Abstract
Short-term monitoring of alpine glaciers is crucial to understand their response to climate change. This paper presents a low-cost multi-camera system tailored for 4D glacier monitoring using deep learning stereo-photogrammetry. Our approach integrates multi-temporal 3D reconstruction from stereo cameras and surface velocity estimation from a monoscopic camera through digital image correlation. To address the challenges posed by wide camera baselines in complex environments, we have integrated state-of-the-art deep learning feature matching algorithms into ICEpy4D, a Python toolkit designed for 4D monitoring. In a pilot study conducted on the debris-covered Belvedere Glacier (Italian Alps), our stereoscopic setup captured daily images from May to November 2022. Our approach utilized SuperPoint and SuperGlue for feature matching, resulting in a daily 3D reconstruction of the glacier terminus. Using dense point clouds with decimetric accuracy, we estimated daily ice volume loss and glacier retreat at the terminus. The total ice volume loss was approximately 63,000 m³ and the retreat was 17.8 m. Surface kinematics revealed three times higher surface velocity during the warm season than in the fall. Daily analyses revealed a significant short-term correlation between air temperature, glacier surface velocity, and ice ablation.
Introduction
Monitoring small alpine glaciers is a logistical challenge. Satellites often lack the spatial or temporal resolution to capture rapid, short-term changes, while Unmanned Aerial Vehicles (UAVs) require frequent, manual field visits. High-end sensors like Terrestrial Laser Scanners (TLS) are prohibitively expensive and difficult to deploy in rugged terrain. Furthermore, traditional photogrammetry often fails in these environments; when cameras are placed far apart to cover a wide area (a "wide-baseline" setup), the perspective change is so drastic that standard algorithms like SIFT cannot find matching points.
This research introduces a low-cost, autonomous 4D monitoring system that leverages off-the-shelf DSLR cameras and deep learning-based photogrammetry. By utilizing advanced feature-matching algorithms, the system achieves daily sub-decimetric 3D reconstructions, revealing a high-resolution link between daily temperature spikes and glacier mass loss.
Methodology and System Design
The researchers developed ICEpy4D, an open-source Python toolkit designed to automate the photogrammetric pipeline. The system consists of two main components:
- Hardware: Two low-cost units (< €2000 each) featuring Canon EOS 1200D DSLRs, Raspberry Pi controllers, and solar power, installed on the moraines of the Belvedere Glacier.
- Deep Learning Integration: Instead of traditional feature matching, the authors integrated SuperPoint (for interest point detection) and SuperGlue (a graph neural network for matching). These models are significantly more robust to the extreme viewpoint changes and lighting shifts common in wide-baseline glacier monitoring.
The system employs dual processing chains. The stereoscopic chain uses both cameras to generate daily 3D point clouds for calculating volume loss and terminus retreat. The monoscopic chain uses Digital Image Correlation (DIC) on a single camera's image sequence to track surface velocity over a larger area.
Study Site and Data Collection
The primary data consists of daily image pairs from the Belvedere Glacier collected between May and November 2022. To validate the precision of the low-cost setup, two high-accuracy UAV surveys served as "ground truth." Meteorological data from a nearby Automatic Weather Station (AWS) provided air temperature and solar radiation metrics for correlation analysis.
Key Results and Insights
The integration of deep learning was the primary enabler for this study. The wide baseline (base-to-height ratio near 1) made the scene unrecognizable to standard mathematical matching; deep learning feature matching was the only way to achieve automated daily 3D reconstruction.
Key findings include:
- High Precision: The system achieved decimetric accuracy (errors < 10 cm) compared to UAV benchmarks.
- Massive Volume Loss: The glacier lost approximately 63,000 m³ of ice at the terminus over one season, with the front retreating by 17.8 meters.
- Velocity Spikes: Surface velocity was three times higher in the warm season (May–Sept) than in the fall, with peak velocities reaching ~0.23 m/day.
- Climate Correlation: A very high Spearman correlation (0.89) was found between daily air temperature and ice volume loss, proving the system's sensitivity to immediate environmental forcing.
Conclusion
While the system is highly effective, it remains weather-dependent; approximately 20% of daily images were discarded due to poor visibility from fog or rain. Future work aims to expand spatial coverage by adding more cameras to create a "multi-view" network. Additionally, the researchers plan to transition the software pipeline away from commercial APIs to fully open-source solvers, making the entire stack free and accessible to the global research community.
Summary updated