This updated guide (October 2026) shows commodity traders and portfolio managers how to convert satellite and remote sensing signals into reliable, tradeable forecasts for Brazilian soy and corn exports. It assumes you already trade futures, FOB swaps or freight and want to institutionalize satellite-derived alpha — not as a curiosity but as an operational signal source. Updates since the original August 2026 version focus on new data product maturity, ML/LLM tools for time-series fusion, and operational lessons from the 2026 austral season.

Prerequisites and context

Before building or adapting a satellite system you should have:

  • Access to core market instruments (CBOT futures/options, regional FOB swaps, freight contracts) and an execution pathway for physical or OTC hedges.
  • Cloud compute and storage (AWS, GCP, Azure) with experience running geospatial workloads or a vendor-managed analytics offering. By Oct 2026, most commodity trading desks use cloud‑hosted analytics to meet latency and scaling needs.
  • Familiarity with key Brazilian sources: CONAB (crop reports), IBGE (agricultural statistics), port authorities (Santos, Paranaguá, Rio Grande) and customs/loadings datasets. Official reports remain the calibration anchor.
  • A data budget: combine free (Sentinel, MODIS, Landsat) and commercial (Planet, Maxar, ICEYE, Capella) feeds. Commercial suppliers now increasingly offer hosted analytics that reduce ingestion engineering, but licensing terms vary — read resale and sharing clauses carefully.

Why this matters now (Oct 2026)

Two short developments since August 2026 that affect traders:

  • Tooling & model maturity: During 2026, multimodal foundation models and time-series LLMs moved from experimentation to production on trading desks. These models accelerate feature engineering (e.g., mapping NDVI/SAR sequences to phenology) and allow rapid scenario queries in plain language, but they require disciplined guardrails for explainability.
  • Operational data fusion: Commercial SAR capacity (ICEYE, Capella) plus expanded hosted delivery of PlanetScope and Maxar products have shortened the time from satellite pass to analytics. Real‑time AIS + brokered port throughput feeds have become standard layers to translate crop signals into export-flow risk within hours rather than days.

Step 1 — Define your trading objective and horizon

Be explicit; successful desks tie technical signals to tradeable outcomes.

  1. Define the outcome: e.g., "forecast November–January soy exportable supply from Mato Grosso with weekly updates" or "flag inland transport congestion likely to delay Paranaguá loadings within a 4–12 week window."
  2. Map to instruments: short-term port/backlog risk → FOB swaps, basis trades, short-charter freight; seasonal production shocks → CBOT futures/options; very short-term logistics shocks → physical spot/tri-party deals.
  3. Set decision timelines and latencies: match data cadence to trade types — daily or sub‑daily for logistics; weekly for in‑season yield adjustments; monthly for strategic procurement.

Step 2 — Select datasets and providers (2026 update)

Combine free and commercial data and prefer hosted delivery where possible:

  • Free: Sentinel‑1 (SAR, 10m), Sentinel‑2 (optical, 10m), MODIS (daily composites) and Landsat archives. ESA Copernicus remains the baseline for reproducibility and audit trails.
  • Commercial and hosted offerings: PlanetScope (3–5m) and Maxar remain the go‑to for field‑level resolution; ICEYE and Capella expanded SAR constellation capacity in 2025–26, improving revisit and timeliness for wet‑season monitoring.
  • Derived and weather products: ERA5/ERA6 reanalysis, ECMWF short‑range forecasts for evapotranspiration, and SMAP/SMOS (soil moisture proxies). In 2026, several vendors began offering pre‑fused soil‑moisture layers optimized for agronomy.
  • Logistics layers: AIS vessel tracking (global and Brazilian harbor feeds), port storage/berth data, ANDA (road) and river gauge hydrology feeds. Brokers and freight indices (e.g., BSI, regional freight desks) provide market context for translation to rate moves.
  • Ground truth: CONAB, IBGE, producer cooperatives and paid field surveys. Use multiple anchors to avoid single‑source bias.

Step 3 — Build a processing pipeline (practical changes)

Operational systems in 2026 emphasize reproducible data ops and explainability.

  1. Ingest: prioritize hosted delivery (commercial providers delivering processed tiles or time-series through APIs) to reduce engineering burden. For Sentinel data, use Google Earth Engine or Sentinel Hub for pipeline orchestration.
  2. Preprocess: standardize reprojection, apply optical cloud masking, and use radiometric and speckle filtering for SAR. In 2026, automated QA pipelines that tag data lineage per record became commonplace — embed those tags early.
  3. Mosaic & aggregate: create multi-resolution composites (7-, 14-, 30-day) and aggregate to your decision unit (farm polygons, municipality, port catchment). Keep both high-frequency raw layers and smoothed composites for model ensembles.
  4. Store & serve: use a time-series friendly DB (PostGIS with time-series extensions, or specialized TSDBs) and an object store with Delta Lake or Iceberg for reproducible snapshots. Version your datasets so audits can tie trades to the exact data used.

