Are Vegetation Indices Really so Different: An Unsupervised Perspective

Remote sensing has been employed to gather rapidly dynamic, nondestructive information on crops and has been utilized across numerous fields. The information richness has been condensed into a plurality of vegetation indices. They offer quantitative data about crop growth and health, supporting a range of precision agriculture practices. The introduction of many indices has been motivated by particular problems unsolved by the previous list. In this paper we investigate if these indices are indeed so different and to quantify the differences among them we assume an unsupervised approach. We focus our study on a PRISMA image from Brasov area. Out of this image, we select four semantic areas indicating agricultural crop, forest, alpine pasture and urban setting and into these four cases we evaluate differences between the most used vegetation index, NDVI, and other several popular choices. The differences are quantified from a statistic viewpoint, using Pearson and Spearman correlation coefficients and from an image consistency, using SSIM. We show that conditioned on semantic category, significant differences arise.