Merge branch 'develop' into fix/chirps-preprocess-dates

Author: sarahkanee

CHANGELOG.md

@@ -10,6 +10,7 @@ For more information about this file see also [Keep a Changelog](http://keepacha
 
 ### Added
 - Added ensemble calibration diagnostics to `inst/ilamb/` in PEcAn.benchmark (rank histogram, spread-skill ratio, coverage, reliability) for assessing whether an ensemble's spread is well calibrated against observations, complementing the ensemble-mean benchmarking.
+- Added regional calibration diagnostics to `inst/ilamb/` in PEcAn.benchmark: break the ensemble calibration assessment down by land cover class and by EPA/CEC ecoregion, reusing the ensemble calibration diagnostics, with figures.
 - Added an ESA CCI biomass benchmark extension to `inst/ilamb/` in PEcAn.benchmark: a converter for the ESACCI Biomass v7.0 product (2015-2024, with per-pixel uncertainty) to ILAMB-compatible netCDF, plus an observation-error test showing the ensemble overconfidence persists against a modern benchmark and after accounting for observation uncertainty.
 - Added `make_scorecard.sh` and documentation to `inst/ilamb/` in PEcAn.benchmark for generating and serving the ILAMB HTML scorecard.
 - New function `PEcAn.utils::netcdf2df()` flattens all dims and vars of a netCDF into a dataframe,

modules/benchmark/inst/ilamb/calibration/.gitignore

@@ -0,0 +1 @@
+*.png

modules/benchmark/inst/ilamb/calibration/README_regional.md

@@ -0,0 +1,72 @@
+# Regional Calibration Diagnostics
+
+Breaks the ensemble calibration assessment of a carbon reanalysis down by region,
+rather than reporting a single domain-wide number. It answers a question the
+domain-wide diagnostics cannot: does the trustworthiness of the ensemble spread
+vary by biome and by ecoregion?
+
+This builds directly on the ensemble calibration diagnostics in
+`ensemble_calibration.py`, computing the same spread-to-error
+ratio and coverage, but on subsets of sites grouped two ways: by land cover class
+and by EPA/CEC ecoregion.
+
+## What it does
+
+At a set of point locations (here, the sites of a state data assimilation
+reanalysis), the ensemble and an observational benchmark are sampled, and the
+calibration diagnostics are computed per group:
+
+- **`regional_diagnostics.py`** samples the per-member downscaled maps and the
+  benchmark at each site and computes calibration per land cover class. Its
+  `stratified_calibration(members, obs, groups, labels)` engine is grouping
+  agnostic and is reused for the ecoregion breakdown.
+- **`ecoregion_join.py`** assigns each site an EPA/CEC North American ecoregion
+  (Level 1 and Level 2) by point-in-polygon, then runs the same engine on the
+  ecoregion grouping.
+- **`regional_figures.py`** renders choropleth maps of the ecoregions shaded by
+  calibration, bar charts by land cover and by ecoregion, and a map of the sites
+  coloured by land cover class.
+
+## Interpretation
+
+The metrics are the same as in the calibration module: a spread-to-error ratio
+near 1 indicates a well-sized ensemble, well below 1 indicates overconfidence,
+and coverage is the fraction of observations inside the ensemble 90 percent band
+(near 0.90 when calibrated). Reporting these per region reveals structure that
+the single domain-wide number hides, for example whether the ensemble is
+overconfident uniformly or only in particular biomes.
+
+## Dependencies and data
+
+The stratification engine is pure NumPy plus the calibration module. The
+ecoregion join additionally requires `geopandas` and the figures additionally
+require `matplotlib`.
+
+The ecoregion breakdown needs the CEC "Ecoregions of North America" Level 1 and
+Level 2 shapefiles, a free download from the EPA
+(https://www.epa.gov/eco-research/ecoregions-north-america, files
+`na_cec_eco_l1.zip` and `na_cec_eco_l2.zip`). Use the full North American files
+rather than the CONUS-only versions bundled with some tools, so that the boreal
+and Arctic north is covered. Unzip them and point `ECO_DIR` in
+`ecoregion_join.py` at the folder containing `NA_CEC_Eco_Level1.shp` and
+`NA_CEC_Eco_Level2.shp`.
+
+## Notes
+
+The paths in these modules (the SDA output directory, the downscaled-map layout,
+the benchmark locations, and `ECO_DIR`) reflect one particular setup and are
+illustrative; adapt them to your own data. The land cover class labels in
+`LANDCOVER_LABELS` are the MODIS MCD12Q1 PFT classification used in the North
+American carbon reanalysis (Zhang et al.), with the class integers 1-8 following
+the standard MODIS PFT coding, confirmed against the site geography. The labels
+are display strings only and do not affect the computed numbers.
+
+## Testing
+
+`test_regional_diagnostics.py` checks the `stratified_calibration` engine on
+synthetic data with known per-group calibration, without touching any real data
+files. Run with:
+
+```
+python test_regional_diagnostics.py
+```