Step 4 — Derive indicators and models (new best practices)

Use an ensemble approach combining interpretable rules and ML models; add LLM-powered diagnostics for human-in-the-loop interpretation.

  • Core indicators: NDVI/EVI time-series, greenup date, peak NDVI, senescence date; VV/VH and VV/VV backscatter ratios from SAR for canopy structure and inundation detection.
  • Advanced features: temporal derivatives (growth rate), phenological stage probabilities from a fitted Hidden Markov Model, soil moisture anomaly indices from fused SAR+SMAP inputs.
  • Modeling: maintain two parallel tracks — (a) interpretable statistical regressions calibrated regionally to CONAB/IBGE yields, and (b) ML ensembles (gradient boosting, temporal CNNs, or transformer-based time-series models) that ingest multi-source sequences. In 2026, time-series LLMs accelerated rapid scenario querying and feature importance extraction, but always validate outputs against interpretable baselines.
  • Explainability: produce per-decision feature attributions (SHAP or integrated gradients) and a natural-language summary for trader consumption (e.g., "SAR backscatter rose 18% in X muni; NDVI lagging 15 days — 60% probability of delayed harvest").

Step 5 — Calibrate with ground truth and reports

Calibration remains essential. Add regionally stratified bias correction and uncertainty quantification.

  • Historical match: validate models on a rolling 5–10 season window and report error metrics by state and soil type. CONAB and IBGE remain critical for official comparators.
  • Real‑time anchors: use port loadings, field surveys, and aggregator reports to update posterior distributions during the season. Incorporate short surveys with cooperatives as rapid validation when anomalies appear.
  • Uncertainty bands: present probabilistic outcomes (P10/P50/P90) and calibrate model spread to observed forecast errors so traders can scale positions by confidence.

Step 6 — Translate crop signals into export and price implications

Mapping production deviations to exportable supply and prices requires explicit flow logic.

  1. Estimate production shock: convert yield delta (t/ha) × sown area → gross production change at municipality/state level.
  2. Adjust to exportable supply: model crush rates (soy → meal/oil), domestic consumption trends, carryover stocks and likely government interventions. Use a simple flow model that can be stress-tested under alternative crush and domestic demand scenarios.
  3. Map to export windows & ports: assign production to likely shipping windows and ports based on historical origin-to-port allocations (Mato Grosso → Santos/Paranaguá via BR‑163/BR‑364; southern states → Rio Grande/Brasil). Timing differences change which futures months and basis contracts are affected.
  4. Price translation: use a calibrated supply-shock → price-impact model (historical regressions or structural simulation). Report a range of price deltas tied to elasticities and logistic constraints rather than a single point estimate.

Step 7 — Execution strategy and instrument selection

Match instrument choice to signal nature and operational constraints.

  • Prompt logistics shocks: use short-dated FOB swaps, prompt futures, and freight forward charters. Coordinate with physical desks to capture basis opportunities.
  • Seasonal production signals: use calendar futures and options to express directional views with defined risk. Options are useful when model uncertainty is wide but directionality exists.
  • When liquidity is limited: prefer smaller futures sizes complemented by options or correlated hedges (e.g., spread trades) to manage execution risk.
  • Operational execution: implement tranche-based scaling tied to real-time confidence bands — increase size as more independent signals confirm (SAR, port AIS, field reports).

Step 8 — Risk controls, sizing and governance

Operational risk rules are as important as model quality.

  • Confidence gating: require ensemble consensus or minimum model confidence before scaling above a pre-defined risk fraction of the desk’s limit.
  • Correlated exposures: actively hedge macro and cross-commodity risks (soy vs. corn correlations, crush margin moves, BRL currency exposure).
  • Liquidity and slippage modeling: include real-world execution costs, slippage, and margin requirements in position-sizing rules.
  • Governance: log datasets, model versions and human approvals. In 2026, internal audit teams expect dataset lineage and explicit trade rationale for every material position opened on satellite signals.

Step 9 — Backtest for economic value

Measure P&L impact, not just predictive metrics.

  • Walk‑forward testing: re-train on rolling windows and simulate live deployment; include seasonality and ENSO-like event scenarios.
  • Transaction cost model: include brokers’ fees, slippage, and margin finance costs. Simulate liquidity shocks (e.g., constrained FOB market) to see worst-case drawdowns.
  • Performance metrics: measure Sharpe, information ratio versus a benchmark (long CBOT soy futures), hit rate and realized drawdowns. Also track alpha persistence across seasons.
  • Operational stress tests: run multi-muni underperformance and port-closure scenarios to understand tail risk and hedge adequacy.

