I.S. Demin – Master Student, 

Department «Computer software, Information Technologies», Kaluga branch of the Bauman MSTU Е-mail: dantlou@mail.com

Yu.S. Belov – Ph.D.(Phys.-Math.), Associate Professor, 

Department «Computer software, Information Technologies», Kaluga branch of the Bauman MSTU Е-mail: ybs82@mail.com

I.V. Chukhraev – Ph.D.(Eng.), Associate Professor, Head of Department «Information Systems and Networks»,  Kaluga branch of the Bauman MSTU E-mail: chukhraev@bmstu-kaluga.ru

The learning of satellite images allows you to recognize objects and determine the various properties of the identified objects. Satellite imagery is used in various areas, for example: urban planning, agricultural and forest management, disaster relief and climate modeling. This article discusses the complete process of recognizing and classifying objects on satellite images of the earth. The project is devoted to the analysis of the use of convolutional neural networks to detect types of the earth's surface according to remote sensing data. For training of convolutional neural networks was used the Urban Atlas tagged image database. Urban Atlas data set contains images of 21 classes. The project used the U-Net architecture of a convolutional neural network, which consists of two parts: a folding conveyor and an expanding network. The network consists of 23 convolutional layers. The pipeline convolution operation is the sequential execution of a convolution layer (3×3) followed by a 2×2 linear straightening module (ReLU) to fix the negative part of the scalar value, followed by the operation of merging the layers with a 2×2 window in step 2 for down-sampling. At each sampling step, the number of channels of the output function is doubled. The expanding network at each stage performs a double roll-up, followed by a 2×2 convolutional layer. This layer reduces the number of functional channels. A 2×3 convolution operation and a 2×2 linear straightening block come after. Cropping is necessary at each stage due to the loss of pixels on the border after each convolution stage. A 1×1 convolution layer is performed to match each 64-component output vector with the classification classes in the last step.

To analyze the accuracy of the object detection algorithm, it is necessary to compare the contours of automatically detected areas with areas of expert marking. To analyze the accuracy of the object detection algorithm, the selected regions were compared with the areas previously noted by experts. For the experiment, 100 satellite images were selected. Then the procedure of automatic detection of objects in the image was carried out, the obtained polygons were compared with expert polygons. As a result, the percentage of intersection was calculated. The largest percentage of intersections of areas detected by objects of the class «Buildings». This is due to the clarity of the borders and the visible visual separation of the surrounding objects. In recognition objects of the «Water» class, many separate parts of the water resource are allocated, which leads to a large number of false positives. This is due to the presence of ice and other objects above the rivers and lakes. Thus, it turns out several waters instead of one water source. Forest territory has an average detection value of 92% of the intersection of areas. This is due to the inaccurate distribution of the borders of the forest area. The accuracy of the final classification is 81% for the «Water» class objects, 92% for the «Forest» class and 96% for the «Building» class objects. The considered algorithm can be applied for: allocation of territories of cities, control over construction, etc.

1. Goodfellow I., Bengio Y., Courville A., Bach F. Deep learning (adaptive computation and machine learning series). Cambridge Massachusetts. 2017. P. 321−359.
2. Géron A. Hands-on machine learning with scikit-learn and tensorflow: concepts, tools, and techniques to build intelligent systems. O'reilly Media. 2017. P. 8−21.
3. Komartsova L.G., Maksimov A.V. Neirokompyutery. Izd-vo MGTU im. N.E. Baumana. 2004. 400 s. (in Russian)
4. Dey S. Hands-on image processing with python: expert techniques for advanced image analysis and effective interpretation of image data. Packt Publishing Ltd. 2018. P. 12−34.
5. Bobyleva E.A., Drach V.E., Rodionov A.V., Chukhraev I.V. Reshenie zadachi klassifikatsii simvolov s pomoshchyu kaskadnykh klassifikatorov. Elektromagnitnye volny i elektronnye sistemy. 2017. T. 22. № 3. S. 61−66. (in Russian)
6. Frolov P.V., Chukhraev I.V., Grishanov K.M. Primenenie iskusstvennykh neironnykh setei v sistemakh obnaruzheniya vtorzhenii. Sistemnyi administrator. 2018. № 9 (190). S. 80−83. (in Russian)
7. Grishanov K.M., Belov Yu.S. Metody vydeleniya priznakov dlya raspoznavaniya simvolov. Elektronnyi zhurnal: nauka, tekhnika i obrazovanie. 2016. № 1 (5). S. 110−119. (in Russian)
8. Rybkin S.V., Nesterov A.Yu. Podkhody v raspoznavanii dorozhnykh znakov. Sb. statei Mezhdunar. nauchno-prakticheskoi konf. «Problemy formirovaniya edinogo nauchnogo prostranstva». 2017. V 4-kh chastyakh. S. 16−19. (in Russian)
9. Vershinin V.E., Grishunov S.S., Loginova M.B. Modelirovanie protsessov raspoznavaniya i klassifikatsii mnogomernykh ob’ektov peresekayushchikhsya klassov na osnove predstavlenii teorii nechetkikh mnozhestv. Nauka, tekhnika i obrazovanie (elektronnyi zhurnal). 2016. № 1(5). S. 120−133. (in Russian)
10. Chityala R., Pudipeddi S. Image processing and acquisition using python. CRC Press. 2014. P. 37−55. Nesterov A.Yu., Belov Yu.S. Cravnitelnyi analiz funktsionirovaniya algoritma raspoznavaniya po kontrolnym tochkam i rezultatov raboty mobilnogo prilozheniya ROADAR. Nauka, tekhnika i obrazovanie (elektronnyi zhurnal). 2017. № CB1. S. 139−145. (in Russian)