Analyzing the Efficiency of the Deep Neural Network in Detecting Urban Changes Using Bi-temporal Landsat-8 Images.

Subject Areas : Applications in earth’s climate change

sahand tahermanesh ¹ , Behnam Asghari Beirami ² , Mehdi mokhtarzade ³

1 - MSc. Student of Remote Sensing, Faculty of Geodesy and Geomatics Engineering, K. N. Toosi University of Technology, Iran
2 - Ph.D. Student of Remote Sensing
3 - Associate Professor of , K. N. Toosi University of Technology

Received: 2022-05-15 Accepted : 2022-08-03 Published : 2024-09-29

Keywords: deep learning, convolutional neural network, Change detection, Artificial Neural Network, Random forest,

Abstract :

Satellite remote sensing images are widely used to monitor the earth's surface phenomena changes at various periods. For accurate change detection, spatial features can be used as the complement information of spectral features. Hand-craft spatial features such as the co-occurrence matrix features are inefficient in detecting the changes due to the complex structure of satellite images. In the present study, a deep learning-based model is proposed as the alternative to address the problems of classical change detection methods. deep neural networks are mainly developed for images and hierarchically extracting spatial-spectral features. In this study, Landsat-8 images between 2013 and 2021 were used to evaluate the changes in Sahand city using the proposed deep network. Pre- and post-classified Landsat-8 images are produced using a deep neural network in the first stage. In the second stage, for producing the change maps, the post-classification approach is used in that change maps are produced based deference of classified images. Finally, the majority voting technique eliminates the noises in change maps. The proposed method results are compared with those obtained by two classical machine learning methods, random forest, and artificial neural networks. According to the change detection results, the proposed deep learning network improves detection accuracy by 13.88% and 12.80% compared with artificial neural networks and random forests. Compared to the random forest and artificial neural networks, the proposed network has improved the overall accuracy of the from-to-change maps by 57.81% and 65.7%, respectively. Final results demonstrate that although Random forest and artificial neural networks have been able to identify the location of changes, they perform poorly in detecting the from-to changes