TerraFlow Roadmap¶

This roadmap tracks completed phases and current state. v0.5.0 (2026-06-30) refocused the surface on the climate-impact flagship: removed GeoAI + H3 wrappers, narrowed validation to spatial-block CV, and shipped scenario × hazard fan-out with CMIP6 NetCDF support. The next release (v0.6.0) makes CMIP6 ingestion first-class in the pipeline (climate.cmip6_scenarios: config block, unit conversion, NetCDF SHA-256 in manifest.json); package restructure into terraflow.io.* + terraflow.climate.* sub-packages slips to v0.7.0 (issue #148).

JOSS pre-review submitted 2026-06-08, rejected 2026-06-24 on impact criteria. Resubmission gated on adoption signal, not code (see internal strategy repo).

Phase Status¶

Phase	Plans	Status	Completed
1. Foundation Hardening	delivered	Released	2026-04-23 (v0.3.0)
2. Sensitivity Analysis	3/3	Released	2026-04-23 (v0.3.0)
3. Model Validation	narrowed to spatial-block CV	Released	2026-06-30 (v0.5.0)
4. H3 Export	cut	Removed	2026-06-30 (v0.5.0; PR #135)
5. GeoAI Engine	cut	Removed	2026-06-30 (v0.5.0; PR #134)
6. Climate-impact assessment	shipped	Released	2026-06-30 (v0.5.0; issue #138)
7. JOSS Resubmission	pending adoption signal	Blocked	—

Phase 1 — Foundation Hardening¶

Close the CRS error and kriging diagnostic gaps that block JOSS reviewer acceptance.

Delivered (v0.2.2 / v0.3.0):

CRSMismatchError raised with both CRS strings when raster and ROI CRS disagree.
kriging_diagnostics block in report.json — model name, psill, nugget, sill, range, range units.
kriging_loocv RMSE in report.json per kriged climate variable.
plotly demoted to optional [viz] extra; trove classifiers + Documentation URL added to pyproject.toml.
interpolation_fallback block in report.json with per-variable fallback-to-mean counts plus aggregate total (v0.3.0).
Coverage floor maintained at 85 % branch coverage with kriging-fallback and Monte-Carlo edge-case tests.

Phase 2 — Sensitivity Analysis ✓¶

Sobol' first-order / total-order indices and Morris elementary effects for ModelParams weights via SALib. Invoked with terraflow sensitivity -c config.yml. Results written to sensitivity_report.json.

Phase 3 — Model Validation (narrowed in v0.5.0) ✓¶

Spatial-block cross-validation (Roberts et al. 2017) via terraflow validate -c config.yml; results appended to report.json under the validation key. Cohen's κ and Moran's I wrappers were removed in v0.5.0 (PR #142 / issue #136) — call sklearn.metrics.cohen_kappa_score and esda.Moran directly on features.parquet. See Migration v0.4 → v0.5.

Phase 4 — H3 Export (cut) ✗¶

Removed in v0.5.0 (PR #135 / issue #135). The module was a thin h3-py + pandas wrapper; downstream users now call h3-py directly on features.parquet (five-line recipe in the migration guide).

Phase 5 — GeoAI Engine (cut) ✗¶

Removed in v0.5.0 (PR #134 / issue #134). Engine bodies were NotImplementedError stubs and downstream citations went to geoai-py, not TerraFlow. A separate terraflow-geoai package is parked post-JOSS pending demand signal.

Phase 6 — Climate-impact assessment ✓¶

Shipped in v0.5.0 (issue #138, PRs #144–#150). Multi-scenario climate driver (historical + CMIP6 SSP windows) × seven WMO/ETCCDI-aligned hazard indicators (annual mean, seasonal mean, growing-degree days, frost days, heat-stress days, precipitation percentiles, simplified Thornthwaite SPEI). Writes a sibling climate_features.parquet alongside features.parquet; the historical artifact contract is unchanged. See examples/demo_config_climate_impact.yml and notebook docs/notebooks/07_climate_impact_crop_suitability.ipynb.

Phase 7 — JOSS Resubmission (blocked on adoption)¶

JOSS pre-review (#10686) rejected 2026-06-24 with the note: "If/when in the future this software is used more widely, please resubmit." The fix is calendar-time adoption work (university outreach, AGU Fall poster, blog posts, CGIAR / AgMIP contact), not code. Resubmission gate criteria and outreach planning live in the internal strategy repo (see project_post_coding_adoption_plan memory).

v2 Ideas¶

These are not planned for v1 (not on the JOSS critical path) but are worth tracking:

Directional variograms for spatial anisotropy detection (scikit-gstat).
Universal Kriging with elevation covariates.
UTM-projected variograms for high-latitude fits.
STAC / COG ingestion for cloud-native rasters.
Polygon ROI support (shapefile / GeoJSON).
Data-driven weight learning from labeled training points.

Out of Scope¶

Feature	Reason
Web UI or dashboard	Library-first; visualization is the caller's responsibility.
Streaming / real-time ingestion	Batch pipeline; streaming adds complexity without research value.
Commercial ag operations tooling	Focus is the research community.
Non-Python client SDKs	Python-first for v1.