Improving UAV-Derived Nearshore Bathymetry Using Geospatial Machine Learning

Abstract

Accurate coastal bathymetry is essential for ocean science and blue economy applications. However, nearshore regions remain undersurveyed due to dynamic environmental conditions and the limitations of traditional multibeam echosounders (MBES). This study presents a framework that integrates high-resolution uncrewed aerial vehicle (UAV) multispectral imagery with geospatial machine learning to address spatial variability in reflectance-depth relationships and to mitigate performance overestimation by explicitly accounting for spatial autocorrelation. The workflow combines exploratory spatial data analysis, spatially explicit model development using the geographical random forest (GRF) algorithm, and comprehensive evaluation through global metrics and local spatial analysis. The framework was applied to a sand-bottom coastal area in southwestern Portugal, using UAV-derived multispectral reflectance as predictors and MBES measurements as ground truth. The GRF model captured spatially varying relationships between spectral bands and depth, outperforming conventional random forest in both predictive accuracy and interpretability, though at a higher computational cost. Spatial analysis indicated that the red band was the most influential predictor at the global scale, while the near-infrared band dominated in shallow areas. The green band proved most relevant within the depth range where predictions were most accurate. Furthermore, the results identified a practical optical depth limit of approximately 5 m in the study area, beyond which prediction reliability rapidly declined due to signal attenuation. These findings indicate that spatially explicit modeling yields improved insight into depth-dependent prediction behavior and feature relevance, supporting more accurate and regionally consistent coastal bathymetric mapping. This framework is reproducible and applicable to other coastal environments.