modules/benchmark/inst/ilamb/calibration/calibration_figure.png

@@ -0,0 +1,116 @@
+"""
+Assign each site an EPA/CEC North American ecoregion (Level 1 and Level 2) by
+point-in-polygon, then run the same stratified calibration engine on the
+ecoregion grouping.
+
+The ecoregion polygons and projection follow PEcAn's EPA_ecoregion_finder. The
+shapefiles are the CEC "Ecoregions of North America" Level 1 and Level 2 files,
+which cover all of North America (the CONUS-only versions bundled with some
+tools drop the boreal and Arctic north). They are a free download from the EPA:
+
+    https://www.epa.gov/eco-research/ecoregions-north-america
+    (na_cec_eco_l1.zip and na_cec_eco_l2.zip)
+
+Unzip them and point ECO_DIR below at the folder containing
+NA_CEC_Eco_Level1.shp and NA_CEC_Eco_Level2.shp. Requires geopandas.
+
+The WATER category is treated as unmatched. Sites that fall outside every
+polygon (open ocean, a few coastal points) are also left unmatched.
+"""
+
+import os
+import numpy as np
+import geopandas as gpd
+from shapely.geometry import Point
+
+import ensemble_calibration as ec
+from regional_diagnostics import (
+    load_sites, sample_ensemble, sample_benchmark,
+    stratified_calibration, VARIABLES, VAR_LABEL,
+)
+
+# Folder holding NA_CEC_Eco_Level1.shp and NA_CEC_Eco_Level2.shp (adapt this).
+ECO_DIR = "ecoregions_na"
+LEVELS = {
+    "L1": ("NA_CEC_Eco_Level1.shp", "NA_L1NAME"),
+    "L2": ("NA_CEC_Eco_Level2.shp", "NA_L2NAME"),
+}
+EXCLUDE = {"WATER"}  # not a real ecoregion
+
+
+def isnull(arr):
+    """Elementwise None-or-NaN test for an object array of region names."""
+    return np.array([(x is None) or (isinstance(x, float) and np.isnan(x))
+                     for x in arr])
+
+
+def assign_ecoregions(lon, lat):
+    """Return {level: name_array} assigning each site an ecoregion by polygon."""
+    pts = gpd.GeoDataFrame(geometry=[Point(x, y) for x, y in zip(lon, lat)],
+                           crs="EPSG:4326")
+    out = {}
+    for level, (fname, name_col) in LEVELS.items():
+        poly = gpd.read_file(os.path.join(ECO_DIR, fname))
+        # the shapefile carries its own equal-area CRS; reproject to lon/lat
+        poly = poly.to_crs("EPSG:4326")[[name_col, "geometry"]].copy()
+        joined = gpd.sjoin(pts, poly, how="left", predicate="within")
+        # a point can fall in overlapping polygon parts; keep the first per point
+        joined = joined[~joined.index.duplicated(keep="first")].sort_index()
+        names = joined[name_col].values.astype(object)
+        names = np.array([None if (n in EXCLUDE) else n for n in names],
+                         dtype=object)
+        out[level] = names
+    return out
+
+
+def _groups_from_names(names):
+    """Map region names to integer ids and a label dict for the engine."""
+    uniq = sorted(set(names[~isnull(names)]))
+    name_to_id = {n: i for i, n in enumerate(uniq)}
+    labels = {i: n for n, i in name_to_id.items()}
+    groups = np.array([name_to_id.get(n, -1) for n in names])
+    return groups, labels
+
+
+def run_level(level_name, names, ens_cache, min_sites=30):
+    """Print the stratified calibration by ecoregion at one level."""
+    groups, labels = _groups_from_names(names)
+    for var, (map_var, bglob, bvar, scale) in VARIABLES.items():
+        members, obs = ens_cache[var]
+        print(f"\n{VAR_LABEL[var]} ({var}) by {level_name} ecoregion")
+        print(f"  {'ecoregion':45s} {'n':>5s}  {'ratio':>6s}  {'cov90':>6s}")
+        rows = sorted(stratified_calibration(members, obs, groups, labels),
+                      key=lambda r: -r["n"])
+        for row in rows:
+            if row["n"] < min_sites:
+                continue
+            lab = row["label"][:44]
+            if np.isnan(row["ratio"]):
+                print(f"  {lab:45s} {row['n']:5d}  (too few)")
+            else:
+                print(f"  {lab:45s} {row['n']:5d}  {row['ratio']:6.3f}  {row['cov90']:6.3f}")
+
+
+def main():
+    lon, lat, lc = load_sites()
+    eco = assign_ecoregions(lon, lat)
+    l1, l2 = eco["L1"], eco["L2"]
+    n_l1 = len(set(l1[~isnull(l1)]))
+    n_l2 = len(set(l2[~isnull(l2)]))
+    unmatched = int(isnull(l1).sum())
+    print(f"assigned ecoregions to {len(lon)} sites: "
+          f"{n_l1} L1 regions, {n_l2} L2 regions, {unmatched} sites unmatched\n")
+
+    ens_cache = {}
+    for var, (map_var, bglob, bvar, scale) in VARIABLES.items():
+        ens_cache[var] = (sample_ensemble(map_var, lon, lat),
+                          sample_benchmark(bglob, bvar, scale, lon, lat))
+
+    print("=" * 72); print("LEVEL 1 ECOREGIONS"); print("=" * 72)
+    run_level("L1", l1, ens_cache)
+    print("\n" + "=" * 72); print("LEVEL 2 ECOREGIONS"); print("=" * 72)
+    run_level("L2", l2, ens_cache)
+
+
+if __name__ == "__main__":
+    main()