Volume IV-5
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., IV-5, 383-390, 2018
https://doi.org/10.5194/isprs-annals-IV-5-383-2018
© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 4.0 License.
ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., IV-5, 383-390, 2018
https://doi.org/10.5194/isprs-annals-IV-5-383-2018
© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 4.0 License.

  15 Nov 2018

15 Nov 2018

A GRAPHIC PROCESSING UNIT FRAME WORK FOR CONVOLUTIONAL NEURAL NETWORK BASED CLASSIFICATION OF REMOTELY SENSED SATELLITE IMAGES

R. A. Ansari1, W. Thomas2, and K. M. Buddhiraju1 R. A. Ansari et al.
  • 1Centre of Studies in Resources Engineering, Indian Institute of Technology Bombay, India
  • 2Department of Electrical Engineering, Indian Institute of Technology Bombay, India

Keywords: GPU, GPGPU, convolutional neural networks, parallel processing, image classification

Abstract. Near real time processing and feature extraction from high-resolution satellite images aids in various applications of remote sensing including segmentation, classification and change detection. The latest generation of satellite sensors are able to capture the data at a very high spatial, spectral and temporal resolution. The processing time required for such a huge data is also large. Disaster monitoring applications such as forest fire monitoring, earthquakes require fast/real time processing of high resolution data to enable response activities. In general, due to the large size of satellite data, the computational time of feature calculation and training neural network is found to be very high. Therefore in order to achieve the aim of near real time processing of such huge data, we developed a parallel implementation. The implementation is performed on NVIDIA’s Graphical Processing Unit. The performance improvement obtained is demonstrated by a GPU implementation on Resourcesat-1 data and compared with the traditional sequential implementation. The results show that the GPU implementation is found to achieve performance improvement in terms of execution time and speedup throughput as compared to the sequential implementation.