Hyperspectral imaging of river bathymetry using an ensemble of regression trees

Abstract

Remote sensing has emerged as an effective tool for characterizing river systems, and machine learning (ML) techniques could make this approach even more powerful. To explore this possibility, we developed an ML-based workflow for hyperspectral imaging of river bathymetry using an ensemble of regression trees (HIRBERT). This approach involves using paired observations of depth and reflectance to select wavelength bands as predictors and then train a depth retrieval model; applying the model to the image yields a spatially continuous bathymetric map. We used data from five rivers with diverse morphologies and optical characteristics to assess whether HIRBERT can (1) provide more accurate depth estimates than a band ratio-based algorithm and (2) extend the range of depths detectable via remote sensing. Relative to single band combinations identified via optimal band ratio analysis (OBRA), regression tree ensembles improved depth retrieval performance, with observed versus predicted (OP) regression R 2 values increasing for all five sites. Similarly, HIRBERT provided more reliable depth estimates than OBRA over the full range of depths present along each river. These results suggest that by incorporating additional spectral information from multiple wavelength bands, ML could enhance bathymetric mapping across a range of river environments. In addition, we show how graphical tools can facilitate interpretation of ML-based depth retrieval models and yield insight regarding relationships between depth and reflectance. The HIRBERT workflow is packaged in free, standalone software developed to support applications in river research and management. Although ML can enhance remote sensing of river bathymetry, the limitations of this approach must also be acknowledged: Field measurements of water depth are required to train a depth retrieval model and the resulting model should only be applied to the image from which the training data were derived. The inherently image-specific nature of this approach implies that developing generalized regression tree ensembles that could be applied at larger scales would require additional research.