System inputs
The inference layer accepts soil-moisture readings, crop and growth-stage context, weather, irrigation-system details, operating constraints, and location inputs. These inputs form the system state behind each recommendation.
Forecasting layer
Helios uses XGBoost multi-output regression to forecast volumetric soil moisture at 24, 48, and 72 hours. The model works across recent moisture levels and deltas, weather and ET signals, irrigation constraints, crop context, and seasonal features.
Runtime enrichment
OpenET monthly ET enrichment adds crop-water-demand context when available. NOAA weather backfill fills missing weather inputs. Runtime results are cached by location and month, with a baked-in fallback path that keeps inference available when external enrichment is unavailable.
Decision layer
A rule-based planning layer converts the forecast into irrigation timing and depth recommendations. It respects pump capacity, budget, infiltration, and water-window constraints instead of treating irrigation as an unconstrained prediction problem.
Product architecture
The product combines a static frontend with a FastAPI backend. The API exposes prediction and nearby-feedback endpoints, while the feedback loop can surface comparable local operating context for future recommendations.