CloudTrack-3D / METEOROLOGY.md
cloudTrack3D's picture
Upload CloudTrack-3D benchmark dataset
f175517 verified
|
Raw History Blame Contribute Delete
12 kB

Additional inputs for satellite cloud learning

Recommendation and project context

Start with ERA5 temperature and humidity profiles, then test dynamics and additional satellite channels separately. This is a proposed historical reconstruction experiment; no change to the training pipeline is included in this download task.

The local folders are ir10.5, vis0.6, and vfms. They contain 2,168 IR PNGs, 1,252 visible PNGs, and 766 ATL09 HDF5 files (including multiple product releases). IR coverage is 2024-11-01 through 2025-09-11. The downloader and training dataset use 25–40°N, 24–43°E. Labels describe cloud occurrence, layer count and heights along the three ICESat-2 beams. The current dataset can include earlier image frames and latitude/longitude channels.

Two approaches are useful:

  • ERA5 historical context (implemented): consistent archive covering this dataset, hourly fields on a 0.25° grid. Good for testing whether atmospheric structure improves cloud reconstruction. ERA5 is a reanalysis that assimilates observations; it is not a forecast available at the image acquisition time.
  • Archived operational forecasts (future work): needed to reproduce real-time inference. Select a forecast cycle actually available before each image, retaining initialization time, lead time and valid time. This needs a separate archive and availability audit for 2024–2025, not merely a current forecast downloader.

ERA5's nominal atmospheric model resolution is about 31 km; the CDS grid is 0.25°. It provides atmospheric context, not resolved cloud structure at satellite pixel scale. Sources: ERA5 pressure-level dataset, ECMWF ERA5 description.

Prioritized features

Priority Inputs Purpose / source
First Temperature, relative humidity, specific humidity, geopotential at pressure levels Thermal structure, moisture and height correspondence; ERA5 pressure levels
First Surface pressure, 2 m temperature/dewpoint, skin temperature, boundary-layer height Below-ground masking, near-surface saturation and low-cloud environment; ERA5 single levels
First Surface geopotential and land/sea fraction Terrain/coastal context; included in single-level download
Next Horizontal wind components and pressure vertical velocity Advection, shear and ascent; script's --dynamics option
Next Solar zenith angle, cyclic local solar time/day of year, satellite viewing angle Illumination and viewing geometry; derive from time, location and satellite geometry
Next FCI water-vapour 6.3/7.3 µm, IR 8.7/12.3/13.3 µm, NIR 1.6/2.2 µm and IR 3.8 µm Test moisture, phase and cloud-height-sensitive spectral information; acquire calibrated FCI products
Later Higher-resolution DEM, sea-surface temperature, CAPE and total-column water vapour Evaluate additional terrain, stability and moisture context after the core ablation

The channel proposal is based on the instrument and retrieval description in EUMETSAT's FCI Optimal Cloud Analysis ATBD. Actual historical product access and channel coverage must be checked before building that downloader. Current WMS PNG values are rendered image intensities; do not treat them as Kelvin or calibrated reflectance. Physical Level-1 products would allow meaningful brightness-temperature differences and temperature-profile comparisons.

Default pressure levels, in hPa: 1000 975 950 925 900 850 800 700 600 500 400 300 250 200 150 100. These emphasize the lower troposphere while spanning upper-tropospheric clouds. Add intermediate levels (e.g. 750, 650, 550, 350, 225, 175, 125) if height performance justifies the larger feature set. Pressure levels are not fixed altitudes.

Do not include ATL09 cloud flags, layer heights or confidence flags as input features: they are labels or label-quality metadata unavailable away from the tracks. Keep model cloud fraction/liquid/ice fields as a separate, explicitly named model-assisted baseline rather than mixing them into the first environmental experiment.

Download

The script uses the two CDS datasets above and ERA5 single levels. No credentials are needed to prepare the complete manifest:

python download_era5.py --dry-run

Default source discovery unions timestamps from both image folders and deduplicates them. It already includes the earlier frames present on disk. It does not validate PNG content or restrict the inventory to matched/quality-approved training granules. Use --dates-json with approved sample dates to narrow the download further.

For the present inventory the dry run produces 1,440 hourly grids and 626 requests (two requests per selected day). At 61 × 77 grid points, 16 levels, four profile variables and seven surface variables, the fields represent approximately 1.92 GB of float32 values before GRIB compression and metadata. This is a planning estimate, not a measured transfer size; CDS queue time may dominate. Avoid concurrent runs writing the same output directory.

