⚡ electrical engineering

A method of neural network identification of thermokarst formations in Central Yakutia using graphic data from satellite remote

This paper presents a convolutional neural network-based method for identifying thermokarst formations in Central Yakutia using satellite imagery, achieving high classification accuracy (0.94–0.96) across four landscape types.

Original authors: Vasilii Zhebsain, Artem Gololobov, Aiaal Poselsky, Nikolai Basharin

Published 2026-10-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vasilii Zhebsain, Artem Gololobov, Aiaal Poselsky, Nikolai Basharin

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, frozen landscapes of northern Russia, the ground is not as solid as it appears. Beneath the surface lies permafrost, soil that has remained frozen for thousands of years. However, as the climate warms, this frozen earth is beginning to thaw, triggering a process known as thermokarst. When the ice within the soil melts, the ground collapses, creating a chaotic terrain of sinkholes, depressions, and water-filled basins. In Central Yakutia, a region in eastern Siberia, these formations often take on a distinct, geometric appearance, resembling a patchwork of polygons stretching across the tundra. This transformation is accelerating rapidly, with new features forming in as little as twenty-five to thirty years. While these changes have occurred naturally for millennia, the current rate of warming is driving them forward with unprecedented speed, threatening infrastructure and altering ecosystems. Monitoring these shifts is critical, but the sheer scale of the landscape makes it difficult for humans to track every change by eye alone.

To address this challenge, a team of researchers from North-Eastern Federal University and several Russian scientific institutes has developed a new way to watch the frozen ground change. They turned to satellite imagery, using the same kind of high-resolution maps found in everyday navigation apps, but applied a sophisticated form of computer learning to analyze them. Instead of a human looking at a screen and trying to spot a depression, the researchers trained a computer program to recognize the visual patterns of thermokarst formations automatically. The goal was to teach the machine to distinguish between four specific types of landscapes: areas with these thawing formations, open fields, forests, and bodies of water. By doing so, they hoped to create a tool that could scan vast areas of the Siberian wilderness and flag the early signs of ground collapse before they become catastrophic.

The researchers began by building a library of images to teach their computer. They created an online tool that allowed them to quickly select and label thousands of small square pictures taken from satellite maps. These pictures came from two specific areas in Yakutia known for having many thermokarst features. They gathered over ten thousand images, dividing them into different sizes to see which worked best for the computer. Some images were small, measuring 70 by 70 pixels, while others were larger at 128 by 128 pixels. To ensure the computer learned correctly, they verified many of the thermokarst spots using drones, flying low over the terrain to confirm what the satellites were seeing. This careful preparation created a robust set of examples, or a training dataset, that the computer could study to learn the difference between a forest, a field, a lake, and a thawing patch of ground.

Once the data was ready, the team built a digital brain, known as a convolutional neural network, to do the recognizing. This type of computer program is designed to process images much like the human eye and brain work together. It looks at an image and breaks it down into layers, searching for simple shapes and textures first, and then combining those simple parts to recognize complex patterns. The researchers tested many different versions of this digital brain, changing the number of layers and the way it processed information. They wanted to find the specific architecture that could identify thermokarst formations with the highest possible accuracy. They ran numerous experiments, feeding the computer thousands of images and watching how well it learned to sort them into the correct categories.

The results showed that the computer could indeed learn to spot these formations with remarkable precision. After testing many different designs, the researchers identified three specific models that performed the best. These top-performing models were able to correctly identify thermokarst formations in about 94 to 96 percent of the cases they were tested on. This level of accuracy is significant because it means the system can reliably distinguish the subtle signs of thawing ground from other landscape features. The computer was particularly good at telling the difference between a forest and a lake, but it sometimes struggled to tell the difference between a thermokarst formation and an open field. This confusion makes sense, as both can look like flat, open ground from above, and the researchers noted that this mirrors the real-world difficulty in distinguishing these features without close inspection.

The study concludes that this method offers a powerful new tool for monitoring the changing Arctic. By using satellite images and trained computer programs, scientists can now scan large areas of Central Yakutia much faster and more consistently than before. The software developed by the team is ready to be used for automated monitoring, provided that the training data is kept up to date and verified. While the system is not perfect and still makes occasional errors, its ability to achieve such high accuracy suggests that machine learning can play a vital role in tracking the rapid degradation of permafrost. This approach allows researchers to keep a closer watch on the frozen ground, offering a way to understand and respond to the accelerating changes in one of the most sensitive environments on Earth.

Technical Summary: Neural Network Identification of Thermokarst Formations in Central Yakutia

Problem Statement
The paper addresses the accelerating degradation of cryolithozone landscapes in Central Yakutia, driven by global warming and manifested as distinctive polygonal thermokarst formations. While these processes have historical precedents, their rate of formation has increased dramatically in recent decades, with an average formation period of 25–30 years. The authors identify a gap in current monitoring capabilities, noting that insufficient attention has been paid to refining methods for detecting these landscape degradation processes using machine learning. Existing studies have focused on other regions (Alaska, Canada, Tibet) or specific features like sinkholes and landslides, but there is a need for automated, early-stage identification of thermokarst features specifically within the Central Yakutia context.

Methodology
The study proposes a method for the automated identification of thermokarst formations using Convolutional Neural Networks (CNNs) trained on satellite remote sensing imagery.

Data Generation: The authors developed an online application to rapidly generate training datasets by allowing users to select specific landscape areas on maps from Google and Yandex services. The study focused on the Churapchinsky district and the village of Vladimirovka in Yakutia. Thermokarst formations were verified using Unmanned Aerial Vehicles (UAVs).
Dataset Construction: Two experimental datasets were created:
- DS0: A balanced dataset containing 3,204 images (70×70 pixels) focused on thermokarst features, fields, forests, and water bodies (though the initial training subset for DS0 only included thermokarst and fields).
- DS1: An unbalanced dataset containing 10,044 images (128×128 pixels, representing 32×32 m areas) with four distinct classes: thermokarst formations, fields, forests, and water bodies.
Model Architecture: The implemented CNN utilizes a sequential architecture comprising convolutional layers, pooling (subsampling) layers, and fully connected layers.
- Operations: The network applies convolution kernels to extract features, followed by ReLU activation functions. Pooling layers reduce dimensionality.
- Training: The network was trained using the backpropagation method with the Adam optimizer. The loss function selected was SparseCategoricalCrossentropy.
- Hyperparameters: Experiments were conducted with 150 epochs. The study tested approximately 180 different model variations across 17 series, varying the number of layers and filter counts (e.g., configurations like $2\times32c, N\times64c, 128$ ).
Evaluation Metrics: Accuracy was evaluated using Recall, Precision, and F1-score, calculated based on True Positives (TP), False Positives (FP), and False Negatives (FN). Confusion matrices were used to analyze classification errors.

Key Contributions

Dataset Creation: The development of a specialized online tool for rapid dataset generation and the creation of two distinct datasets (DS0 and DS1) containing verified satellite imagery of Central Yakutia landscapes.
Model Selection: The systematic evaluation of 17 model series to identify three specific CNN architectures that yield the best performance for this specific task.
Feature Visualization: The analysis of layer-by-layer feature map generation, demonstrating that significant differences between landscape classes (thermokarst vs. fields vs. forests vs. water) become evident as early as the fourth layer of the network.

Results
Numerical experiments revealed that model performance is highly dependent on the specific architecture and the number of layers.

Dataset DS0: Initial experiments on the balanced dataset showed high accuracy, reaching 1.0 for training and 0.97–0.98 for validation subsets.
Dataset DS1: On the more complex, unbalanced dataset, the study identified three optimal model series:
1. $2\times32c, N\times64c, 128$
2. $2\times64c, N\times32, 128$
3. $2\times32c, 2\times64c, N\times128c, 128$
Accuracy Levels:
- The $2\times32c, N\times64c, 128$ series achieved maximum Recall of 0.95–0.96 with 10–12 layers.
- The $2\times32c, 2\times64c, N\times128c, 128$ series achieved a maximum Recall of approximately 0.96 with 13 layers.
- The $2\times64c, N\times32, 128$ series showed maximum Recall around 0.95 with 14–15 layers.
- Overall, the best models achieved accuracy levels (Recall, Precision, F1-score) in the range of 0.94–0.96.
Error Analysis: Confusion matrices indicated that the primary source of error was distinguishing between thermokarst formations and fields, a difficulty attributed to the actual geographical distribution and visual similarity of these features.

Significance and Claims
The paper claims that the proposed method, utilizing the identified CNN models and the developed application software, is effective for the automated monitoring of cryolithozone landscapes. The authors state that the method successfully identifies early-stage thermokarst features in satellite imagery with high accuracy. The study emphasizes that the approach is viable provided that appropriate, verified training datasets are prepared. The work contributes a practical tool for addressing the urgent need to monitor landscape degradation in Central Yakutia, offering a scalable alternative to manual observation.

More like this