Presentation Information
[ACC45-P11]High-Frequency Monitoring of Short-Lived Glacial Lakes in the Tienshan Mountains Using Deep Learning
*Okamura Takemasa1, Chiyuki Narama1, Yusuke Iida1, Ryuken Uda2 (1.Niigata University, 2.Niigata Univ., Graduate Student)
Keywords:
Short-lived glacial lake,Deep learning,Digital disaster reduction,High-resolution satellite images
1. Introduction
In the Tien Shan mountain range of the Kyrgyz Republic in Central Asia, glacial lake outburst floods (GLOFs) caused by the drainage of small glacial lakes have been reported in recent years. Many of these glacial lakes are of the ‘short-lived glacial lake’ type, which form and drain within a short period of several months (Narama et al., 2018). Even within the same mountain range, cases of dozens of short-lived glacial lakes appearing annually have been observed. On the other hand, GLOFs involving large-scale outflows have been observed at a frequency of once every few years.
al., 2018), with dozens of such short-lived glacial lakes observed annually within the same mountain range. Conversely, GLOFs involving large-scale outflows occur every few years, causing extensive damage to downstream settlements and infrastructure. Therefore, from a disaster mitigation perspective, establishing a monitoring system capable of frequently and widely detecting small glacial lakes that fluctuate rapidly and detecting anomalies at an early stage.
Traditional glacial lake monitoring using remote sensing has often relied on medium-resolution satellite data (10–30 m scale) and manual visual interpretation. Limitations such as missed detections of small lakes, boundary errors, and constraints on update frequency have been noted. Particularly in mountainous regions, strong terrain shading and seasonal snow/ice cover induce false detections, complicating continuous, large-scale monitoring. Recent advances in high-resolution commercial satellites (e.g., PlanetScope) and automated methods based on deep learning offer promising new solutions to these challenges. High-resolution data readily captures the spatial characteristics of minute water bodies, while deep learning enables improved reproducibility through pixel-level segmentation.
spatial characteristics of small water bodies, while deep learning can enhance reproducibility through pixel-level segmentation.
However, merely combining high-resolution data and models is insufficient for constructing a monitoring system robust enough for practical operation.
Robustness against noise factors such as clouds, shadows, and snow cover, generalisation performance across regions, and operational aspects like labelling costs and processing time must be simultaneously satisfied.
Against this backdrop, this study targets the Teskey Mountains within the Tianshan Mountains. Using PlanetScope satellite imagery as the primary data source, it aims to verify the feasibility of high-frequency, wide-area monitoring. This is achieved by developing a small glacial lake detection model based on the U-Net architecture and automating its inference pipeline. Specifically, specifically, it aims to incorporate multi-band inputs including NIR (near-infrared), balance detection accuracy with operational feasibility, and establish a time-series analysis foundation contributing to GLOF early detection.
2. Methods
The target area is the Teshkey Mountains in the Kyrgyz Republic, utilising PlanetScope satellite imagery as the primary data source. Most recently, GLOFs occurred in this region from late June to July 2025. This study aims to develop methods for detecting and monitoring such sudden events. Approximately 200 datasets were created from optical satellite imagery, with 80% allocated for training and 20% for validation. Each dataset comprises satellite imagery and ground truth (GT) indicating glacial lake areas, with an image size of 256×256 pixels. The U-Net architecture (Ronneberger et al., 2015) was adopted for the model. Initially, preprocessing including saturation adjustment and shadow region detection was applied to the 3-band input. To improve accuracy, a 4-band input including near-infrared (NIR) and data augmentation (including addition of non-lake areas) were introduced.
3. Results
Figure 1 shows the glacial lake area prediction results using the trained model (4-band input). The U-Net model combining 4-band input including NIR with data augmentation demonstrated a clear improvement in detection performance for small glacial lakes compared to the conventional 3-band model. The automated inference pipeline enabled batch processing of large-area data, demonstrating the feasibility of weekly to daily updates.
. However, false detections under strong shading or snow cover conditions, and limitations in generalisation performance across regions persist. These issues require addressing through additional data and the introduction of time-series analysis.
In the Tien Shan mountain range of the Kyrgyz Republic in Central Asia, glacial lake outburst floods (GLOFs) caused by the drainage of small glacial lakes have been reported in recent years. Many of these glacial lakes are of the ‘short-lived glacial lake’ type, which form and drain within a short period of several months (Narama et al., 2018). Even within the same mountain range, cases of dozens of short-lived glacial lakes appearing annually have been observed. On the other hand, GLOFs involving large-scale outflows have been observed at a frequency of once every few years.
al., 2018), with dozens of such short-lived glacial lakes observed annually within the same mountain range. Conversely, GLOFs involving large-scale outflows occur every few years, causing extensive damage to downstream settlements and infrastructure. Therefore, from a disaster mitigation perspective, establishing a monitoring system capable of frequently and widely detecting small glacial lakes that fluctuate rapidly and detecting anomalies at an early stage.
Traditional glacial lake monitoring using remote sensing has often relied on medium-resolution satellite data (10–30 m scale) and manual visual interpretation. Limitations such as missed detections of small lakes, boundary errors, and constraints on update frequency have been noted. Particularly in mountainous regions, strong terrain shading and seasonal snow/ice cover induce false detections, complicating continuous, large-scale monitoring. Recent advances in high-resolution commercial satellites (e.g., PlanetScope) and automated methods based on deep learning offer promising new solutions to these challenges. High-resolution data readily captures the spatial characteristics of minute water bodies, while deep learning enables improved reproducibility through pixel-level segmentation.
spatial characteristics of small water bodies, while deep learning can enhance reproducibility through pixel-level segmentation.
However, merely combining high-resolution data and models is insufficient for constructing a monitoring system robust enough for practical operation.
Robustness against noise factors such as clouds, shadows, and snow cover, generalisation performance across regions, and operational aspects like labelling costs and processing time must be simultaneously satisfied.
Against this backdrop, this study targets the Teskey Mountains within the Tianshan Mountains. Using PlanetScope satellite imagery as the primary data source, it aims to verify the feasibility of high-frequency, wide-area monitoring. This is achieved by developing a small glacial lake detection model based on the U-Net architecture and automating its inference pipeline. Specifically, specifically, it aims to incorporate multi-band inputs including NIR (near-infrared), balance detection accuracy with operational feasibility, and establish a time-series analysis foundation contributing to GLOF early detection.
2. Methods
The target area is the Teshkey Mountains in the Kyrgyz Republic, utilising PlanetScope satellite imagery as the primary data source. Most recently, GLOFs occurred in this region from late June to July 2025. This study aims to develop methods for detecting and monitoring such sudden events. Approximately 200 datasets were created from optical satellite imagery, with 80% allocated for training and 20% for validation. Each dataset comprises satellite imagery and ground truth (GT) indicating glacial lake areas, with an image size of 256×256 pixels. The U-Net architecture (Ronneberger et al., 2015) was adopted for the model. Initially, preprocessing including saturation adjustment and shadow region detection was applied to the 3-band input. To improve accuracy, a 4-band input including near-infrared (NIR) and data augmentation (including addition of non-lake areas) were introduced.
3. Results
Figure 1 shows the glacial lake area prediction results using the trained model (4-band input). The U-Net model combining 4-band input including NIR with data augmentation demonstrated a clear improvement in detection performance for small glacial lakes compared to the conventional 3-band model. The automated inference pipeline enabled batch processing of large-area data, demonstrating the feasibility of weekly to daily updates.
. However, false detections under strong shading or snow cover conditions, and limitations in generalisation performance across regions persist. These issues require addressing through additional data and the introduction of time-series analysis.