Multiple kernels learning for classification of agricultural time series data

A new multiple kernel learning (MKL) framework is presented for classification of satellite remotely sensed time series for agricultural analysis. In this MKL framework, a new composite kernel is constructed with a weighted sum of some predefined kernels. The problem of proper estimation of weights is modeled as an optimization problem of maximizing the kernel alignment between composite kernel and an ideal kernel. Two heuristics and genetic algorithm are used to solve this optimization problem. The proposed MKL framework is tested with SVM algorithm using a time series of Soil Adjusted Vegetation index data for mapping of five crops. The results obtained from the MKL method showed higher performances in comparison with the SVM trained with a single kernel.