<p>Ocean remote sensing datasets often have the problem of missing values due to various reasons. However, many scientific applications require spatiotemporal seamless data. Data reconstruction methods are commonly used to obtain such gap-free datasets. In reconstructing satellite remote sensing data, randomly masking original data for progressive cross-validation is a common method to indicate the performance of reconstruction. In this study, the accuracy of this validation method is analysed. We artificially constructed two data missing patterns using the sea surface temperature (SST) data in the East China Sea, one simulating natural cloud coverage and the other randomly masking the same percentage of original data. The results of reconstruction for the two types of masking were compared. The root mean square error (RMSE) of dataset that simulate real cloud coverage is more than 50% higher than that of the dataset randomly masking data, regardless of the data missing rate. This result implies that the error of satellite data gap-filling is underestimated when random masking of original data is applied for progressive cross-validation, which should be treated with care in applications.</p>