1:45 PM - 2:00 PM
[ACC26-01] Mapping Glacial Lakes in the Tienshan ranges, Kyrgyzstan, Using Deep Learning
Keywords:Glacial lake outburst flood, Deep learning, Remote sensing
Introduction
GLOFs caused by the outflow of numerous small glacial lakes in the Kyrgyz Tien Shan Mountains in Central Asia have been reported (Erokhin et al., 2008, 2017; Daiyrov et al., 2018, 2022). The presence of glacial lakes in the region poses a monitoring challenge due to their short lifespan (Narama et al., 2010, 2018). High-resolution satellite images are necessary for accurate tracking. Current glacial lake monitoring mainly relies on remote sensing technology, which enables wide-area observations. A major obstacle in using this technology for glacial lake identification is the challenge of accurately detecting water bodies. Typically, a normalized index is employed for automatic extraction of water areas, with a determined threshold value. However, conventional rule-based image processing techniques struggle to determine the threshold value due to the complex mountainous terrain and natural conditions surrounding glacial lakes. Deep learning has been recently introduced to solve this problem (J.E. Ball et al., 2017). Glacial lakes in the Indian Himalayan region and Greenland have been accurately segmented and classified using deep learning techniques (Qayyum et al., 2020). However, the glacial lakes in the study region are smaller and have a different surrounding environment compared to glacial lakes in other regions. Therefore, in this study, we developed a model to classify glacial lake areas using satellite imagery and deep learning. We focused on small glacial lakes in the target area and used optical satellite imagery for accurate and frequent data collection. The model was tested by measuring the size of glacial lakes, and then the lakes were mapped in the area where the model was meant to be used.
Method
Our research focuses on the Kyrgyz and Teskey Ranges, which are situated in the Tien Shan Mountains of Kyrgyzstan. Over 300 glacial lakes have been identified in the Teskey Range alone (Daiyrov et al., 2018), and the most recent GLOF event occurred in its eastern region (Narama et al., 2010a, 2018).
We utilized optical satellite images obtained from PlanetScope during the period from 2017 to 2023. A total of 1300 datasets were generated from the optical images, with 80% used for training and 20% for validation. We used an image size of 256×256. The widely-used U-Net model structure (Ronneberger et al., 2015) was utilized for image segmentation. In order to assess the accuracy of the model, we calculated the evaluation index for each pixel and compared the predicted model results with the actual glacial lake area measured by the UAV.
Evaluation of the model's accuracy and issues
It was found that the model accuracy was about 97%, precision was about 85%, and recall was about 80%. Figure 1 displays the model predictions for the glacial lake areas based on the learned model. In general, the glacial lake detection was mostly accurate. All the glacial lakes had high values, and the predicted results matched the actual lake areas. The data set was generated from the glacial lakes data set. The small number of glacial lake areas in the dataset, which accounted for only 3% of the image, may have caused an imbalance in the data labels during model training. This issue could be attributed to the small number of glacial lake areas, causing an imbalance in the data labels. In order to make the model more accurate in the future, we will focus on reducing the number of missed glacial lake regions by adjusting the data imbalance, such as using a weighted loss function.
GLOFs caused by the outflow of numerous small glacial lakes in the Kyrgyz Tien Shan Mountains in Central Asia have been reported (Erokhin et al., 2008, 2017; Daiyrov et al., 2018, 2022). The presence of glacial lakes in the region poses a monitoring challenge due to their short lifespan (Narama et al., 2010, 2018). High-resolution satellite images are necessary for accurate tracking. Current glacial lake monitoring mainly relies on remote sensing technology, which enables wide-area observations. A major obstacle in using this technology for glacial lake identification is the challenge of accurately detecting water bodies. Typically, a normalized index is employed for automatic extraction of water areas, with a determined threshold value. However, conventional rule-based image processing techniques struggle to determine the threshold value due to the complex mountainous terrain and natural conditions surrounding glacial lakes. Deep learning has been recently introduced to solve this problem (J.E. Ball et al., 2017). Glacial lakes in the Indian Himalayan region and Greenland have been accurately segmented and classified using deep learning techniques (Qayyum et al., 2020). However, the glacial lakes in the study region are smaller and have a different surrounding environment compared to glacial lakes in other regions. Therefore, in this study, we developed a model to classify glacial lake areas using satellite imagery and deep learning. We focused on small glacial lakes in the target area and used optical satellite imagery for accurate and frequent data collection. The model was tested by measuring the size of glacial lakes, and then the lakes were mapped in the area where the model was meant to be used.
Method
Our research focuses on the Kyrgyz and Teskey Ranges, which are situated in the Tien Shan Mountains of Kyrgyzstan. Over 300 glacial lakes have been identified in the Teskey Range alone (Daiyrov et al., 2018), and the most recent GLOF event occurred in its eastern region (Narama et al., 2010a, 2018).
We utilized optical satellite images obtained from PlanetScope during the period from 2017 to 2023. A total of 1300 datasets were generated from the optical images, with 80% used for training and 20% for validation. We used an image size of 256×256. The widely-used U-Net model structure (Ronneberger et al., 2015) was utilized for image segmentation. In order to assess the accuracy of the model, we calculated the evaluation index for each pixel and compared the predicted model results with the actual glacial lake area measured by the UAV.
Evaluation of the model's accuracy and issues
It was found that the model accuracy was about 97%, precision was about 85%, and recall was about 80%. Figure 1 displays the model predictions for the glacial lake areas based on the learned model. In general, the glacial lake detection was mostly accurate. All the glacial lakes had high values, and the predicted results matched the actual lake areas. The data set was generated from the glacial lakes data set. The small number of glacial lake areas in the dataset, which accounted for only 3% of the image, may have caused an imbalance in the data labels during model training. This issue could be attributed to the small number of glacial lake areas, causing an imbalance in the data labels. In order to make the model more accurate in the future, we will focus on reducing the number of missed glacial lake regions by adjusting the data imbalance, such as using a weighted loss function.