Sentinel-1/-2 ML Model for Vegetation Cover

Project overview We are building an operational satellite-based agricultural compliance service for a national government agency in Europe. Pilot launches autumn 2026. We need an experienced Earth-observation ML practitioner to deliver the model component as a trained, runnable artefact we deploy on our side. You deliver the model and its Sentinel ingestion; we own the production infrastructure, the regulatory rule layer, and the customer-facing platform. What we need A trained model that: - Takes a field-parcel polygon and an evaluation date as input - Returns estimated fractional ground cover (%) of the catch crop, calibrated to a visual percentage scale - Returns calibrated uncertainty alongside the point estimate - Optional / nice-to-have: above-ground biomass estimate Consumes Sentinel-1 + Sentinel-2 time series plus minimal static parcel context (main crop, declared sowing date). Architecture, feature engineering, training framework, and Sentinel ingestion are your call. What we provide - Labelled training data: several thousand georeferenced field inspections across multiple recent seasons. Each is a polygon with a graded interval-valued label (cover %), naturally suited to ordinal / interval-censored regression - Static parcel metadata for the full declared population — polygons, main crop, declared sowing date - Optional: Google Cloud compute credits for training We do not provide pre-processed Sentinel features. You source, ingest, and preprocess S1 + S2 yourself. Required experience - Concrete prior work on Sentinel-1/-2 time-series modelling for vegetation cover, crop monitoring, or biomass. Share examples — papers, code, production projects - Own working pipeline for S1/S2 ingestion (CDSE, AWS Open Data, Sentinel Hub, GEE — any stack is fine, just production-ready) - Ordinal regression / interval-censored modelling experience, or a clear case for an alternative Production-minded: clean inference path, dependency management, documentation - Experience with Sentinel-2 cloud-cover gaps in autumn at high latitudes Academic credentials welcome but not required; we evaluate on delivered, working ML. Deliverables — a private GitHub repo containing: - Trained model artefacts (weights, scalers, transforms) - Clean Python inference code (polygon + date → predictions + uncertainty) - Sentinel ingestion / preprocessing code in the inference path, runnable on our infrastructure - Documentation for production-scale inference (tens of thousands of parcels per weekly batch on Google Cloud) - Brief technical write-up — approach, validation results, known limitations Intellectual property Trained model artefacts and any code written specifically for this engagement become our property on delivery, with full rights to modify, update, and redistribute. You retain ownership of general-purpose methodology or framework code you bring from prior work, under any licence that permits our production use and downstream modification. Timeline (please confirm you can hit it) Project start: early June 2026 on contract signing First trained checkpoint (validation review): mid-July 2026 Final production-ready delivery: early August 2026 Hand-over and integration support: through end of August 2026 Budget and milestones Fixed-price. Scope your bid against the deliverables, not hourly. Higher bids welcome if backed by prior work that shortens the path — pre-trained components, validated pipelines, your own Sentinel infrastructure. Three payment milestones: - First trained checkpoint accepted - Final production-ready delivery - Successful integration on our side How to apply Please include: -Under 300 words on relevant prior work — links to papers, code, production projects. Generic "I do remote sensing" pitches will be skipped -Your high-level proposed approach in 1–2 paragraphs (include how you'll handle Sentinel ingestion) -Your fixed-price bid -Timeline against the milestones above -Any clarifying questions We will shortlist 2–3 candidates for a 30-minute call before awarding. This is real, scoped, production work — not research. We're looking for a practitioner who has done this before, has their own Sentinel stack, and can deliver a model we can deploy with confidence.