Analysis of land cover mapping performance of deep learning based PAN-sharpening methods

Pansharpening enhances both spatial and spectral characteristics of a high-resolution panchromatic (PAN) image by fusing its spatial features with the spectral properties of a lower-resolution multispectral (MS) image, resulting in an improved MS image. Over the past forty years, numerous pansharpening methods have been developed, and in recent years, deep learning (DL) based approaches have emerged for pansharpening, similar to other applications in remote sensing. Therefore, it is important to examine the performance of DL-based methods and assess their applicability for various remote sensing applications. The proliferation of various pansharpening methods makes it challenging for researchers to select the best-performing technique. Hence, the objective of this study is to evaluate the performance of several DL-based pansharpening methods utilizing pre-trained models, along with traditional methods, in terms of enhancing spatial sharpness and preserving spectral color accuracy in images. A total of 29 pansharpening methods were employed to generate pansharpened images, and their spectral and spatial characteristics were qualitatively and quantitatively assessed to investigate classification performance. The study utilized datasets from IKONOS and WORLDVIEW2 (WV2) satellites. The pansharpened images were classified using the Support Vector Machine (SVM) algorithm to investigate the effects of pan-sharpening on the land cover mapping performance. The variational optimization (VO)-based techniques generally outperformed other methods, with a few exceptions. The DL-based methods were found to result in superior pansharpening and classification accuracy compared to many other traditional approaches. Additionally, it was revealed that the performance of DL-based methods is highly dependent on the success of training and the hyperparameters employed.