Using multi-source data and decision tree classification in mapping vegetation diversity

Abstract

This study acknowledges the problem of land cover demarcation in diverse vegetation condition. The Normalized Difference Vegetation Index is used for the preparation of base map. Further identification of mix and incorrect classes was done using ground truth. Radar data in combination with optical indices are used. In different NDVI classes, σ°RV with additional criteria on Normalized Difference Water Index successfully demarcated waterlogged area, polarization ratio σ°RV/σ°RH and backscattering coefficient σ°RH are found suitable to separate bare land from dry grass land, sparse and dense scrub could be separated by − (σ°RV + σ°RH)/2 and NDVI is efficient to identify dense vegetation. The study area is taken as Keoladeo National Park in Bharatpur, India. Statistical similarity between ground truth and classified class has been assessed using Jaccard coefficient (JC), Jaccard distance (JD), Dice coefficient (DC) and F-score. High similarity values of JC, JD, DC and F-score are achieved for all land cover types except bare land. Although, dry grassland showed low value of F-score; the reason could be low precision of class. The overall accuracy (87.17%), producer’s accuracy (86.39%), user’s accuracy (85.81%) and Kappa Coefficient (0.84) are also utilized to analyze performance of classifier. © 2018, Korean Spatial Information Society.