A. A. Yuriev

This paper presents the results of a comprehensive study of a catastrophic karst-suffosion sinkhole that occurred on September 19, 2024, in the Khadakhan-Melkhituy karst massif (Southern Angara region) and resulted in a human fatality. Using UAV‑derived data, the morphometric parameters of the sinkhole were determined (depth 9.7 m, volume 412 m³), and a comparison was made with historical monitoring data from the 1980s. Laboratory tests performed on samples collected from the sinkhole walls revealed that the 9.7‑m thick cover sequence consists of structurally unstable loess‑like loams and sandy loams, with anomalously high silt content (up to 84 %) and porosity (up to 64 %). The interaction between karst and suffosion processes plays a key role in the sinkhole formation. The long‑term (57‑year) backwater effect of the Bratsk Reservoir and its cyclic level fluctuations caused leaching of sulfate‑carbonate rocks, promoting mechanical removal of silt‑sized particles from the cover sequence by percolating water. A two-stage mechanism of karst-suffosion sinkhole development was identified: from the formation of a loosening zone above an ancient karst cavity to the instantaneous (brittle) roof collapse triggered by a technogenic factor – the weight of heavy agricultural machinery. The study of the reactivation of karst‑suffosion processes in the backwater zone of the Bratsk Reservoir contributes to the understanding of modern geodynamic processes and ground surface deformation mechanisms, including those induced by technogenesis.

This paper deals with the process of developing and training a U-Net neural network for image segmentation of lakes and hillocks based on synthetic aperture radar and interferometric synthetic aperture radar data. The main goal of the work is to create an effective deep learning model capable of automatically identifying lakes and heave mounds based on complex radar images. To achieve this goal, several stages were carried out, including data collection and annotation, selection of the neural network architecture, training and validation of the model, as well as evaluation of its performance. At the beginning of the work, the process of creating a training dataset is described, which includes annotating images, highlighting features, and preparing data for training. Next, we consider the U-Net architecture, which was chosen because of its ability to efficiently segment objects in images. The choice of hyperparameters, such as the number of filters, the size of the convolution core and activation functions, is justified, and the Adam optimizer is used to achieve fast and stable convergence of the model. The learning and validation process of the model is described in detail with an emphasis on using the validation subset to monitor performance. Regularization methods, including early stopping, are used to prevent overfitting and improve the generalizing ability of the model. As a result, the importance of using deep learning for synthetic aperture radar and interferometric synthetic aperture radar data analysis is demonstrated, as well as confirmation of the effectiveness of the U-Net model for solving segmentation problems.