An honest breakdown of what's scientifically validated, what's a useful proxy, and when a number needs field verification before you rely on it.
Water extent is detected using NDWI thresholding at Sentinel-2's native 10m resolution — meaning the smallest reliably-detected water feature is roughly a 10m × 10m patch.
| Factor | Effect on accuracy |
|---|---|
| Shoreline / edge pixels | A pixel that's 50% water, 50% land gets classified as one or the other, not both — this is the main source of boundary error. |
| Typical accuracy vs. surveyed boundaries | ±2–5% for a well-defined open lake; worse for narrow rivers or vegetated shorelines. |
| Reeds / emergent vegetation | Can be misclassified as land even when growing in shallow water — undercounts true wet area. |
| Cloud / haze on the selected date | Can distort the apparent boundary if not fully masked by the cloud filter. |
These indices have decades of peer-reviewed field validation behind them. Relative differences (higher/lower, more/less) are trustworthy at face value.
| Index | Confidence | Why |
|---|---|---|
| NDVI — vegetation health | HIGH | One of the most validated indices in all of remote sensing. |
| NDWI / MNDWI — water extent | HIGH | Very reliable for open water detection under normal conditions. |
| Nighttime lights (VIIRS) | HIGH | A direct physical radiance measurement — real data, not a derived proxy. |
| NDBI — built-up index | MEDIUM-HIGH | Well-established, but can confuse bare soil with built-up surfaces. |
These indices correlate strongly with the physical property they name, but converting them to a certified lab-grade number (mg/L, NTU, µg/L) requires a calibration sample from that specific water body. Treat them as a strong signal for where and when to investigate further, not as a lab result.
| Index | Confidence | What it actually is |
|---|---|---|
| NDTI — turbidity | PROXY | A spectral ratio that correlates with turbidity — not the same as an NTU reading. |
| NDCI — chlorophyll | PROXY | Indicates relative bloom intensity, not a lab-grade µg/L concentration. |
| TSS — suspended solids | PROXY | A band-ratio approximation; real TSS in mg/L needs local calibration. |
| Secchi depth | PROXY | The least reliable of the water indices — a statistical estimate, not a physical disk measurement. Can be meaningfully off in unusual water chemistry. |
| Geology indices (iron, clay, carbonate, evaporite) | PROXY | Spectral signatures suggesting mineral presence — good for narrowing survey targets, not a substitute for lab mineral analysis. |
No amount of image processing can see detail smaller than the sensor's native pixel size. This sets the ceiling on precision regardless of which index you choose.
| Sensor | Resolution | Revisit | Best for |
|---|---|---|---|
| Sentinel-2 | 10m / pixel | ~5 days | Water quality, vegetation, shoreline detail |
| Landsat 9 | 30m / pixel | ~16 days | Geology, large-area built-up trends, long historical baseline |
| VIIRS DNB | 500m / pixel | Nightly | City-scale nighttime lights, not individual buildings |
Comparing the same location across two dates. Spotting which of several regions is worse/better. Tracking a multi-year trend direction. Screening many candidate sites to prioritise where to actually send a team.
Regulatory compliance reporting. Legal or insurance evidence requiring certified values. Any decision with safety consequences (drinking water, livestock, swimming advisories). Converting a proxy index to a specific mg/L, NTU, or µg/L figure.