Follow the official CDS API setup: create an account, accept each dataset's terms on its download page, and put the personal token in ~/.cdsapirc:

url: https://cds.climate.copernicus.eu/api
key: YOUR_PERSONAL_ACCESS_TOKEN

Install the client in your project environment and try one day before the full run:

python -m pip install 'cdsapi>=0.7.7'
python download_era5.py --start 2024-11-01 --end 2024-11-01
python download_era5.py
# Optional experiment with the additional profile variables:
python download_era5.py --dynamics

Other selections:

python download_era5.py --dates-json dates.json --offsets-minutes 0 -15 -30 -45 --dry-run
python download_era5.py --levels 1000 925 850 700 500 300 200 100 --dry-run
python download_era5.py --area 40 24 25 43 --output ./era5 --dry-run

Dates JSON follows the existing downloader's unusual ordering: [[year, day, month, hour, minute], ...], in UTC. Date filters select source samples inclusively, not every hour of a continuous date range. Offsets and bracketing can request the adjacent day outside the filter. Image paths default relative to the script, independent of the shell working directory.

Each request fetches only the required hours for one day and a regional subset. Files use native GRIB with JSON receipts and SHA-256 checksums. Request-specific filenames prevent changed variables/levels/areas from silently reusing old files. Interrupted transfers are not published; rerunning verifies completed files and retries incomplete work. A failure stops the run with a nonzero exit status. GRIB header/trailer checks detect basic transfer errors; ingestion must additionally verify expected variables, levels, timestamps and coordinates. Changed time selections can produce overlapping files: deduplicate valid times during ingestion.

Alignment and experiment plan

  1. Time: interpolate between the surrounding ERA5 hours at each image or beam observation time. The script requests floor/ceiling hours, including midnight rollover. For final track-level matching, retain each profile's actual timestamp; a filename start time or whole-granule mean is only an approximation. Audit the ATL09 GPS-to-UTC conversion against the product metadata before integration. If beam times exceed the image-time coverage, supply their times in a manifest too.
  2. Space: bilinearly sample ERA5 at track coordinates, or interpolate to each satellite patch. Preserve north/south orientation and use coordinates, not assumed array indices. Prefer a coarse context encoder or patch-level feature vector over expanding every weather field to a full 4096² image. Upsampling adds no information.
  3. Vertical coordinates and units: preserve temperature in K, RH in percent, specific humidity in kg/kg, winds in m/s, omega in Pa/s and pressure in Pa. Mask pressure levels where level_hPa * 100 > surface_pressure and provide the mask to the model. Convert geopotential to geopotential height with z / 9.80665; reconcile geometric/geopotential height and ATL09's vertical reference before comparing to cloud height. Use surface elevation when constructing AGL features. Keep missingness explicit. Do not blindly clamp ice-supersaturated RH to 100%.
  4. Derived features: test temperature lapse rates, humidity maxima, wind shear, freezing-level height and lifting-condensation-level estimates after raw profiles. Fit normalization only on the training split. Missing/nighttime visible imagery needs an explicit mask; darkness is not evidence of clear sky.
  5. Ablations: compare satellite-only; satellite + T/RH/geopotential; plus q and surface features; plus dynamics; plus extra spectral channels. Also train a weather-only baseline to quantify how much skill comes from meteorological context. Keep samples and splits identical across experiments.
  6. Evaluation: split by contiguous days/orbits with a temporal gap, deduplicate ATL09 releases, and keep all neighboring frames of a sample in one split. The existing every-fifth-granule split can share weather systems across partitions. Report cloud precision/recall, height MAE/bias conditional on valid cloudy labels, and results by altitude, day/night and season. Quality/attenuation flags should control supervision; an obscured lower layer is not necessarily clear sky.

ERA5 uses retrospective information; even choosing only the preceding hour does not make it an operationally available input. Any claim about real-time performance requires a separate test using forecasts available at the actual inference time.

Offline checks: python -m unittest -v test_download_era5.py. A live CDS download and scientific field inspection remain required after setup.

Complete date range, all hours

download_era5_all_hours.py implements the complete-range request: every date from 2024-11-01 through 2025-09-11 inclusive, at all 24 UTC hours (7,560 timestamps). It uses the same study area and 16 pressure levels, with all seven profile variables (temperature, relative/specific humidity, geopotential, U/V wind and vertical velocity). The seven surface variables listed above are included as separate single-level jobs. This selection includes dates and hours without a local VFM observation.

python download_era5_all_hours.py --dry-run  # Prepare era5_all_hours/requests.json
python download_era5_all_hours.py            # Submit and download; rerun to resume

The manifest contains 154 requests, each covering one month and one parameter: seven pressure-level and seven surface requests per month. Each pressure-level request contains at most 11,904 fields (31 days × 24 hours × 16 levels), below the 120,000-field limit documented in the CDS documentation. The prepared manifest has been checked offline for complete date coverage, valid calendar dates, all hours, 16 levels and seven profile variables; it has not been submitted to CDS as part of preparing the request.

Run ./download_era5_monthly.sh to process each month separately. Add --dry-run to prepare the 11 monthly manifests without submitting requests. Each parameter is saved to its own GRIB file. Failed parameters are reported while remaining parameters and months continue; rerunning verifies and skips completed files. September still ends on September 11. Existing combined-variable files are retained, but the new parameter-specific requests download their own files.