Data Model & Outputs

Input contract, dekad convention, and what the pipeline writes

Purpose

This page is the practical bridge between:

  • Quick Start / Tutorials (how to run),
  • User Guide (how to interpret), and
  • Methodology / Implementation (why each design choice was made).

If you are unsure whether your data is “ready” for the LSEQM+DL pipeline, what config.yml expects to find, or what each output NetCDF contains, start here.

Input Data Contract

Supported containers

  • Gridded inputs: xarray.Dataset (preferred) or xarray.DataArray, served from CF-compliant NetCDF.
  • Station inputs: pandas.DataFrame loaded from semicolon-separated CSV (location) and comma-separated CSV (daily observations).

Required coordinates and dimensions

Your gridded inputs must include:

  • time (datetime64)
  • lat (decimal degrees, ascending preferred)
  • lon (decimal degrees)

Preferred dimension order: (time, lat, lon). The loader checks lat orientation and flips descending grids automatically.

Temporal resolution

  • Daily timestep across all gridded inputs.
  • Continuous time axis: gaps are tolerated but flagged.
  • Recommended minimum: 20 years for stable GPD tail estimation.

Spatial conventions

  • Resolution: 0.1 deg for IMERG and the regridded CPC. CPC native is 0.5 deg.
  • Coordinate names: lat, lon (not y, x).
  • Mask alignment uses reindex(method='nearest'), not interp, so float32 / float64 dtype mismatches at AOI boundaries do not silently drop edge cells.

Required input files

File Grid Variable Units Role
imergl_file 0.1 deg precipitation mm/day The satellite field being corrected (IMERG Late Run V07)
imergf_file 0.1 deg precipitation mm/day Gauge-adjusted satellite baseline (IMERG Final Run V07) - comparison only
cpc_file 0.1 deg precip mm/day CPC-UNI regridded - per-pixel EQM target
cpc_native_file 0.5 deg precip mm/day CPC-UNI native - BCSD parameter fitting source
mask_file 0.1 deg land (1 = land, 0 / NaN = sea) – Applied at load time; ocean cells become NaN
station_file CSV (; separator) ID, ID_WMO, Station, Lon, Lat, Elevation, region, a1name, a2name – Station metadata for confidence mask + validation
station_data_file CSV (, separator, wide format) one column per WMO ID, indexed by ID, Date, JD mm/day Daily station observations for validation

Paths are declared in config.yml and support {main_dir} / {input_dir} placeholders. See Configuration.

Variable name conventions

Configurable through config.yml:

variable_names:
  imerg_precip_var: "precipitation"   # IMERG-L and IMERG-F
  cpc_precip_var:   "precip"          # CPC-UNI (both 0.1 deg and native 0.5 deg)
  mask_var:         "land"            # 1 = land, 0 / NaN = sea

Verify with ncdump -h <file> or xarray.open_dataset(<file>).data_vars before running.

Missing values

  • IMERG / CPC: NaN for missing.
  • Mask: ocean is NaN; land is 1.0. Applied via xr.where(mask == 1) so ocean pixels become NaN, not zero - this prevents ocean contamination of mean calculations.
  • BMKG station observations: sentinel value 8888.0 for missing (configurable via station_validation.missing_sentinel).
Quick sanity checks (recommended) 
import xarray as xr
from src.config import initialize_config

config = initialize_config('config_bali.yml')

for label, path in [
    ('IMERG-L', config.imergl_file),
    ('IMERG-F', config.imergf_file),
    ('CPC 0.1', config.cpc_file),
    ('CPC 0.5', config.cpc_native_file),
    ('Mask',    config.mask_file),
]:
    ds = xr.open_dataset(path)
    has_time = 'time' in ds.coords
    print(
        f"{label:8s}: dims={dict(ds.sizes)}, "
        f"time={str(ds.time.values[0])[:10] if has_time else 'n/a'}.."
        f"{str(ds.time.values[-1])[:10] if has_time else 'n/a'}"
    )
    ds.close()

If the time ranges line up across IMERG and CPC, and the mask grid matches the IMERG grid, you are ready to run Full Pipeline starting from nb02.

Data quality thresholds (per dekad per pixel)

The pipeline enforces these minimums during fitting. If a pixel fails any check, the output for that pixel becomes NaN (or zero) for the dekad.

Check Threshold What happens if violated
Non-NaN IMERG time series any non-NaN day Pixel skipped (NaN output)
Wet days in IMERG >= 5 (above wet_day_threshold = 1.0 mm/day) Output forced to zero (preserves dry climatology)
GPD exceedances >= 30 recommended K-Fold split may fail; configurable via n_splits_gpd_crossvalidate
Gamma fit shape / scale > 0 Pixel returns NaN
Very dry dekads

In peak dry season in highly seasonal regions, some pixels may have fewer than 5 wet days across the dekad. The pipeline outputs zero for those, which preserves the dry climatology but means the LSEQM step is effectively a no-op there. Distribution metrics will still be computed; they just reflect a degenerate distribution.

