Optimized Feature-Level Fusion of Hyperspectral Thermal and Visible Images in Urban Area Classification

Remote sensing utilization has become a new and well-accepted trend in the use of multisource images at different processing levels for numerous applications such as the classification of the urban area. Throughout this study, in order to fully exploit information for the classification of the urban land cover, feature-level data fusion of a coarse resolution hyperspectral long-wave infrared (LWIR) image and a very high-resolution visible-light image are employed. However, optimum parameter determination for the support vector machine (SVM) classifier and the feature subset pick strongly affect the classification performance of these data. Taking into consideration the complex relationship between these two obstacles, the parameters of SVM and the feature subset by particle swarm optimization (PSO) are the simultaneous determinants proposed in this paper. For this purpose, on the one hand, vegetation index, spectral and textural features, and morphological building index (MBI) obtained from visible data are extracted. On the other hand, PCs are derived from hyperspectral LWIR. Experimental implementations of the 2014 Data Fusion Contest dataset showed that the suggested approach improved the classification efficiency by up to 7% compared to SVM without PSO. Furthermore, the obtained findings illustrate the superiority of the suggested technique compared to other data fusion experiments with the same data.