Method of Improving Instance Segmentation for Very High Resolution Remote Sensing Imagery Using Deep Learning
| Autor: | Volodymyr V. Hnatushenko, Vadym Zhernovyi |
|---|---|
| Rok vydání: | 2020 |
| Předmět: | |
| Zdroj: | Communications in Computer and Information Science ISBN: 9783030616557 DSMP |
| Popis: | In current work a complex solution for very high resolution remote sensing data processing is suggested. The approach is based on modern deep learning techniques such as fully convolutional neural networks and solves multiple significant remote sensing problems such as detection and instance segmenetation of manmade land objects on very high resolution hyperspectral map scenes and taking an advantage of non-visible bandwidths. The solution consists of three parts: dataset development, neural network development, neural network fine-tuning. The first part of the solution suggests a semi-automated method of rapid dataset development using modern tools and keeping the advantages of remote sensing imagery such as very high resolution and using the variety of hyperspectral imagery bandwidths, both - visual and thermal. In order to prove application of a dataset, current article suggests an approach to the problem of remote sensing imagery segmentation using deep fully convolutional neural networks. The best deep learning architectures for instance segmentation are considered and investigated, and the binary classifier for a dataset is built based on Mask RCNN. The final neural network architecture with the new classifier is fine-tuned using alternative configurations and optimizers, and altered to benefit more from the developed hyperspectral remote sensing dataset. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|