| Abstract [eng] |
The application of deep neural networks for satellite analysis is researched in this work. Due to various natural processes and human behavior, deforestation is happening, and water boundaries are changing. These changes must be monitored in case harmful processes can be stopped. Also, it would help to plan reforestation and water bodies „recovery“. This problem can be solved by using deep neural network which would help to segment satellite images automatically and monitor changes. There are also presented various types of deep neural network architectures in this work. U-net and SegCaps architecture networks are selected to be analyzed in more detail. These networks seem to be the best in medical images segmentation and U-Net architecture deep neural networks was also used in satellite images segmentation in the past. The main problem of analysis of satellite images is that objects in satellite images are small and photographed from above that is the reason why it is hard to obtain accuracy while doing segmentation separate objects. U-Net and SegCaps deep neural networks have shown different results. U-Net has been distinguished as a more precise artificial neural network. U-Net architecture network was better in image segmentation than SegCaps architecture, especially water bodies. It could happen for different reasons, first of all SegCaps was designed to segment 3D medical images which had many layers, and satellite images are 2-dimension images. |
