Point anywhere on Vietnam's map. Culvera reads your satellite vegetation data, soil profile, and 14-day weather forecast — then delivers your top 3 crops with a full action plan.
Free to use · No API key required · 3 seasons
Generalized advisories that don't account for this season, this location, or this soil — published months in advance and applied nation-wide.
A wrong crop choice in the Mekong Delta can mean losing 30–50% of potential revenue. With limited capital, one bad season has lasting consequences.
National statistics hide the variance between provinces 40 km apart. Soil pH, salinity intrusion, and rainfall patterns differ dramatically across Vietnam's four agroecological zones.
Culvera changes this.
Every recommendation is built live from three public scientific data pipelines — no canned responses, no stale data.
Click any point on Vietnam's interactive map. Culvera snaps to the nearest of the 13 supported provinces and confirms your location instantly.
The system reads today's date and your region. Rice zones get Đông Xuân, Hè Thu, or Mùa. The Central Highlands perennial belt uses an annual cycle.
Three data pipelines fire in parallel: Open-Meteo 14-day weather, SoilGrids 2.0 soil properties, and Sentinel-2 NDVI vegetation index. 28 features assembled in seconds.
XGBoost ranks your top 3 crops with confidence scores. Each comes with yield forecast, 6-month price outlook, NPK fertiliser schedule, and harvest timing window.
No black box. Each output is explainable, sourced, and calibrated to your specific location and season.
XGBoost classifier ranks all supported crops for your location and season. Confidence percentage shown for each; results below 35% trigger a visible caution flag.
A separate XGBoost regressor predicts expected yield in tonnes per hectare. Validated at 0.43 t/ha RMSE against GSO historical province-level data.
Forward price model using GSO farmgate baselines × empirically-grounded seasonal multipliers (2018–2024 average patterns). Trend direction: rising, stable, or falling.
MARD-aligned application rates adjusted for your soil pH and organic carbon. Split into phased timings — basal, week 3, week 6 top-dressing — with product examples.
SoilGrids 2.0 data scored across pH, organic carbon, nitrogen, and texture class. Plain-language fix suggestions — e.g. lime application rates for acid-sulfate soils.
Enter your farm size in hectares. Culvera allocates area across the top crops by confidence-weighted share, computes total expected revenue, and applies a weather-hedge buffer when high-risk conditions are detected.
Culvera integrates openly-licensed scientific data — no proprietary feeds, no vendor lock-in, no API keys required.
14-day hourly weather per location — temperature, precipitation, solar radiation, humidity. Cached 6 hours to minimize latency. Falls back to NASA POWER for historical growing-season aggregates.
14-day horizon · 6h cache · No key requiredNDVI vegetation index computed from Sentinel-2 L2A imagery via Microsoft's Planetary Computer STAC API — anonymous access, no registration. Captures crop greenness, peak growth timing, and growth stage.
~10m resolution · 5–14 day revisit · Public STACSoil pH, organic carbon, total nitrogen, texture class, bulk density, and CEC at 250m resolution from ISRIC World Soil Information. Annual vintage, no registration or API key needed.
250m resolution · Annual vintage · Open REST API
All data sources are public, free, and actively maintained by global scientific institutions.
No proprietary data. No vendor lock-in. Culvera's entire data stack can be self-hosted.
Covering Vietnam's major commodity crops across four agroecological regions — from Mekong rice paddies to Central Highlands coffee and pepper.
| Dataset: | 270 rows × 22 columns |
| Training set: | 263 rows × 28 features |
| Classifier accuracy: | 0.812 |
| Regressor RMSE: | 0.431 t/ha |
| Bootstrap time: | ~12.3 min (Sentinel-2 bottleneck) |
Seven design decisions behind every recommendation
MVP scope acknowledged. The training set is small (~270 observations, 13 provinces). Classifier accuracy of 0.6–0.9 is normal for this data size. The goal is a working, honest end-to-end pipeline — not a black box with inflated claims. Confidence scores are shown to users at every step.