Performance assessment of ESTARFM with different similar-pixel identification schemes

Abstract

Remote sensing images with high-spatial and high-temporal (HSHT) resolution are difficult to be consistently achieved with a single polar orbit satellite because of the trade-off between spatial and temporal resolution. However, blending algorithms have been developed to synthesize HSHT images from low-spatial but high-temporal resolution images and high-spatial but low-temporal resolution images. One example is a widely used weight-function-based method known as the enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM). Similar-pixels play a key role in the prediction accuracy of this model, and their identification is influenced by the method used, the number of classes, and the moving window size. However, influence of these parameters on the final prediction accuracy has not been thoroughly examined. Therefore, this study assesses the accuracy of ESTARFM when similar-pixels are identified by two separate methods using different numbers of classes and moving window sizes. The results indicated that the class-image-based method outperformed the threshold-based method in most cases and the increase of moving window size generally improved the accuracy for both methods. However, the improvement was minor and accuracy degraded in some cases if the moving window was too large. In addition, the ESTARFM program was optimized in this study, and its computational efficiency was improved compared with previous studies.