Improving Super-Resolution Mapping by Combining Multiple Realizations Obtained Using the Indicator-Geostatistics Based Method

Zhongkui Shi,Xianfeng Zhang,Huiran Jin,Yan Chen,Yugang Tian,Peijun Li

doi:10.3390/rs9080773

Abstract

Indicator-geostatistics based super-resolution mapping (IGSRM) is a popular super-resolution mapping (SRM) method. Unlike most existing SRM methods that produce only one SRM result each, IGSRM generates multiple equally plausible super-resolution realizations (i.e., SRM results). However, multiple super-resolution realizations are not desirable in many applications, where only one SRM result is usually required. These super-resolution realizations may have different strengths and weaknesses. This paper proposes a novel two-step combination method of generating a single SRM result from multiple super-resolution realizations obtained by IGSRM. In the first step of the method, a constrained majority rule is proposed to combine multiple super-resolution realizations generated by IGSRM into a single SRM result under the class proportion constraint. In the second step, partial pixel swapping is proposed to further improve the SRM result obtained in the previous step. The proposed combination method was evaluated for two study areas. The proposed method was quantitatively compared with IGSRM and Multiple SRM (M-SRM), an existing multiple SRM result combination method, in terms of thematic accuracy and geometric accuracy. Experimental results show that the proposed method produces SRM results that are better than those of IGSRM and M-SRM. For example, in the first example, the overall accuracy of the proposed method is 7.43–10.96% higher than that of the IGSRM method for different scale factors, and 1.09–3.44% higher than that of the M-SRM, while, in the second example, the improvement in overall accuracy is 2.42–4.92%, and 0.08–0.90%, respectively. The proposed method provides a general framework for combining multiple results from different SRM methods.

Highlights

Land cover classification is a major technique of mapping land cover types and monitoring their dynamics using remote sensing data
This paper proposed a new method of combining multiple super-resolution mapping (SRM) realizations from Indicator-geostatistics based super-resolution mapping (IGSRM) [33]
Several existing methods were used for both individual steps and the whole SRM process to fully evaluate the proposed method

Summary

Introduction

Land cover classification is a major technique of mapping land cover types and monitoring their dynamics using remote sensing data. Owing to the common occurrence of mixed pixels in remote sensing images, even in images with very high spatial resolution [5], spectral unmixing has been widely studied and is used in various applications [6,7,8]. Super-resolution mapping (SRM), or sub-pixel mapping has recently been developed and is widely accepted for estimation of the spatial distribution of land cover classes at sub-pixel scale [14]. SRM generates higher-resolution classification results from coarse-resolution class proportion images produced by spectral unmixing [15,16,17]. SRM provides a way to obtain land cover information at finer resolution from images with relatively coarse resolution. It has been used to monitor land cover change [28,29], to refine ground control points [30] and to calculate landscape pattern indices [31]

Results

Discussion

Conclusion