Dekad temporal convention

The pipeline correction parameters are fit per dekad (10-day window). Each calendar month has three dekads:

  • Dekad 1: days 1-10
  • Dekad 2: days 11-20
  • Dekad 3: days 21-end (28, 29, 30, or 31)

That gives 36 dekads per year. Fitting independently per dekad lets the framework follow the wet-dry seasonal cycle without smearing the monsoon onset / retreat.

Figure 1: The dekad temporal unit: each month splits into three windows (days 1-10, 11-20, 21-end), giving 36 dekads per year. All correction parameters are fit independently per dekad.

Calibration period

  • The full overlap between IMERG and CPC time axes is used (typically 2001-onward).
  • Calibration is applied per pixel, per dekad (each cell fitted independently for each of the 36 windows).
  • For Indonesia (2001-2025) this means ~25 samples per (pixel, day-in-dekad) - enough to fit gamma + GPD with K-Fold cross-validation.

Output Contract (What You Get Back)

The pipeline writes each correction stage to its own directory so any stage can be evaluated or replaced independently.

data/output/
  corrected_ls/         per-dekad NetCDF (Linear Scaling)
  corrected_lseqm/      per-dekad NetCDF (LS + EQM + GPD)
  corrected_lseqmdl/    per-dekad NetCDF (LSEQM + CNN refinement)
  trained_models/       .keras + .json per dekad
  station_density/      confidence_mask_station_density.nc4
  metrics_{method}/     per-dekad metrics NetCDF
  quality_{method}/     per-dekad QA NetCDF
  station_validation/   per-method CSVs
  figures/              QA / Taylor / station validation PNGs

Corrected precipitation

Return type: one NetCDF per dekad per method.

Filename pattern: {prefix}_{method_abbr}_corrected_imergl_month{MM}_dekad{D}.nc4 (e.g. bali_cli_lseqmdl_corrected_imergl_month01_dekad01.nc4).

Variable Type Units Description
precipitation float32 mm/day Corrected daily precipitation
lat, lon float64 deg Cell-centre coordinates on the 0.1 deg grid
time datetime64 UTC Daily timestamps for the dekad x all years

Global attributes record the correction method (ls / lseqm / lseqmdl), source files, framework version, and run timestamp.

Trained CNN models

trained_models/
  {prefix}_cnn_month{MM}_dekad{D}.keras       # Keras model
  {prefix}_cnn_month{MM}_dekad{D}_norm.json   # Normalisation constants

Load with tf.keras.models.load_model(path). Both files are required to apply the model.

Confidence mask

station_density/confidence_mask_station_density.nc4 - a single 2-D field on the 0.1 deg grid with values in [0, 1]. Built once per AOI from the station network; cached for the rest of the session.

Metrics

One per-dekad NetCDF per method:

metrics_{method}/{prefix}_metrics_{method}_month{MM}_dekad{D}.nc4

Contains the full WMO/TD-1485 + WMO-1317 metric set: RMSE, MAE, ME, Pearson r, Spearman rho, NSE, KGE, POD, FAR, CSI, percentile errors at 50/75/90/95/99, standard deviation ratio.

QA framework (CQI)

One per-dekad NetCDF per method:

quality_{method}/{prefix}_quality_{method}_month{MM}_dekad{D}.nc4
Variable Range Meaning
basic_statistical_quality 0-1 Bias / correlation / efficiency component
distribution_quality 0-1 Distribution-shape component
temporal_quality 0-1 Event-timing component
continuous_quality (CQI) 0-1 Weighted sum of the three components
categorical_quality 1-4 Poor / Fair / Good / Excellent classification
confidence_level 0-1 Station-density confidence

Station validation

station_validation/
  {prefix}_validation_by_station_{method}.csv         # 31 metrics x stations
  {prefix}_validation_by_region_{method}.csv          # aggregated by region
  {prefix}_multi_threshold_summary_{method}_month{MM}_dekad{D}.csv
  {prefix}_taylor_by_station_{method}.png

Figures

figures/
  qa/                  QA spatial maps (8 plot types x 36 dekads x 2 quality modes)
  taylor/              Pooled + per-dekad Taylor diagrams
  station_validation/  Per-station scatter, threshold curves, regional boxplots

Configuration tokens

config.yml uses three placeholder tokens that initialize_config() resolves at load time:

Token Resolves to
{main_dir} Repo root (directories.main_dir from config)
{input_dir} {main_dir}/data/input (or override via directories.input_dir)
{output_dir} {main_dir}/data/output (or override)
{method} Expanded per call: ls / lseqm / lseqmdl

So metrics_path_template: "{output_dir}/metrics_{method}" becomes data/output/metrics_lseqm/ when the pipeline runs for method='lseqm'.

NetCDF encoding (CF-1.8)

All output NetCDFs follow CF Conventions 1.8:

Property Value
Data type float32
Compression zlib level 4
_FillValue NaN
Conventions CF-1.8
lat.standard_name latitude, units degrees_north
lon.standard_name longitude, units degrees_east
time.units days since 2001-01-01
precipitation.standard_name lwe_precipitation_rate, units mm day-1

Reproducibility metadata on each output: framework_version, git_commit, config_file, run_timestamp, blend_alpha, gpd_threshold_percentile, saturation_count.

See Technical > CF Compliance for the full attribute table.

Where to go next

See Also

Back to top