ERROR FINGERPRINT AND ENTROPY MAPS AS A WAY TO VISUALIZE THE EVALUATION OF THE QUALITY OF PERFORMANCE OF DEEP LEARNING MODELS FOR AERIAL IMAGE PROCESSING

Abstract

The work is devoted to the research process and the development of our own approach to create a standardized error map of the quality of deep learning models and entropy in cases of analysis of aerial images obtained from drones. With the help of such maps, it is possible to understand in which areas of the images the models are most often mistaken in the test data and obtain the corresponding uncertainties. Building such maps helps to identify weaknesses in models and is especially useful for cases where the average standard metrics are almost the same. To conduct research and experiments, the most popular neural network architectures U-Net, DeepLabV3+ and Feature Pyramid Network, which are most often used in image segmentation problems in the field of computer vision, were used. Semantic segmentation plays a key role in the analysis of spatial data in the tasks of environmental monitoring, mapping and land management. The high accuracy of popular architectures does not eliminate the problem of analyzing errors and uncertainties. The paper presents an approach to building error heat maps, which show where models most incorrectly classify pixels on test data, and to estimate uncertainties, it is proposed to build special entropy maps. In addition to determining quantitative indicators of the quality of models (accuracy, Jaccard index, F-metrics), such visualizations allow for a qualitative analysis of weaknesses, for example, in cases where images have heterogeneous terrain or fuzzy contours. In the course of experiments on our own set of aerial photographs, it was determined that the effectiveness of the proposed imaging approaches can significantly increase the understanding of the results of the trained models. The results of such studies can have a significant impact on improving segmentation architectures, training data preparation, and error analysis in general. Future research may focus on expanding the aerial imagery dataset to evaluate model performance and reduce errors.

Keywords: deep learning, semantic segmentation, U-Net, DeepLabV3+, FPN, neural networks, error maps, entropy maps, visualization method, drones, aerial images.