Optimizing SfM workflows for continuous river bank monitoring: evaluating image alignment accuracies across diverse environmental conditions

River bank erosion represents a dynamic geomorphic hazard, particularly in meandering channels where migration rates threaten critical infrastructure and agricultural land. While our previous work on the Sajó River (Hungary) established a novel, low-cost monitoring framework utilizing Raspberry Pi (RPi) cameras for near-continuous observation, the reliability of photogrammetric reconstruction under uncontrolled outdoor conditions remains a critical challenge. This study presents a systematic evaluation of the accuracy constraints inherent in automated Structure-from-Motion (SfM) processing pipelines, with a specific focus on optimizing image alignment across a wide range of scene conditions.

To determine the robustness of RPi imagery, we conducted a comprehensive sensitivity analysis of the SfM-based image alignment phase. We systematically tested over 120 variations of processing parameters, manipulating keypoint and tie-point limits, upscaling factors, and masking strategies. The implementation of rigorous masking was critical, as the imagery is geometrically challenging: the moving river surface in the foreground and the sky in the background occupy the majority of the field of view, leaving only a narrow, static fraction of the image relevant for reliable 3D reconstruction. These combinations were evaluated against a dataset representing the full range of environmental variability, including clear, cloudy, dark, foggy, overexposed, and rainy conditions, as well as distinct hydrological states such as low flows, flood events, and snow cover.

Preliminary results indicate that a specific balance of 30,000 keypoints and 5,000 tie points (ratio 6.0) optimizes reconstruction fidelity, achieving an RMS error of 0.75 pixels under clear weather conditions. Notably, the system demonstrated unexpected robustness in low-light scenarios, maintaining consistent error margins of 1.17–1.18 pixels across various configurations. Conversely, scaling up these limits beyond the optimum yielded diminishing returns, confirming that higher computational loads do not necessarily equate to improved geometric accuracy. Furthermore, we applied gradual selection algorithms to filter sparse point clouds, removing unreliable points based on reconstruction uncertainty to isolate the most geometrically valid features.

The crucial final phase of this research bridges the gap between digital reconstruction and physical reality. We validate the optimized SfM-based point clouds by comparing them directly against high-precision Terrestrial Laser Scanning (TLS) data acquired during two previous campaigns and upcoming field surveys. This multi-temporal comparison allows us to quantify specific error margins for volumetric and horizontal material displacement calculations. By defining these accuracy constraints, we establish a validated protocol for calculating erosion volumes during high-flow events, ensuring that automated, low-cost monitoring systems can provide actionable, high-precision data for river management even under adverse environmental conditions.

---

The research was funded by the DAAD-2024-2025-000006 project-based research exchange program (DAAD, Tempus Public Foundation).