Fractal-Based Pattern Extraction From Time-Series NDVI Data for Feature Identification

Abstract

The spatial coarseness of time-series satellite data (≥250 m) and land cover mixing within a pixel poses difficulties in identifying a unique reference profile for each land cover class in regional level mapping. This problem is exaggerated further in a hilly terrain with different topographic, climatic, environmental conditions, and landscape composition. In this regard, this study aims to evaluate the usefulness of two reference-independent fractal dimension (FD) estimation techniques (box counting and rescaled range) for identifying land cover classes from the MODIS (MOD13Q1-v5) 16-day interval time-series normalized difference vegetation index profile at the pixel level. We have analyzed the FD images from each technique using Welch's t -statistics to evaluate the interclass mean FD variations in Himachal Pradesh—a mountainous state in Northern India. Furthermore, a support vector machine-based classification is performed using time-series principal components in conjunction with FD measures to assess the effectiveness of FD images in improving the accuracy of land use and land cover classification. The experiments revealed that the use of FD images has considerably increased the user accuracy and helped in accurately detecting mixed pixels which are dominated by agriculture class.