Evaluation of Three Error-Correction Models Based on the Matched Pixel Scale Ground “Truth”: A Case Study of MCD43A3 V006 Over the Heihe River Basin, China

Abstract

Validation merely provides the “impression of accuracy” of satellite products but does not deal with their errors. There have been several error correction/adjustment models. Nevertheless, their performances have not been evaluated and compared. In this study, three typical models, namely random forests (RF), cumulative distribution function (CDF), and Kalman Filter (KF) were comprehensively evaluated based on the pixel scale ground “truth” regarding their ability to correct errors of coarse-resolution satellite albedo products. MODIS albedo product (i.e., MCD43A3 V006) was utilized as an example due to its widespread use and application. These three models all show significant improvements regarding the accuracy of the corrected MCD43A3. RMSEs decreased from 0.037 to 0.020, 0.021, and 0.025 for RF, CDF, and KF, respectively. Biases were reduced from -0.018 to 0.004, -0.001, and -0.001 for these three models, respectively. And R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> was increased from 0.585 to 0.849, 0.823, and 0.764 for RF, CDF, and KF, respectively. Generally, RF shows the best overall performance, followed by CDF and KF. These three models are more adept at handling the bias of MCD43A3 than their consistency with respect to the pixel scale ground “truth”, and the improvement is the most significant at the sites with large errors. Nevertheless, the performance of RF shows dependence on both the number and representativeness of training samples. When these conditions were not satisfied, CDF performs best in this situation. Regarding the stability of their performance, RF performs better in reducing RMSE while CDF performs better in reducing Bias and improving consistency.