Training concepts in Noise Clustering Classifier -A case study for Pigeon Pea crop mapping

Abstract

In this research work, a novel approach of ‘Individual Sample as Mean’ (ISM) based Noise Clustering has been proposed, where each sample is taken as an input parameter in training the model. When compared to the existing method which uses statistical mean for training, this proposed technique is far superior and robust, the efficiency of which has been demonstrated using Pigeon Pea crop mapping. The issue of spectral overlap of Pigeon Pea with other crops in vicinity is overcome by adopting a temporal approach. In order to fill the void in multi-temporal data availability, dual sensor approach was implemented. Different vegetation indices have been tested in reducing data dimensionality and the temporal datasets were optimized so as to capture the unique growth phenology of the Pigeon Pea crop. A new technique of ‘Class Based Sensor Independent – Modified Soil-Adjusted Vegetation Index 2 (CBSI-MSAVI2)’ has been proposed in this research work for the purpose of dimensionality reduction which ensures the maximum enhancement of target crop class in temporal domain. Based on an evaluation of classification results using Mean Membership Difference (MMD) method, the proposed ISM-Based NC appears to have produced superior results for Pigeon Pea since the test fields had membership values that were closer to those of the training fields. The MMD for ISM-Based NC was around 0.0013 whereas for the usual Mean-Based approach the value was at around 0.0071 respectively for Pigeon Pea crop fields. It was also noted that variance was zero in outputs generated by the ISM-based NC approach thereby proving the efficiency of the same in suppressing heterogeneity within the field.