A neural network approach for land-cover change detection in multi-temporal multispectral remote-sensing imagery

  • Authors:

  • Victor-Emil Neagoe;Mihai Neghina;Mihai Datcu

  • Affiliations:

  • Depart. Electronics, Telecommunications & Information Technology, Polytechnic University of Bucharest, Bucharest, Romania;Depart. Electronics, Telecommunications & Information Technology, Polytechnic University of Bucharest, Bucharest, Romania;Depart. Electronics, Telecommunications & Information Technology, Remote Sensing Technology Institute, German Aerospace Center, Germany

  • Venue:
  • GAVTASC'11 Proceedings of the 11th WSEAS international conference on Signal processing, computational geometry and artificial vision, and Proceedings of the 11th WSEAS international conference on Systems theory and scientific computation
  • Year:
  • 2011

Quantified Score

Hi-index 0.00

Visualization

Abstract

This paper presents a neural network approach for land-cover change detection in remote-sensing imagery. One has considered the following supervised neural classifiers: Multilayer Perceptron (MLP), Radial Basis Function Neural Network (RBF), and Supervised Self Organizing Map (SOM). For comparison, we have chosen two well-known statistical classifiers (Bayes and Nearest Neighbour (NN)). The proposed model of change detection in multispectral satellite images has two main processing stages: (a) feature selection (using one of the three techniques: concatenation algorithm (CON), the algorithm based on absolute differences of pixels (ADIP), and the algorithm based on difference of reflectance ratios (DIRR)); (b) classification, using one of the above mentioned classifiers. The considered techniques are evaluated using a LANDSAT 7 ETM+ multitemporal image, corresponding to a set of two images of the same zone (400 × 400 pixels) in the region Markaryd, Sweden taken in 2002 and 2006. One has the change reference map; we have used 2000 pixels for training and the rest of 158 000 pixels for test. The best experimental result leads to the change detection rate of 88.24% for the test lot, proving the advantage of neural network models over the statistical ones.