Enhancing UAV Photogrammetry-Derived Shallow Water Bathymetry Accuracy Through Regression-Based Refraction Correction
I GD Yudha Partama, I Gede Gegiranang Wiryadi, I Dewa Gede Agung Pandawana

Universitas Mahasaraswati Denpasar

Abstract

Accurate shallow-water bathymetry is essential for coastal planning, habitat conservation, and disaster mitigation. Uncrewed Aerial Vehicle (UAV) photogrammetry offers a cost-effective method to generate high-resolution Digital Surface Models (DSM) in nearshore environments. However, optical distortions particularly from light refraction at the water surface introduce significant depth errors. This study evaluates the effectiveness of four regression-based models: Simple Linear Regression (SLR), Polynomial Regression, Generalized Additive Models (GAM), and Support Vector Regression (SVR) in correcting refraction-induced errors in UAV-derived DSMs. Ground-truth depth data were collected using Real-Time Kinematic GPS (RTK-GPS) and used to train each model, with DSM elevation as the predictor variable. A k-fold cross-validation approach was applied to assess model robustness, and performance was evaluated using Root Mean Square Error (RMSE) and Mean Error (ME). Results show that GAM achieved the lowest RMSE (0.261 m) and the smallest ME (-0.0063 m), indicating high accuracy and low bias. SLR performed comparably (RMSE = 0.262 m, ME =-0.0139 m), validating its utility as a simple yet reliable model. SVR also showed good performance (RMSE = 0.275 m), though with slightly higher bias (ME =-0.051 m). Polynomial Regression performed the poorest (RMSE = 0.736 m), suggesting its limited ability to model the complexity of refractive distortion. Spatial visualization of the corrected depth rasters confirmed the quantitative findings, with GAM and SVR producing more realistic bathymetric patterns. The study highlights the potential of non-linear and machine learning-based models, particularly GAM and SVR, to enhance the accuracy of UAV-based bathymetry in optically shallow coastal zones. These methods offer scalable, low-cost solutions for improving nearshore depth mapping and support informed coastal management decisions.

Keywords: UAV-photogrammetry, bathymetry, regression model, coastal mapping, refraction correction

Topic: Topic C: Emerging Technologies in Remote Sensing