Geoscience-Data-Quality-for-Machine-Learning

A problem exists when building broad scale models, for example, Australia.

Disparate datasets, breaking them down into broad domains:

• Geophysics (Gravity, Magnetics, Radiometrics, Seismic, Electromagnetic, Induced Polarisation, Magnetotelluric...)
• Geology (Lithology, Stratigraphy, Structure, Hydro..)
• Remote Sensing (Landsat, ASTER, Sentinel...)
• Geochemistry (Rock, Soil, Water, Assay techniques...)

Variety of data layers:

• Direct observations
• Gridded Data
• Interpretations (Solid geology, SEEBase...)
• Derivations (e.g. ASTER band ratios, Rolling up of rock units...)
• Machine Learning Models (Regolith Depth...)
• Inversions

Quality

• Age of science
• Technology used
• Resolution (Pixel size, map scale, survey spacing, detection limits..)
• Survey Type
• Human ratings? e.g. 1-10
• Downsampling/Upsampling
• Missing data (Geophysic survey blanks, Remote sensing gaps on old satellites..)

Dimensionality

• 1
• 2
• 3
• 4
• more? (Depth Slices...)

Scale

• World
• Country
• State
• Region
• Local

Outputs

• Variance of different model runs

Categorisation

How, thinking in a raster fashion, to get a combined per-pixel Data Quality rating for a map output.

• Some sort of normalised ranking for each quality area?

• Weightings?

• Simple qualitative (3/2/1, Good/Average/Bad, High/Medium/Low or other ordinals).

• Exists / Missing

  Reference

  • https://www.researchgate.net/profile/Alan_Aitken/publication/326193704/figure/fig1/AS:646297606443016@1531100765653/