Multiset Kernel CCA for multitemporal image classification

The analysis of multitemporal remote sensing images is becoming an increasingly important problem because of the upcoming scenario of multispectral satellite constellations monitoring our Planet. Algorithms that can analyze such amount of heterogeneous information are necessary. While linear techniques have been extensively deployed, this work considers a kernel method that finds nonlinear correlations between all image sources and the class labels. We introduce in this context the Kernel Canonical Correlation Analysis (KCCA) to exploit the wealth of temporal image information and to handle nonlinear relations in a natural way via kernels. To achieve this goal, we use the generalization of KCCA for several datasets, Multiset KCCA (MKCCA), which allows to analyze more than two datasets simultaneously. MKCCA can map several datasets to a high dimensional space where features have maximum correlation. Using several images from a time series together with labeled information, these nonlinear features obtained with MKCCA can be used with simple linear classifiers, and yield good results. We also propose an easy approach to use multitemporal information via composite kernels together with KCCA and linear classifiers. We illustrate the use of MKCCA and composite kernels plus KCCA for land-use classification in two multitemporal settings. The obtained results show the validity of the proposed methods, and open new possibilities for future algorithms and applications.