Use of Artificial Neural Networks for Automatic Categorical Change Detection in Satellite Imagery

Change detection techniques based on post-classification comparison are used for monitoring the land-cover change using multi-temporal satellite imagery. A methodology for automatic land-cover change detection has been developed based on Artificial Neural Networks (ANNs) for multispectral image classification. The method developed is based on a classification step of the available images and a change detection step which makes use of the classified images. Three supervised learning ANNs are used: Multilayer Perceptron (MLP), Radial Basis Function (RBF) and Non-Extensible Radial Basis Function (NERBF). In addition, an unsupervised learning ANN technique called Self-Organized Map (SOM) is also used in this work. A texture analysis, using a gray-scale image, and a statistical analysis, using a RGB image, are applied to identify the inputs to the ANN. The texture analysis is done using Grey-Level Co-Occurrence Matrix (GLCM). The statistical analysis is done by calculating the average and standard deviation of a pixel with respect to its eight neighbor pixels in each of the bands composing the image, red (R band), green (G band), and blue (B band). Two images acquired from different periods of time (years 2003 and 2014) from Casalvieri (Frosinone province, Italy), have been employed to track the changes caused by the human action on the terrain. Our results demonstrate the potential of using ANNs for automatic land-cover change detection.