E.A. Antokhin1, A.A. Zalishchuk2, V.A. Nenashev3, S.А. Nenashev4

1-4 St. Petersburg State University of Aerospace Instrument Engineering (St. Petersburg, Russia)

1 fresguap@mail.ru; 2 sacha1501@yandex.ru; 3 nenashev.va@yandex.ru; 4 nenashev_sergey178@mail.ru

Problem. The task of classifying images of the earth surface is a particularly significant one in the field of data processing and analysis for aviation systems with technical vision. Nowadays there is a necessity to classify separate zones of the Earth surface on a large amount of data in order to assess its condition. As the number of available images of the Earth's surface increases every year, the possibility to improve the accuracy of automatic classification of individual zones of the Earth's surface increases as a consequence. Extraction of useful information from large amounts of data for the classification task without application of neural network technologies is a long and subjective process. It follows that there is a problem of both acceleration of the process of processing large data sets and selection of the most suitable neural network providing a high share of correctly classified zones of the Earth's surface from optical images.

Purpose. To carry out a comparative analysis of different neural network models used for classification of Earth surface zones from optical images. Further on the basis of the comparative analysis it is required to evaluate the investigated models in order to identify the most suitable one for this type of big data sets in terms of speed and accuracy of the output result.

Results. A study of modern neural network models has been carried out on the basis of comparative analysis of their quality metrics. It was found that the greatest efficiency in the classification of earth surface zones is shown by the neural network “InceptionV3”, which uses a pre-trained architecture and has a wide feature extraction capability, which allows it to capture important patterns and structures on a large set of images. The neural network models “ResNet” and “MobileNetV2” also showed quality metrics for their performance. However, with slightly lower accuracy than “InceptionV3”, making them a less acceptable choice for this task. Thus, the neural network model “InceptionV3” shows the fastest performance and the highest percentage of correctly classified land surface areas, which is more than 99% of those considered.

Practical relevance. The results of the study are essential for analyzing land surface images in various fields such as land surface monitoring, vegetation mapping, urban planning, and rapid response to natural disasters. Accurate classification of land surface images enables operational decisions to be made automatically. The use of pre-trained neural network models provides the required indicators of quality metrics when solving the problem of classification of earth surface zones from optical images.

Antokhin E.A., Zalishchuk A.A., Nenashev V.A., Nenashev S.A. Comparative analysis of neural network models for classification of earth surface zones from optical images. Radiotekhnika. 2024. V. 88. № 8. P. 35−44. DOI: https://doi.org/10.18127/j00338486-202408-04 (In Russian)

1. Nenashev V.A. Osobennosti klassifikacii podstilajushhih poverhnostej zemli po harakteristikam jeho-signalov v bortovyh RLS. Trudy MAI. 2021. № 118. DOI: 10.34759/trd-2021-118-11 (In Russian).
2. Bakunov D.V., Pljushkin S.A., Solov'ev D.A. Model' mashinnogo obuchenija dlja klassifikacii zemnoj poverhnosti na osnove sputnikovyh izobrazhenij. Izvestija RAN, 2017. T. 72. № 3. S. 425-438 (In Russian).
3. Vinogradov S.A., Kuznecov D.V., Perelygin V.P. Ocenka kachestva modelej mashinnogo obuchenija dlja zadach analiza sputnikovyh izobrazhenij. Izvestija RAN. 2017. T. 72. № 3. S. 439-452 (In Russian).
4. EuroSAT: Land Use and Land Cover Classification with Sentinel-2. GitHub URL: https://github.com/phelber/EuroSAT (data obrashhenija: 21.11.2023).
5. El-Sayed A., et al. Assessing the potential of machine learning algorithms for land cover classification using remote sensing data. Remote Sensing. Sep. 2019. V. 11. № 9. Р. 1153–1176.
6. Alzahabi S., Alzahabi A., AlZahabi A. A Comprehensive review of machine learning techniques for land cover classification. 2020.
7. Dandin D.D., Shrivastava S.K., Shukla S.K. A comparative study of land cover classification techniques in remote sensing: A review. Journal of Remote Sensing. 2018. V. 42. № 3. Р. 1450–1476.
8. Shrivastava S.K., Shukla S.K., Mishra S.K. Machine learning for land cover classification using remote sensing data. A Review. 2017.
9. Rich feature hierarchies for accurate object detection and semantic segmentation. arxiv URL: https://arxiv.org/abs/1311.2524 (data obrashhenija: 11.11.2023).
10. Howard A.G., Zhu M., Chen B., Kalenichenko D., Wang W., Weyand T., Andreetto M., Adam H. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861. 2017.
11. He K., Zhang X., Ren S., Sun J. Deep residual learning for image recognition. CVPR. 2016. Р. 770-778.
12. Liu S., Deng W. Very deep convolutional neural network-based image classification using small training sample size. 2015. 3rd IAPR Asian Conference on Pattern Recognition (ACPR) (Kuala Lumpur. Malaysia. 03–06 November 2015). 2015. Р. 730-734.
13. Szegedy Ch., Vanhoucke V., Ioffe S., Shlens J., Wojna Z. Rethinking the inception architecture for computer vision. 2016. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (Las Vegas. NV. USA. 27–30 June 2016). 2016. Р. 2818-2826.
14. Gong J., Liu X., Zhang X. A review of land cover classification methods based on remote sensing images using machine learning algorithms. Journal of Remote Sensing. 2018. V. 42. № 3. Р. 1350-1373.
15. El-Sayed A., El-Bakry M. A survey of methods for land cover classification using remote sensing data. 2018.
16. Zhang H., et al. Accurate land cover classification using deep learning and its application in remote sensing. Remote Sensing. 2019. V. 11. № 8. Р. 1075-1092.
17. Nenashev V.A., Hanykov I.G. Formirovanie kompleksnogo izobrazhenija zemnoj poverhnosti na osnove klasterizacii pikselej lokacionnyh snimkov v mnogopozicionnoj bortovoj sisteme. Informatika i avtomatizacija. 2021. T. 20. № 2. S. 302-340. DOI: 10.15622/ia.2021.20.2.3.
18. Nenashev V.A., Khanykov I.G. Formation of fused images of the land surface from radar and optical images in spatially distributed on-board operational monitoring systems. Journal of Imaging. 2021. V. 7. № 12. DOI: 10.3390/jimaging7120251.
19. Nenashev V.A., Shepeta A.P., Kryachko A.F. Fusion radar and optical information in multiposition on-board location systems. 2020 Wave Electronics and its Application in Information and Telecommunication Systems (WECONF 2020. Saint-Petersburg. 01–05 June 2020). Saint-Petersburg, 2020. P. 9131451. DOI: 10.1109/WECONF48837.2020.9131451 (In Russian).
20. Sencov A.A., Nenashev V.A., Ivanov S.A., Turneckaja E.L. Sovmeshhenie sformirovannyh radiolokacionnyh izobrazhenij s cifrovoj kartoj mestnosti v bortovyh sistemah operativnogo monitoringa zemnoj poverhnosti. Trudy MAI. 2021. № 117. DOI: 10.34759/trd-2021-117-08 (In Russian).