Class Based Sensor Independent Indices and Training Parameter Approach in Fuzzy Machine Learning Model for Psyllium Husk, Medicinal Crop Mapping

Anam Sabir, Anil Kumar*

Abstract

The use of spectral indices is prevalent in remote sensing data processing for a variety of applications. However, the spectral bands considered to formulate conventional spectral indices may not be the most appropriate ones for a particular application. With the improvement in technology, the available spectral resolution has increased up to a large extent which can be explored to enhance a class much better way for specific class mapping. This study tests the performance of Modified Soil Adjusted Vegetation Index (MSAVI2) for mapping Psyllium Husk (Plantago Ovata), a medicinal crop with a relatively small size which increases the impact of soil brightness. Three variants of MSAVI2 with different bands combination were tested i.e. Conventional (Red-NIR), RedEgde1 (705 nm)-NIR, and CBSI-MSAVI2 (Class Based Sensor Independent). The classification technique used was MPCM (Modified Possibilistic c-means), for which two training approaches were made use of i.e. Fuzzy mean-based MPCM and Fuzzy Individual-sample-as-mean (ISM) based MPCM. The classifications were carried out considering a range of sizes of training samples starting from 5 to 50. The accuracy of different combinations of index, bands, and number of samples were assessed using Mean Membership Difference, Variance, Fuzzy Error Matrix, and Sub-pixel Confusion Uncertainty Matrix by making use of soft classified CubeSat (Dove) 3m data. The overall accuracy for different test cases achieved were between 83%-99% while the Kappa coefficient varies from 0.6 to 0.99.

Keywords: Modified Soil Adjusted Vegetation Index; Modified Possibilistic c-Means; Individual-sample-as-mean; Class Based Sensor Independent; Fuzzy Error Matrix

Full Text: