[Insert Figure 1 about here]
Figure 1: Location, across Eurasia, of the 1,098 lakes and reservoirs (blue) selected for the study and the climatic zones where the lakes and reservoirs are situated. Typical lakes within each climatic zone are shown in red.
2.2 MODIS LST product preprocessing
Daily MODIS-Terra/Aqua LST (level 2, collection 6) is an option to characterize LWST. However, in order to track thermal dynamics of water bodies, cloud contamination must be accounted for and it may cause a large proportion of invalid pixels in the MODIS data product. Therefore, MODIS LST 8-day composite product (level 3, MOD11A2), which integrates averaged daily LST observations of eight days with a nominal resolution of 1 km2, was considered suitable for this investigation. The 2015 MOD11A2 data for the Eurasian continent were downloaded from NASA’s Land Processes Distributed Active Archive Center (LP-DAAC) using the Warehouse Inventory Search Tool (WIST). To track inter-annual temperature variation, 18 great lakes (Figure 1) with obvious water surface area variations in the 15-year period from 1 January 2001 to 31 December 2015 were identified. Then, the MODIS Reprojection Tool (MRT) (LP DAAC, 2001; Sioux Falls, South Dakota) was used to extract 8-day composite LST data. MODIS LST over lake surfaces was evaluated against in situ -measured temperature with absolute differences in the range of 0.8–1.9 K (O’Reilly et al., 2015). As for other lakes in China, MODIS LST products were validated with direct temperature measurements of lake waters made in 2015.
2.3 LST data analysis
After mosaicking using MRT, the MOD11A2 images were projected onto Albers in GeoTIFF format using nearest neighbor interpolation. The daytime LST (overpass time is circa 10:30 am local time), nighttime LST (overpass time is circa 22:30 pm local time), and corresponding quality control images (QC-daytime, QC-nighttime) were extracted. Due to the strong contrast of land and water spectral features in the near infrared (NIR) and green band, thresholding was used to extract boundaries of water bodies based on the spectral ratio of NIR/Green using Landsat OLI imagery. The shape file for water surface areas greater than 25 km2 was generated using Landsat-OLI imagery acquired in 2010, by referencing it to MODIS reflectance product (MOD09Q1) in seeking a robust-matching MODIS LST product. To minimize seasonal or inter-annual difference on Eurasian lakes measurement, the algorithm LakeTime, which is suitable for lake mapping at continental and global scales, was used to select Landsat scenes based on seasonally-defined climatic and hydrological variables (Lyons & Sheng, 2017; Sheng et al., 2016). LakeTime is robust and reliable for delineating lake boundaries using lake-specific NDWI thresholds (Sheng et al., 2016). LakeTime can not only reduce the impacts of seasonal and inter-annual variability, but also mitigate cloud contamination effects, and significantly reduce labor costs in the subsequent QA/QC process (Sheng et al., 2016). To avoid possible land-water interface contamination and fluctuation in lake surface areas, a 500-m offshore buffer zone was generated to exclude LST pixels along shoreline zones (Ke & Song, 2014). Despite using the 8-day composite product (MOD11A2), invalid pixels could still arise from cloud contamination, so additional procedures were taken to remove questionable pixels in the final MODIS LST product (Ke & Song, 2014; Deo & Şahin, 2017). To do this, unreliable LST pixels over lakes were removed using QC information, followed by median filtering of time-series LST data stacks (Ke & Song, 2014). In the first step, all pixels having an average LST error of < 1 K (i.e., QC = 0, 1, 5, 17, and 21) were kept, while pixels with other QC values were removed through referencing to LST data quality flags stored in the QC file. In the second step, a data cube was generated by stacking 46 extracted LST layers according to the Julian day sequence of the generated product, to which a median filtering algorithm iterated pixel by pixel to eliminate any questionable pixels in the LST images (Figure 2). In most cases, the 8-day composites consisted of at least 70% valid pixels. Over the 16-year study period, only 97 images (out of 3,840 images or 2.5%) had less than 70% valid pixels. The months with affected images were consistent over the years. Most affected images were from April to June, and from September to mid-December. Furthermore, there were more valid images in the nighttime MODIS LST product (Ke & Song, 2014). Finally, to track temporal trends in LWST for the selected lakes in Eurasia, annual rates of temperature change from MODIS LST data were regressed against the data acquisition year using linear regression (Zhang et al., 2014).
y = a + bx +e (1)
where y represents the MODIS LST, x denotes the time-series of years, e represents the residuals, a is the intercept, and b is the rate of temperature change. The coefficients a and b were determined through least squares fitting.