GrapeExpectations
Precision-viticulture pipeline modeling vine canopy health across Washington State wine country at 10 m resolution — nine years of Sentinel-2 imagery, terrain, soil, and climate fused into a gradient-boosted model, then pushed through three different tools to see what each one is actually built for.
1. The pipeline — Python + Google Earth Engine
Vineyard blocks digitized from KML are retiled into 32,978 equal-area hexagonal cells (~10 m across, one Sentinel-2 pixel each). Each cell gets a terrain profile from the USGS 1 m DEM (elevation, slope, aspect, curvature), soil properties from USGS gSSURGO, daily weather from gridMET, and nine years of Sentinel-2 NDVI pulled through Google Earth Engine — then a stacked gradient-boosted model (PySpark MLlib) predicts each cell's harvest-window NDVI anomaly from terrain, soil, and climate alone.
Real problems the GEE pull actually hit
FeatureCollection blew past GEE's 10 MB
request cap. Fixed by uploading the tile centroids as a GEE asset
once, then referencing that asset ID in every export task instead
of re-sending geometry each time.
filterBounds() against the uploaded asset returned
zero images — GEE didn't recognize a spatial extent on an asset
built from bare points. Fixed by computing an explicit
ee.Geometry.BBox from the centroid CSV's own
lon/lat min/max instead of trusting the asset's implicit bounds.
reduceRegion() returned null.
sampleRegions() against the polygon set, not a
buffered point, was the fix.
Full run order and every notebook: RegressionRidge/spark_pipeline/.
2. GIS branch — QGIS
The same 32,978-cell grid that crashed a browser BI tool's map
chart (see below) is exactly what a desktop GIS tool is built for.
export_qgis.py joins the real hex polygon geometry to
terrain and a 9-year NDVI trend and writes one GeoPackage —
no aggregation, no row-count wall.
The authored QGIS map — graduated symbology on the 9-year trend, styled and screenshotted from an actual QGIS session — isn't built yet. The export above is real and ready to open; this callout gets replaced with the real screenshot once that session happens.
What the full-resolution surface looks like
These are the matplotlib fallback — not QGIS output, but the same underlying grid rendered at full density, which is the point: neither of these figures aggregates or samples down from 32,978 cells.
3. BI branch — Looker Studio
Power BI Service rejected personal-email signup outright, and its "free" workaround wanted a phone number and a subscription plan. Looker Studio works with a plain Gmail account — but its map chart draws one marker per row client-side and crashed the browser tab past a few thousand points, which this grid clears by a wide margin. The star-schema CSVs below are the real fix: same data, same tool, joined by hand in Looker Studio's Blend editor instead of trying to force 32,978 markers onto one map.
dim_cells.csv dim_years.csv fact_ndvi_yearly.csv Loading + blend walkthroughThe published Looker Studio report isn't live yet. The CSVs and walkthrough above are the real, working export — the report itself slots in here once it's built.