On the validation of pansharpening methods

TL;DR

This paper discusses the challenges in validating pansharpening methods, compares different validation approaches, and investigates various component substitution and high pass filtering methods on remote sensing data.

Contribution

It provides an analysis of validation issues in pansharpening and evaluates several methods with enhancements on real satellite data.

Findings

Validation approaches often yield inconsistent results.

Component substitution and HPF methods show varying effectiveness.

Enhancements like haze correction and spectral response functions improve performance.

Abstract

Validation of the quality of pansharpening methods is a difficult task because the reference is not directly available. In the meantime, two main approaches have been established: validation in reduced resolution and original resolution. In the former approach it is still not clear how the data are to be processed to a lower resolution. Other open issues are related to the question which resolution and measures should be used. In the latter approach the main problem is how the appropriate measure should be selected. In the most comparison studies the results of both approaches do not correspond, that means in each case other methods are selected as the best ones. Thus, the developers of the new pansharpening methods still stand in the front of dilemma: how to perform a correct or appropriate comparison/evaluation/validation. It should be noted, that the third approach is possible, that is to perform the comparison of methods in a particular application with the usage of their ground truth. But this is not always possible, because usually developers are not working with applications. Moreover, it can be an additional computational load for a researcher in a particular application. In this paper some of the questions/problems raised above are approached/discussed. The following component substitution (CS) and high pass filtering (HPF) pansharpening methods with additive and multiplicative models and their enhancements such as haze correction, histogram matching, usage of spectral response functions (SRF), modulation transfer function (MTF) based lowpass filtering are investigated on remote sensing data of WorldView-2 and WorldView-4 sensors.