Kernel-based regularized learning for time-invariant detection of paddy growth stages from MODIS data

Abstract

Most current studies have been applying high temporal resolution satellite data for determining paddy crop phenology, that derive into a certain vegetation indices, by using some filtering and smoothing techniques combined with threshold methods. In this paper, we introduce a time invariant detection of paddy growth stages using single temporal resolution satellite data instead of high temporal resolution with complex cropping pattern. Our system is a kernelbased regularized learner that predicts paddy growth stages from six-bands spectral of Moderate Resolution Image Spectroradiometer (MODIS) satellite data. It evaluates three Kernel-based Regularized (KR) classification methods, i.e. Principal Component Regression (KR-PCR), Extreme Learning Machine (KR-ELM), and Support Vector Machine with radial basis function (RBFSVM). All data samples are divided into training (25%) and testing (75%) sampling, and all models are trained and tested through 10-rounds random bootstrap re-sampling method to obtain more variety on hypothesis models during learning. The best model for each classifier method is defined as the one which has the highest kappa coefficient during testing. The experimental results show that the classification accuracy of each classifiers on testing are high competitive, i.e. 84.08%, 84.04%, and 84.95% respectively.