A View Through Time: a Data Science Approach on the Spectral Alignment of Landsat 5, 7 and 8

Abstract

The Landsat satellite series provide geoscientist with over half a century with sensing data. Differences in sensor properties on these satellites create variations between sensing. A spectral alignment of the Landsat satellite series would be required to properly conduct longitudinal research on the development of the earth’s surface. In this thesis four different models, two ordinary least square regressions, a polynomial regression and a single layer neural network regression, were compared in their performance on the transformation of Landsat 8 and Landsat 5 on Landsat 7 in a spatially and non-spatially dependent situation. Furthermore was tested if the azimuth and elevation formed a meaningful contribution to the transformation model. Here we show that in the transformation from Landsat 8 to Landsat 7 and Landsat 5 to Landsat 7 the single layer neural network regression, with the addition of the azimuth and elevation, has the lowest root mean squared error on spatially independent data. On spatially dependent data did the ordinary least squares regression without the azimuth and elevation as predictors contain the lowest root mean squared error. However, on local spatially dependent data did the single layer neural network contain the lowest root mean squared error as well. The difference in performance between a single neural network regression and complex linear regressions could shows that even in its simplest form, the neural network architecture forms a better fit for these ransformations. Future research could study the implementation of backwards compatibility models, assessing the structural validity of the model and allowing more complex neural network regressions.