earth-climate

name	earth-climate
description	Use for earth, ocean, atmosphere and ecology data work: NetCDF/CF, GRIB, Zarr, HDF5, GeoTIFF and vector formats, coordinate reference systems, regridding and land/sea masks, non-standard calendars, anomalies and baselines, trends with autocorrelation, extremes and return periods, CMIP scenarios and ensemble spread, remote-sensing preprocessing, and field-ecology survey data.
version	1.0.0
author	Remedy
tags	["research","climate","geospatial","remote-sensing","ecology","netcdf"]
requires	[]
tools	["analysis_env","analysis_run","analysis_ledger","data_profile","data_diff","stats_assumptions","stats_effect_size","lit_search","cite_add","skill_activate","file_read","file_write"]
triggers	["\\b(NetCDF|\\.nc4?\\b|\\bGRIB2?\\b|CMIP[56]|ERA5|MODIS|Landsat|Sentinel-[12])\\b","\\b(reanalysis (?:data|product|dataset)|climate (?:model|projection|scenario)|RCP\\d(?:\\.\\d)?|SSP\\d-?\\d)\\b","\\b(detrend\\w+|seasonal (?:cycle|decomposition)|anomaly time series|teleconnection|\\bENSO\\b|\\bNAO\\b index)\\b","\\b(shapefile|geospatial (?:raster|vector)|EPSG:\\d+|remote sensing|zonal statistics|reprojection)\\b"]

Earth, climate and ecology data

Run skill_activate(skill="research-method") first and work from that spine — question framing, evidence standards, citation honesty, what verification means. Do not restate it. This pack covers what is different when the data is gridded, projected, dated on a strange calendar and produced by somebody else.

Before any arithmetic

1. analysis_env(path). These stacks live in the project, not in this process: xarray, netCDF4, cfgrib, rasterio, GDAL, cdo, nco, geopandas, R terra/sf. Everything heavy runs through analysis_run in the project's own interpreter. Missing tool -> say so and name the install; do not silently substitute a worse method.
2. Record the product identity. Name, version/DOI, spatial and temporal resolution, variable long_name and units, download date, and processing level (L1/L2/L3). cite_add the dataset by its own DOI — datasets are citable objects, and "ERA5" alone is not a citation.
3. Print the metadata before you trust the array. ncdump -h, gdalinfo, xr.open_dataset(...).attrs. Read units, _FillValue/missing_value, scale_factor/add_offset, calendar, the CRS, and whether latitude runs north-to-south. references/data-formats-and-cf.md.
4. Say which kind of data it is and never mix silently. Observation, reanalysis (a model constrained by observations — not an observation), and free-running model output are three different epistemic objects. A figure that overlays them must label each one.

Working rules

• CRS. Every dataset carries or lacks a CRS; assuming EPSG:4326 because the numbers look like degrees is the classic silent error. Reproject deliberately, compute areas and distances in an equal-area or local projection, never in degrees. references/crs-and-projections.md.
• Area weighting. A global or regional mean of a lat/lon grid without cos(latitude) weighting is wrong, and wrong in a direction that flatters polar amplification. Use the grid cell areas the tool gives you.
• Regridding. Bilinear for smooth fields you will look at; conservative for anything that must integrate correctly (precipitation, fluxes, area totals). Regrid once, at a documented step, and keep the source grid. references/regridding-and-masks.md.
• Land/sea masks differ between products; apply the mask that belongs to the grid you are on. Model calendars can be 360-day or no-leap: never compare a 360-day series to a Gregorian one by index.