify how it would affect the predictions. The results showed that the accuracy of the estimator was reliable, however, due to data complexity and imbalance, rapids were harder to be classify accurately, with an overprediction of the slow-flowing class. Combining the datasets and having the presence of lakes lessened the shortcomings of the data imbalance. Using the feature importance and permutation importance methods, the three most important features identified were the channel slope, the median of the roughness in the 100-m buffer, and the standard deviation of the planform curvature in the 100-m buffer. This finding was supported by previous studies, but other variables expected to have a high participation such as lithology and valley confinement were not relevant, which most likely relates to the coarseness of the available data. The most frequent errors were also placed in maps, showing there was some overlap of error hotspots and areas previously restored in 2010.