High-frequency automatic 3D glacier monitoring using low-cost time-lapse cameras and Deep Learning algorithms

High-frequency monitoring of hardly-accessible glaciers is usually challenging. Though, it is critical for understanding and modeling infra-seasonal glacier dynamics. Fixed time-lapse camera are often used for retrieving high-frequency qualitative and quantitative information on glaciers' evolution. Nevertheless, only one camera is usually employed for estimating glacier surface velocity by Digital Image Correlation (DIC) techniques and an approximate DSM is required [1]. Using multiple cameras can step up in-situ glacial monitoring, as 3D scene reconstruction can be obtained by photogrammetry and Structure-from-Motion (SfM). Indeed, two (or more) cameras allows for estimating glacier surface flow velocity in a 3D world, volume variations, ablation and glacier terminus retreat.

This work presents a pilot study for implementing a low-cost image-based stereoscopic system for automatic high-frequency monitoring of an alpine glacier. Each hand-made monitoring station includes a DSLR camera, an Arduino microcontroller for camera triggering, a Raspberry Pi Zero with a SIM card for sending images to a remote server. The two cameras were installed in summer 2021 on each side of the Belvedere Glacier north-west terminus (Italian Alps), with a wide baseline (i.e., ∼260 m). The cameras have been operating taking daily images for one and a half year. Every day, the acquired stereo-pair was processed by SfM. Due to the wide baseline, which is typical of complex mountain environments, finding corresponding points across different viewpoints was troublesome [2]. Commercial SfM software packages based on traditional feature matching (e.g., Agisoft Metashape) failed to find enough and well distributed matches, while state-of-the-art deep learning-based algorithms for wide-baseline matching, such as SuperGlue [3], outperformed traditional feature matching. Therefore, an automatic open-source Python pipeline for finding matches, orienting image-pairs, solving Bundle Adjustment with Ground Control Points (GCPs) and building 3D point clouds was developed from scratch. Although alternative open-source solutions are under study, dense 3D reconstruction is currently carried out at every epoch by Agisoft Metashape, exploiting Python API to fully integrate dense matching in the processing pipeline. Results were validated at three epochs by UAV-based ground truth, obtaining RMSE of point clouds of ∼15 cm.

Overall, the monitoring system is simple and low-cost (less than €2000 per camera), requires minimum in-situ operations (limited to cameras’ installation and materialization of few GCPs), and an automatic 3D processing of stereo-pairs can improve in-situ glacier monitoring. Indeed, from daily point clouds, glacier volume reduction and retreat speed can be estimated by computing cloud-to-cloud distances. This, combined with surface velocities estimated by DIC, may help glaciologists to better understand glacier dynamics and quantify mass balances. The full Python pipeline will be released as open-source code, together with a documentation to make it reproducible for other study cases.

[1] Messerli, A., & Grinsted, A. (2015). Image georectification and feature tracking toolbox: ImGRAFT. Geosci. Instrum. Meth., 4(1), 23–34.

[2] Yao, G., Yilmaz, A., Meng, F., & Zhang, L. (2021). Review of Wide-Baseline Stereo Image Matching Based on Deep Learning. Remote Sens., 13(16)

[3] Sarlin, P. E., Detone, D., Malisiewicz, T., & Rabinovich, A. (2020). SuperGlue: Learning Feature Matching with Graph Neural Networks. Proc. CVPR IEEE, 4937–4946.