Potential of Using the Random Forest Machine Learning Algorithm for Water Surface Interpretation

Abstract

Abstract Terrestrial water bodies serve as reliable geoindicators that enable the identification of hidden properties and structural features of geosystems based on remote sensing data. Analysis of their spatiotemporal dynamics provides valuable information for studying landscape transformation, global carbon cycles, and climate change. However, such analysis requires processing considerable arrays of multi-temporal data, making labor-intensive visual interpretation methods ineffective. Therefore, the task of automatically identifying water bodies in satellite images remains highly relevant. This article presents the results of automatic water body identification for three key areas using the Random Forest machine learning algorithm. Multispectral Landsat-8/9 images were used as the initial data. Pixel values were converted into physical characteristics of spectral reflectance of the surface. Evaluation of the classification accuracy demonstrated the high efficiency of the method: the probability of correct identification was 97.35% for water bodies and 99.52% for land. The obtained results demonstrate high reliability and confirm the effectiveness of the algorithm for automatic mapping of terrestrial water bodies.