Automatic Procedure for Monitoring Urban Land Cover Changes by Multitemporal ASAR Images

In this paper an automatic procedure for change detection from multitemporal advanced synthetic aperture radar (ASAR) images is presented. The model is structured in four main blocks: a set of standard preprocessing procedures, a feature-extraction module, a neural-network classifier and a change detection module. To improve the capabilities of the neural network, the feature-extraction module derives a set of features from a series of multitemporal ASAR images, such as mean and standard deviation of four backscattering intensity images (one per each season of the year) and their textural features (Energy and Contrast) based on the Gray Level Co-Occurrence Matrix (GLCM) method [2,3]. The supervised neural network algorithms [4,5] is used to classify urban areas versus not urban areas. It is based on the multilayer perceptron MLP approach and trained with the error back-propagation (EBP) learning algorithm. Finally, the change detection map is obtained by comparing two classified images taken at two different times. To validate the model, test images acquired by the ENVISAT ASAR over Rome, Italy and its surroundings are used. Changes occurred from 2005 to 2009 are automatically recognized with an accuracy of about 75%, while the classification accuracy obtained over the urban areas is higher than 90%. The developed algorithm is capable of efficiently identifying urban growth on big scale areas with short processing time.