Abstract:
Determining the spatial resolution (GSD) of remote sensing images is in demand in various fields, from environmental monitoring to urban planning and agriculture, which makes it relevant to analyze both urbanized and natural landscapes. This study focuses on deep learning methods for evaluating the GSD of still images, and its purpose is to study the impact on the quality of the GSD assessment of approaches such as pre-training the autoencoder on still images (SSL) and taking into account features from different levels of the model. All the models under consideration were trained on images of different types of terrain. Scaling augmentations were also used to train them to expand the range of GSD in data samples. ResNet18 was used as the base model, which, based on the available data, demonstrates a relative error in the range from 2.73% to 15.6%. Using the relative loss function directly in training allows you to improve performance on data with low spatial resolution from 15.6% to 14.7%. At the same time, using the SSL approach slightly improves performance at high GSD values. The combination of SSL with FPN reaches 3.61% on a set with a mixed terrain type. The results obtained make it possible to use trained models to solve applied problems. However, it is necessary to choose the most appropriate method for each specific case.
