Thermodynamic Entropy: A Natural Metric for Landcover Classification

Abstract

Most landcover (LC) studies employ physical characteristics involving spectral reflectance. This study employs the concept of energy and entropy for LC classification. Satellite imageries from Landsat-8, the Operational Land Imager (OLI) and the Thermal Infra-Red Sensor (TIRS) captured at 100 m for three years (2013, 2015 and 2017) and resampled to a spatial resolution of 30 m for multispectral measurements were acquired. The normalized difference vegetation index (NDVI) of the study area was computed for both the wet (September/October) and dry (March) seasons of years 2013, 2015 and 2017, respectively; and the vegetation cover maps and landcover classification of the area based on the NDVI values were generated. Land surface temperature (LST) and net radiation of the surface area were computed and used as inputs for computing the surface entropy flux (SEF) of the study area. Overall, the dry season of 2017 had the highest vegetation cover while the wet season vegetation cover was highest in 2015. The high dry season vegetation cover in 2017 is attributable to higher level of living biomass than previous years. NDVI-based LC classification of the area showed an overlap in distinguishing between built-up areas and vegetation cover as well as between bare (land) surface and free water bodies. This is attributable to skewed measure of centrality for the sample distribution possibly resulting from the quality (resolution) of data used or sampling techniques. Free water bodies had the highest SEF values that fall within expected behavior for bodies of such state (liquids). The vegetation covers had the second highest SEF values for the entire site and the highest on land surface, which could be attributed to higher latent heat fluxes of such covers resulting from the evapotranspiration processes. The bare surfaces and built-up area of the study area were observed to have the lowest SEF values. The LC of the study area was reclassified based on SEF values via ground truthing i.e., by taking statistical distribution of well-known surface classes across the years under study. This was successful in distinguishing vegetation cover, surfaces of free water bodies and bare surfaces. A regression analysis revealed that LST influences SEF by up to 69 – 75% while vegetation cover has no appreciable influence on SEF. A two-sample t-test showed a significant (P = 0.05) difference between the SEF for bare surfaces and that for vegetation cover. All these presents SEF as a reliable natural metric for LC classification.