Day 4

Detailed paper information

Back to list

Paper title A new MODIS-based global lake ice cover data record (2000-2020): A contribution to ESA’s Lakes_cci project
  1. Yuhao Wu H2O Geomatics Inc. Speaker
  2. Claude R. Duguay University of Waterloo
  3. Daoyuan Zhang H2O Geomatics Inc.
Form of presentation Poster
  • A5. Climate
    • A5.02 The role of Earth Observation in climate services
Abstract text Lake ice cover (LIC), a thematic variable under Lakes as an Essential Climate Variable (ECV) that is a robust indicator of climate change and plays an important role in lake-atmosphere interactions at northern latitudes (i.e. heat, moisture, and gas exchanges), refers to the area (or extent) of a lake covered by ice. Ice dates and ice cover duration at the pixel scale (ice-on and ice-off) and lake-wide scale (complete freeze-over (CFO) and water clear of ice (WCI)) can be derived from lake ice cover data (Duguay et al. 2015). Determination of ice onset (date of the first pixel covered by ice), CFO, melt onset (date of the first pixel with open water), and WCI are of most relevance to capture important ice events during the freeze-up and break-up periods. Duration of freeze-up and break-up periods and duration of ice cover over a full ice season can be determined from these dates. The generation of a LIC product from satellite observations requires the implementation of a retrieval algorithm that can correctly label pixels as either ice (snow-free and snow-covered), open water, or cloud. The LIC product v2.0 generated for Lakes_cci ( uses MODIS Terra/Aqua data to provide the most consistent and longest daily historical record globally to date (2000-2020). The new product provides three bands: Band 1 - lake ice cover flag (lake forms or does not form ice); Band 2 - lake ice cover class (open water, ice, cloud, and bad); and Band 3 - lake ice cover uncertainty (% accuracy for each of open water, ice and cloud classes).
In the first step of production, the Canadian Lake Ice Model (CLIMo) was applied to help determine which lakes of the Lakes_cci harmonized product (total 2024 lakes), which includes four other variables (water level, water extent, surface water temperature, and water-leaving reflectance), could have formed ice or have remained ice-free at any time over the 2000-2020 period. This step can correct false detection of ice in summer in the situation of dry lakebeds and reduce the computational cost of the production. CLIMo (Duguay et al. 2003) is a one-dimensional thermodynamic model capable of simulating ice phenology events, ice thickness and temperature, and all components of the energy/radiation balance equations during the ice and open water seasons at a daily timestep. Input data to drive CLIMo include mean daily air temperature (°C), wind speed (m s-1), relative humidity (%), snowfall (or depth) (m), and cloud cover (in tenth). Here, European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis hourly data on single levels (0.25-degree grid) were used to generate inputs required for CLIMo simulations for each of the 2024 lakes. Lake ice depth data provided by ERA5 were also utilised to check for the possible formation of ice on any of the lakes. Ice cover was deemed possible to have formed on a lake if ice depth was determined to have reached a thickness greater than 0.001 m on any day from either CLIMo or ERA5. Additionally, as a third check, a number of lakes (largely located at the southern limit of where ice could potentially form during a cold winter in the Northern Hemisphere and in mountainous regions of both the Northern and Southern hemispheres) were inspected manually through interpretation of MODIS RGB images to determine if any of these lakes had formed ice between 2000 and 2020. As a result of the process described above, presented in the variable of lake ice cover flag of the LIC product v2.0, 1391 of 2024 lakes were flagged as forming an ice cover and 633 not forming any ice over the 2000-2020 period. Once flagged, only lakes determined to form ice were selected to perform lake ice classification from MODIS data by the main processing chain.
MODIS TOA reflectance bands and the solar zenith angle (SZA) band are used for feature retrieval (i. e. for labeling as water, ice, or cloud) (Wu et al. 2021). The reflectance bands are MOD02QKM at 250 m (band 1: 0.645 µm and band 2: 0.858 µm) and MOD02HKM at 500 m (band 3: 0.469 µm; band 4: 0.555 µm; band 5: 1.240 µm; band 6: 1.640 µm; band 7: 2.130 µm) resolutions. Prior to retrieval, pixels of interest are identified as “good” or “bad” using quality bands from the original MODIS TOA reflectance product. The pixels with SZA greater than 85 degrees are identified as “bad”. Pixels of interest are classified and labelled as either cloud, ice, or water from a random forest algorithm (Wu et al. 2021). Labelled pixels are resampled to the output grid. The processing chain has been revised for Lakes_cci to generate the output grid based on specifications of the harmonized product (1/120th degree latitude/longitude; ca. 1 km). Aggregation is performed by taking a majority vote between ice and water, ties broken by selecting water. If there are zero ice and water pixels, then the cell is labelled as cloud if there are non-zero cloud pixels; otherwise, the output cell is labelled as “bad”. The variable of lake ice cover class presents the retrieved labels.
Validation of the LIC V2.0 product has been performed through the computation of confusion matrices built on independent statistical validation. The reference data for validation were collected for water, ice, and cloud as AOIs from the visual interpretation of the MOD02/MYD02 false color composite images (R: band 2, G: band 2, B: band 1) with a 250 m spatial resolution. A total of 10,075,081 pixels taken from 229 MOD02 swaths over Great Slave Lake and Lake Onega were used to conduct classification assessment of the LIC product generated by MODIS Terra. There is no notable difference in the accuracy of the product between the break-up (98.14% overall accuracy) and freeze-up (96.83% overall accuracy) period. Additionally, 1,665,188 samples collected from MYD02 false color composite images were applied for the validation of the LIC product produced from MODIS Aqua. The overall accuracy of 97.68% reached with Aqua data is comparable to that obtained with MODIS Terra data. Further evaluation of the Lakes_cci LIC V2.0 product and its comparison with other products is planned in the future, and with input from the user community.

Duguay, C. R. Bernier, M. Gauthier, Y. & Kouraev, A. (2015). Remote sensing of lake and river ice. In Remote Sensing of the Cryosphere, Edited by M. Tedesco. Wiley-Blackwell (Oxford, UK), 273-306.
Duguay, C.R., Flato, G.M., Jeffries, M.O., Ménard, P., Morris, K. & Rouse, W.R. (2003). Ice cover variability on shallow lakes at high latitudes: Model simulations and observations. Hydrological Processes, 17(17), 3465-3483.
Wu, Y., Duguay, C.R. & Xu, L. (2021). Assessment of machine learning classifiers for global lake ice cover mapping from MODIS TOA reflectance data. Remote Sensing of Environment, 253, 112206,