Machine Learning Application for Evaluating River Plume Spreading: Identification and Modeling

Based on machine learning (ML) methods, a new problem-solving procedure is proposed here to enhance river-plume research, including data collection, data-mining processes, and characteristic analyses. A plume shape recognition model is developed, structured on convolutional neural networks (CNNs), and is named “PlumeCatcher.” This model is shown to extract various river-plume features from satellite images. The random forests (RFs) data-mining method was applied to investigate the significance of the controller on plume development, including dynamic factors such as wind, current, water discharge, and tides, all of which are potential factors that can influence plume properties on daily, seasonal, and annual timescales. Focusing on the Magdalena River in Columbia (South America), the results of the methodology indicate that wind is the factor with maximum sensitivity dominating plume movement starting from the river mouth. Moreover, the plume features including area, direction, and spreading range are varied from short timescales (tidal cycle) to long timescales (yearly) and demonstrate the dynamic nature of plume spreading. ML methods provide an effective and convenient way to map plume features and extract and analyze their properties, which is beneficial for future scientific analyses.