wet_snow_threshold_localization

This repository is for SAR binary wet snow threshold localization google slides with some ideas and background: https://docs.google.com/presentation/d/1y1C19CQyOyy0HFLKyf9x4ZlzIvoxL7Bz3Yzis7T1P6w/edit?usp=sharing

the problem with with binary wet snow detection

for SAR binary wet snow detection, there are a ton of methods for thresholding (e.g. for SAR: -1dB to -2dB to -3dB, VV, VH, VV/VH combinations weighted on LIA, vegetation, different reference images, etc), but very little spatially distributed wetness data to validate against :(

an idea for an empirical approach

• for a ratio image with wet snow (choose a scene where most/all of snow is wet), compare histograms of dB change values
  • compare dB drop within snow / no snow classes using optical imagery (figure on the right from Nagler et al. 2016)
  • hopefully find bimodal distribution in dB drop (snow / no snow) which would allow us to optimize a wet snow threshold
• iterate at locations with different landcover/veg, different snow classification types, snow depths, variable topography (to get at incidence angle dependence)
  • build up a huge datacube containing [VVdB drop,VHdB drop, snow/nosnow, LIA, LCC/veg, snowclass, where in the melt season are we, fSCA, othervars?] for each pixel at all locations
• find trends, optimize, or use ML/DL approach
  • will help us characterize how binary wet snow threshold changes with these variables allowing localization of this threshold
  • end goal would be some sort of heuristic: given a pixel with x vegetation/LC type, y snow class, z incidence angle, we expect at least a backscatter change of -XdB in VV and -YdB in VH from the reference image to the wet snow image
  • if it works, could be the foundation of a “smart” binary wet snow algorithm, where the binary wet snow threshold is variable across a scene as a function of [landcover class, snow class, local incidence angle, etc]

To do list:

• notebook to build dataset
  • data sources
    • for Sentinel-1 RTC data: https://github.com/egagli/sar_snowmelt_timing/blob/main/dev/collaborations/sierra_nevada/sierra_nevada_binary_wet_snow_map_timeseries.ipynb
    • LIA: https://github.com/egagli/generate_sentinel1_local_incidence_angle_maps
    • LC: https://planetarycomputer.microsoft.com/dataset/esa-worldcover
    • FCF: ?
    • Snow classification: https://nsidc.org/data/nsidc-0768/versions/1
    • Snow cover (snow / no snow): https://planetarycomputer.microsoft.com/dataset/modis-10A1-061 (or 8 day)
  • loop over a bunch of areas, multiple years, save a bunch of these as zarr files?
    • scene selection criteria
      • global? what are the extents of our individual datasets
  • questions to consider
    • reference scene
      • time span?
      • mean or median?
      • multi-year?
    • how to find when most snow is melting
      • morning / afternoon overpass?
      • temperature data?
      • remote sensing thermal?
      • in-situ weather station?
      • reanalysis?
• [] notebook to ingest saved zarr files, stitch all together. possibly ouput as singular large dataframe with each row being a single pixel
• [] notebook to analyze full datas
  • aggregate statistics
  • histograms analyzing each variable
    • for example, LIA on x axis, dB change on right axis, then two box plots for each LIA, one for snow and one for no snow, maybe each of those broken into VV and VH
  • test VV and VH for higher wet snow seperability
  • given snow / no snow distributions, which threshold provides the most seperability between distributions
    • create confusion matrix based on optimal threholds
  • analysis of variance
  • multiple linear regression
  • unsupervised machine learning
    • kmeans clustering
  • deep learning approach

critiques: ratio values can be influenced by (that we don't analyze) time of year snow depth!! stratigraphy roughness grain size temperature