| Convolutional Autoencoder for Remote Sensing Change Detection |
| Paper ID : 1107-ICRSSSA-FULL |
| Authors |
|
MennaTullah Mamdouh El-Kholy *1, Dina Elsayad2, Marwa Mostafa3, Hala Moshir Ebeid4, Mohammed Fahmy Tolba4 1Scientific computing, Computer Science Faculty, Ain Shams University, Cairo, Egypt 2Scientific computing, Faculty of Computer and Information science, Ain Shams University 3Data Reception, Analysis, and Receiving Station Affairs, National Authority for Remote sensing and space science, Cairo, Egypt 4Faculty of Computer and Information Sciences, Ain Shams University, Cairo, Egypt |
| Abstract |
| Change Detection (CD) is one of essential tools for the accurate identifying and analyzing regions that have spatial or spectral changes. Binary change detection system aims to assign changes and no changing per pixel based on a pair of coregistered images of a given region taken at different times, since improving the quality of the binary CD map is an important issue in the areas of image processing, computer vision and remote sensing images. Due to the good performance of Deep Learning in the domain of nonlinear problem modeling, image processing and pattern recognition, Deep Learning is becoming popular to resolve the Change Detection problem using remote sensing imageries. Deep Learning and especially convolutional neural networks (CNN) monitor the environmental change in Binary change system. In this paper, we propose a new version of Siamese Convolutional Autoencoder which use heuristics for CD problem. We introduce three versions of Siamese architectures. We also studied the effect of order of the layers, and pooling layer to the accuracy of CD map. To evaluate the proposed architectures, we use a benchmark dataset called LEVIR-CD. Experimental results show that the proposed approach exceed the original accuracy of Siamese by around 3% in terms of accuracy. |
| Keywords |
| Remote sensing, Change Detection (CD), Convolutional Autoencoder, Deep Learning |
| Status: Accepted (Oral Presentation) |