Step 10 — Alerts, monitoring and operational rollout

Turn signals into actionable workflows and maintain human oversight.

  • Automated alerts: trigger when composite indicators breach thresholds (e.g., NDVI percentile fall + SAR flood signature + AIS backlog increase).
  • Human‑in‑the‑loop: require trader or risk manager sign-off for material trades and provide a one-page decision pack (imagery, time-series, confidence metrics, trade rationale).
  • Audit trail: store the alert payload, model version, and decision outcome for compliance and post‑trade review.
  • Cost monitoring: track data subscription ROI; use free Copernicus as baseline and commercial tasking selectively for confirmation.

Worked (updated hypothetical) example — Mid‑Oct 2026 operational chain

Scenario: A 10‑day cluster of high SAR backscatter and delayed greenup appears across a belt of three Mato Grosso municipalities during late vegetative stage.

  1. Signal detected: 14‑day composite shows NDVI at the 12th percentile and SAR indicates surface saturation in low‑lying fields.
  2. Confirmations: AIS shows rising vessel dwell times at Paranaguá and a local cooperative report notes harvest machinery delays.
  3. Model output: ensemble estimates a 3–5% yield reduction for affected muni cluster → 0.8–1.2Mt reduction in exportable soy for the November–January window (probabilistic P10/P50/P90 bands reported).
  4. Execution: A desk could (a) buy a modest size of prompt CBOT futures scaled to risk budget, (b) overlay a small call option for convexity, and (c) negotiate a short-dated FOB swap if counterparty access exists. Exit rules: scale down if NDVI recovers to seasonal median or if port AIS clears within two weeks.

Common mistakes and how to avoid them

  • Overfitting to historical weather regimes: use rolling windows and stress tests for ENSO-like events. Keep simpler interpretable models as sanity checks.
  • Equating crop anomaly to immediate exports: always layer logistics and loading capacity into your flow model before trading exportable-supply shocks.
  • Ignoring data licensing: some commercial feeds restrict redistribution; ensure trade-station dashboards and audit logs comply with vendor terms.
  • Blind faith in "AI outputs": LLMs and fusion models speed analysis but can hallucinate causal links — require quantitative attribution before executing.

Pro tips

  • Use regional bias correction: municipality-level bias terms materially improve yield forecasts versus a single national model.
  • Triangulate with freight desks: small inland delays can amplify into large basis moves if freight supply is tight — keep freight and port desks integrated into execution plans.
  • Task selectively: use commercial high-res tasking to confirm anomalies flagged in free data rather than as ongoing coverage to control costs.
  • Version everything: tag data, model, and decision versions to enable post-trade attribution and regulatory compliance.

FAQ

How accurate are satellite-derived yield estimates compared with CONAB or IBGE?

Satellite estimates are increasingly accurate for relative deviations (e.g., identifying municipalities under- or over-performing the seasonal norm). Absolute yield accuracy depends on local calibration; well‑calibrated systems typically achieve a lower RMSE for inter-annual deviations at state level than at field level. Use official CONAB/IBGE figures for annual benchmarking and maintain municipality‑level bias corrections to improve absolute estimates.

Can I trade solely on satellite signals?

No — satellite signals are a valuable input but should be one of multiple confirmation layers. Best practice is to require at least one independent confirmation (port AIS, field survey, aggregator report) for material positions and gate position sizing to model confidence bands.

Which data combination gives the best signal-to-noise for Brazil soy/corn?

A fused approach typically performs best: Sentinel‑optical for phenology, Sentinel‑1/ICEYE/Capella SAR for all‑weather structure and inundation, and weather/soil moisture reanalysis for stress context. Add AIS and port throughput for export-flow translation. The marginal benefit of high-res optical (Planet/Maxar) is highest when you need field-level confirmation.

How should I manage data cost versus alpha?

Start with free Copernicus baselines (Sentinel‑1/2) and implement cheap, targeted commercial tasking only for high‑value confirmation. Monitor alpha per data product by backtesting with and without the commercial layer; scale spend where incremental P&L justifies the subscription.

What are regulatory or compliance risks when using satellite data?

Satellite imagery is generally public or commercially licensable, but vendors’ licensing terms may restrict redistribution or sharing with counterparties. Maintain dataset licenses in your compliance repository, log data lineage for trade rationale, and ensure traders’ use aligns with internal policies on confidential information and market manipulation.

Final advice (Oct 2026)

Satellite-driven forecasting is now a mature operational tool on many desks. The edge comes from rigorous calibration to local conditions, disciplined uncertainty quantification, and tight integration with logistics and execution teams. Use multimodal ensembles and LLM-assisted diagnostics to accelerate interpretation, but enforce strong governance and human sign-offs for material positions. If you adopt these practices, satellite foresight can reliably turn days of advance warning into economic value in Brazil’s soy and corn complex.