A tabular dataset ready for machine learning applications

On Zenodo, I have uploaded a dataset (DOI: 10.5281/zenodo.10048875) containing snow depth measurements retrieved from ICESat-2. In my previous work, I described the process as follows:

This formula represents the discrepancy between ICESat-2 and DEMs. To obtain precise snow depth measurements, the primary effort involves minimizing skewness and noise of \(dh\), as detailed in my previous post. Consequently, we derived snow depth from well-corrected DEMs, as available in the Zenodo dataset:

This dataset is a test run on snow depth retrieval workflow based on ICESat-2 (publication is under preparation).

Advantages

• Open Access and Free: The ICESat-2 dataset is globally open-accessible and free of charge.
• Subgrid Variability: The dataset offers relatively high resolution, with an ideal footprint size of 13 - 17 m. It may be the only data source that contains subgrid variability of snow depth.

Limitations

   

   

   

• Sparse Sampling: The sparse nature of the sampling makes it challenging to produce a completed snow depth map. In a year cycle, there is a 4 times repeat orbit with revisited tracks at 7.2 km and 14.4km away in different seasons. For this reason, we have to use a reference ground to bridge the gap between snow-on and snow-off segments to retrieve snow depth. The sparse sampling also leads to an unbalance of the dataset, making it more challenging to predict the correct distribution of the snow depth. Figure 1 shows the sampling in Hardangervidda each month, there are obvious gaps (15 km wide) during the snow season. Therefore we can expect more difficulty in this area.
• Sensitivity to bias correction. The accuracy of ICESat-2 snow depth is highly dependent on the bias correction process.
  • Vegetation introduces significant uncertainties in both ICESat-2 and DEMs.
  • Permanent ice and inland water areas, which cannot be used as reference surfaces, must be excluded. However, many DEMs include seasonal snow patches, particularly in high elevations, which are not indicated by current permanent snow/ice masks. Similarly, lake masks may not be entirely suitable for all DEMs that are acquired in different seasons.
  • In practice, many DEMs are old enough to accumulate many changes on the surface, such as landslides, erosion, and brushwood…All these require careful co-registration and bias correction.
• ICESat-2 Ground-Finding Algorithm:
  • The parameterization of ICESat-2’s ground-finding algorithm significantly impacts retrieved snow depth. This algorithm uses a probability distribution function (PDF) of photons reflecting off the surface to determine exact elevations in the ATL08 product. If the algorithm or applied slope correction tends to capture or ignore specific signals, or if the footprint size averages out specific landforms, this can lead to systematic or scaling biases larger than the native footprint size. Under unideal conditions, the algorithm has produced lower estimation/interpolation, leading to negative skewness.
• Lack of Validation:
  • Due to scaling and representativeness issues, it is challenging to compare the retrieved snow depth with ERA5 Land data (native 9km resolution, including permanent ice), or other gridded snow products.
  • Large-scale snow surveys for validation are costly, making it difficult to validate the retrieved snow depth across various terrains.

Example of Application

I am excited to unveil an application based on this dataset, focusing on extracting subgrid variability and its utilization in downscaling processes. The following is a flowchart illustrating this application:

   

   

   

Example of Downscaled Snow Depth