Improving change detection of urban areas with optimal selection of textural and spectral features using genetic algorithm

Analyzing multi-temporal remotely sensed images is a powerful technique for monitoring land use and land cover changes in urban areas. Apart from the change detection (CD) technique, the features space have an enormous impact on the CD accuracy.To achieve satisfactory CD results in urban area, optimum selection of textural and spectral features is necessary. Although an exhaustive search guarantees the optimality of the selected features, but it is computationally prohibitive. Data reduction techniques such as PCA considers the independence of the data to find a smaller set of variables with less redundancy without intending to improve the CD accuracy. Difficulty in setting the best threshold for JM distance in separability analysis algorithm (SAA) reduces its efficiency. The aim of this paper is finding the optimal textural and spectral features to enhance the CD accuracy using genetic algorithms (GA) and Bayesian classifier. To evaluate the effectiveness of the proposed approach, a case study using IRS-P6 and GeoEye1 satellite imagery acquired on July 15, 2006, and September 1, 2013, respectively, from Sahand New Town (Northwest of Iran) was performed. All the mentioned feature selection methods (PCA, SAA and proposed GA-based method) were implemented in MATLAB R2013a. Results show that, textural features provides a complementary source of data for CD in urban areas. Results show that features selection is an effective procedure in change detection based on textural and spectral features. Each of feature selection methods has its own limitation and advantages, but in general they increase the CD accuracy. The proposed GA-based feature selection approach was found to be relatively effective when compared to PCA and SSA approaches. Overall accuracy and Kappa coefficient of CD were increased from 53.66% to 88.49% and 58.94% to 90.39%, respectively using proposed method in compared with that using only spectral information.