Modification of Statistical Metric Biases in Large-Region and Long-Time-Series Landsat Dataset Due to Insufficient Observations

Landsat time series, as the longest fine resolution dataset, has the most significant limitation of relatively low temporal frequency (16 days). However, the presence of clouds, cloud shadows, snow, and the failure of the sensor further reduces the amount of clear data, resulting in insufficient observations, which can easily lead to biases in statistical metrics (i.e., maximum, mean, percentiles, etc.). In this study, we took advantage of Google Earth Engine and proposed a Statistical Time-series biAs Modification Model (STAMM) that can generate real Landsat statistical metrics (i.e., Landsat statistical metrics based on sufficient observations) in large regions. STAMM can also quantitatively evaluate the bias of statistical metrics due to insufficient observations. Results show that the original Landsat NDVI 75 ^th percentiles over both the northeast and northwest of China overestimated about 10%~30% compared with real Landsat percentiles; while in the southwest of China, it underestimated about 10%. This issue has persisted for the past 20 years. As for other percentiles, the amount of bias depends on the data distribution. The bias of the 95 ^th percentile is relatively small when there are more clear observations in the high-value period. However, at this time, the bias of the 50 ^th percentile bias is relatively large. Taking the Sentinel percentiles as references, STAMM effectively improves the accuracy of Landsat percentiles (i.e., the RMSE decreased from 0.065 to 0.045). It is supposed to be able to provide consistent and accurate Landsat percentiles in large-region and long-time-series for better studying the inter-annual change of Earth's surface.