Phenology-based crop classification from multi-frequency dual-pol SAR data utilizing Gaussian Processes

This paper proposes a novel multivariate Gaussian Process Regression (GPR) approach for multi-class crop classification. Backscatter information from E-SAR L- and C-band dual-polarimetric data acquired during the AGRISAR 2006 campaign were used to train and validate the proposed Gaussian Process Classifier (GPC) model. The model’s accuracy was assessed separately for mono-temporal and multi-temporal data for both frequencies. The mono-temporal analysis reflected the inter-crop misclassifications because of high crop phenology dynamics. The overall multi-temporal accuracy of the proposed model for both L- and C-band was 96% with Kappa Score (κ) = 0.95. Finally, the optimized classifier model was utilized to generate the classification maps of the major crop types.