SEA-LAND SEGMENTATION MODELS IN DEEP LEARNING FROM REMOTE SENSING DATA

Background. Coastline changes can have a significant impact on coastal landscape, ecosystems and communities. Therefore, monitoring of such a highly dynamic system as sea-land is an urgent task that can be solved both by traditional methods and by using depth learning techniques to improve the efficiency of processing such as class of tasks. The object of the authors' research is the coastline along the coast of the western part of the Crimean Peninsula, the study of which by traditional methods has become impossible due to the temporary occupation of the Crimean Peninsula since 2014. The paper considers the main coastal indicators and methods of coastline digitization. The main types of satellite images as well as their combinations are compared for effective utilization of the shoreline mapping task. Many methods are used to recognize and extract shorelines in satellite images, which are generally divided into three groups: indexing, edge detection and classification methods. Methods. Authors compared the main depth learning models that can be used to efficiently recognize the coastline and its boundaries in satellite images, which include ISODATA (Iterative Self-Organizing Data Analysis Technique), Maximum Likelihood Estimation (MLE), Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), U-Net, and Segment Anything Model (SAM). Results. The outlines of the Crimean Peninsula coastline were obtained on the basis of PlanetScope images using ISODATA, MLE, RF, KNN, SVM, U-Net, SAM methods. The obtained images and their performance were compared. The study included the development of a Python code to automatically generate reports including information on five evaluation metrics, such as accuracy (98.96), recall (99.45), precision (97.27), F1-score (98.34), and IoU (96.74), which facilitated the evaluation of different approaches and methods. Conclusions. The comparative analysis highlights the advantage of the U-Net model for shoreline extraction from remotely sensed images. U-Net consistently provides the most accurate and detailed segmentation in different scenarios, demonstrating robustness and accuracy.