Platform Notes

    Known Limitations

    Every data source has constraints. Understanding these helps you build robust integrations and set accurate expectations with your users.

    Cloud Cover Gaps

    Optical satellites (Sentinel-2, Landsat) cannot image through clouds. Persistently cloudy regions may have weeks-long data gaps. The API returns the best available cloud-free composite, but some dates may have no usable imagery.

    Impact: Tropical and monsoon regions are most affected during rainy seasons.

    5-Day Revisit Cadence

    Sentinel-2 revisits the same location approximately every 5 days (at the equator). Events that occur and resolve between revisits — such as flash floods or short-duration fires — may not appear in the data.

    Impact: Sub-daily or intra-day events are not captured.

    10–30 m Spatial Resolution

    Sentinel-2 provides 10 m/pixel for visible bands and 20 m for red-edge/SWIR. Landsat provides 30 m. Objects smaller than one pixel (individual trees, small structures) cannot be resolved.

    Impact: Fine-grained features like individual crop rows or small buildings are not distinguishable.

    NDVI Measures Greenness, Not Causation

    Normalized Difference Vegetation Index measures chlorophyll reflectance. A drop in NDVI indicates less green vegetation but does not identify the cause — it could be harvest, drought, disease, fire, or seasonal dormancy.

    Impact: Automated anomaly alerts should be verified with on-ground or contextual data.

    Vegetation Classification Edge Cases

    The ESA WorldCover dataset provides 10 m land-cover labels. At biome boundaries (e.g., forest-grassland transitions), the dominant class may not represent the full area. Fallback NDVI thresholds are used if WorldCover data is unavailable for a location.

    Impact: Mixed land-use areas may be labeled with the single most-common class.

    Anomaly Detection in Arid Regions

    Anomaly detection uses NDVI z-scores to compare a current period against a baseline. In arid or semi-arid environments with low NDVI variability, small real changes produce small z-scores that may fall below detection thresholds.

    Impact: Subtle or gradual degradation may not trigger an anomaly alert. Consider using 'high' sensitivity for arid areas.

    Trend Projection vs. Machine Learning Forecasting

    The temporal forecast uses damped linear extrapolation with a seasonal overlay derived from historical observations. This is a statistical projection, not a trained ML model. Accuracy degrades for forecast windows beyond 30–60 days.

    Impact: Forecasts are best used as directional indicators, not precise predictions.

    Risk Scores Are Indicative, Not Actuarial

    Risk scoring combines elevation, flood history, wildfire fuel load, and precipitation data into a composite score. These scores are useful for screening and prioritization but are not a substitute for site-specific actuarial or engineering assessments.

    Impact: Do not use risk scores alone for insurance underwriting or regulatory compliance.

    Confidence Scores Are Relative, Not Absolute

    Confidence values (0–1) returned alongside detections reflect the model's internal certainty relative to available data quality and cloud-free pixel count. A confidence of 0.85 does not mean "85% probability of being correct" in a statistical sense — it indicates strong internal agreement among spectral bands and temporal composites. Scores below 0.5 warrant manual verification.

    Impact: Treat confidence as a ranking signal to prioritize review, not as a calibrated probability.

    Decision-Support Tool, Not Ground Truth

    All Satalyse outputs — change maps, anomaly alerts, risk scores, and vegetation indices — are derived from remote-sensing models and should be treated as decision-support inputs. They are not a substitute for ground-truthing, field surveys, or professional assessments. Outputs may contain false positives or miss events that fall below detection thresholds.

    Impact: Always validate critical decisions with on-site data or domain-expert review before taking action.

    Have questions about data quality?

    Our team can help you evaluate whether Satalyse fits your use-case.