Satellite Image Mapping Refinement Method Based on Siamese Neural Networks

Geographically referencing satellite images plays a key role in geospatial data processing, as the accuracy of referencing directly affects the quality of subsequent data analysis and interpretation. Traditional geo-referencing methods, such as manual annotation using datum points or the use of descriptor algorithms (SIFT, SURF, ORB), have certain disadvantages. These methods are time-consuming and labor-intensive, especially when dealing with large amounts of data, and often depend on the human factor, which can lead to subjective errors. Often, existing geo-referencing methods do not take into account complex non-linear relationships that arise in conditions of changing illumination, shooting angles and seasonal changes in the landscape. Ignoring these non-linear factors can lead to significant distortion in the binding process, reducing the overall accuracy of data processing. It is necessary to develop a method of geo-referencing satellite images that would combine high accuracy, speed and resistance to changing shooting conditions, taking into account possible non-linear dependencies, to increase the efficiency of processing geospatial data in real conditions. Purpose – develop a method to refine the geographic mapping of satellite images based on Siamese neural networks. It has been established that the use of Siamese neural networks for the task of geo-referencing satellite images demonstrates accuracy comparable to descriptor methods, while significantly exceeding them in processing speed. It was found that the average time for processing one image was 0.112 s. It was noted that the best binding accuracy was achieved at a resolution of 10 m per pixel, where the average absolute error (MAE) was 18.3 m in the autumn. At higher resolutions, the accuracy decreased, but the method still showed stable results even at a resolution of 50 m per pixel. It has been shown that the proposed method using Siamese neural networks is able to take into account and adapt to non-linear processes, such as the variability of illumination and relief, which makes it more resistant to changes in shooting conditions compared to traditional algorithms. The proposed method makes it possible to significantly speed up the process of geo-referencing, which is especially important when processing large amounts of data in automatic mode. This opens up the possibility of using the method in monitoring and mapping tasks, where high accuracy is required with minimal time spent. The application of Siamese nets can reduce processing time and improve accuracy in complex landscapes such as forests and meadow areas, with practical implications for ecology, agriculture and area management.