Lake shoreline recession and drought-driven surface-area decline
Multi-decadal Landsat records combined with Sentinel-2 imagery and satellite altimetry let analysts separate genuine long-term lake decline from seasonal noise, quantifying shoreline retreat and surface-area loss with documented accuracy.
Sensors
- Landsat 4-9 (Collection 2): 30 m multispectral resolution, 16-day repeat at the equator (8-day effective with Landsat 8 and 9 operating together). Archive from 1982 onward enables four-decade trend analysis. SWIR and green bands support MNDWI and DSWE water indices.
- Sentinel-2 MSI: 10 m resolution in visible and near-infrared bands, 20 m in SWIR. Five-day revisit at mid-latitudes with both Sentinel-2A and 2B active. Resolves smaller lakes and captures finer shoreline geometry than Landsat, but archive begins only in 2015.
- Sentinel-3 SRAL (radar altimeter): Along-track water-surface elevation measurements to roughly 3-4 cm precision over lakes wider than approximately 1 km. 27-day repeat. Corroborates area-change signals with independent volume proxies. Covered by the Copernicus Hydrology Service.
- ICESat-2 ATL13: Photon-counting lidar delivering inland water-surface elevations at approximately 13 m along-track footprint spacing, with stated precision of a few centimetres on calm water. Repeat cycle of 91 days limits temporal density, but the product resolves smaller water bodies than radar altimetry.
What the water index actually measures, and what it does not
The Modified Normalised Difference Water Index (MNDWI) uses the ratio of green reflectance to shortwave infrared reflectance. Water absorbs strongly in the SWIR; most land surfaces do not. A pixel with MNDWI above a chosen threshold is classified as open water. The Dynamic Surface Water Extent (DSWE) product, distributed by the USGS as part of Landsat Collection 2, extends this by combining multiple spectral tests and assigning confidence levels rather than a binary mask.
Both indices measure surface reflectance, not water depth. A lake that has lost half its volume but spread into a shallow basin may show only modest area decline. Conversely, a lake retreating into a steep-sided caldera can lose volume rapidly while the satellite records relatively little area change. This ambiguity is why corroborating altimetry matters. Area alone is an incomplete proxy for water storage; the combination of area and elevation change gives a far more defensible estimate of volume loss.
Separating seasonal swing from secular decline
Most lakes breathe with the seasons. A single-date comparison between a wet-season image and a dry-season image can fabricate an apparent crisis or conceal a real one. The standard approach is to build an annual time series using same-season composites: for each calendar year, stack all cloud-free observations within a defined seasonal window and derive a median or percentile surface. Comparing July composites across forty years, for instance, removes the seasonal signal and isolates the trend.
Trend detection then applies a non-parametric method such as the Mann-Kendall test to the annual area series. This test is distribution-free and tolerates the gaps and outliers that satellite records inevitably contain. The Sen slope estimator gives the median rate of change in hectares per year. Statistical significance thresholds should be reported honestly: a lake with high interannual variability may show a physically real decline that remains statistically ambiguous over a short record.
Step changes complicate trend analysis. A dam built upstream, a canal opened for irrigation abstraction, or a single exceptional drought year can introduce a structural break that a linear trend model misrepresents. Breakpoint detection methods, applied to the area time series before fitting trends, identify where the series changes character. This matters for attribution: a step change coinciding with a known infrastructure event points to abstraction; a gradual monotonic decline correlates better with reduced precipitation or rising evaporation.
Cloud contamination in humid regions: the honest accounting
In humid tropical regions, persistent cloud cover is the dominant practical constraint. Landsat's 16-day revisit means that in a persistently overcast month, zero usable observations may exist. Annual composites built from a handful of partly cloudy scenes can be biased toward the dry season, when cloud clears, systematically underestimating wet-season maxima and distorting trend estimates.
Several compositing strategies partially address this. Seasonal percentile compositing uses the 10th or 90th percentile of the MNDWI stack rather than the median, which can recover near-maximum or near-minimum extents even from sparse observations. Gap-filling with Sentinel-2 adds observations in years after 2015. Sentinel-1 SAR, though not the subject of this page, can penetrate cloud entirely and is sometimes used to anchor the water mask in persistently overcast periods. None of these strategies eliminates the fundamental problem: if cloud is present on every overpass for three months, no optical sensor recovers that period. Analysts should report data density per year alongside area estimates, so readers understand which years rest on thin evidence.
How altimetry closes the argument
Satellite radar altimetry from Sentinel-3 SRAL provides water-surface elevation along fixed ground tracks. Where a ground track crosses a lake, the elevation record extends back through earlier missions: Envisat, TOPEX/Poseidon, Jason-2 and Jason-3 collectively provide a record from the early 1990s for many large lakes. The Copernicus Global Land Service publishes near-real-time lake level products for several hundred monitored water bodies.
ICESat-2's ATL13 product offers a different geometry: dense along-track sampling at very high elevation precision, useful for smaller lakes that fall between Sentinel-3 ground tracks. The 91-day repeat limits its use for seasonal monitoring but it provides valuable snapshots for calibration and for lakes too small for radar altimetry to resolve cleanly.
The practical value of combining area and elevation is the hypsometric curve: the relationship between lake surface elevation and surface area. Once established from a period of good data, this curve allows volume change to be estimated even when only one of the two observables is available. It also allows analysts to flag inconsistencies: if area is stable but elevation is falling, the lake is deepening at its centre while the shoreline holds, which implies a different management or geological story than uniform recession.
From pixels to a number a water-resources ministry can use
The output of this analysis is not a single map. It is a time series of annual area estimates with uncertainty bounds, a trend rate with statistical confidence, a set of georeferenced shoreline polygons at key dates, and, where altimetry is available, a volume-change estimate. Each of these has a different audience. The trend rate is the headline for a minister. The shoreline polygons are inputs to irrigation planning and infrastructure siting. The volume-change estimate feeds water-balance models.
Satellize runs this workflow on open constellations for clients who need a defensible, auditable record rather than a one-off consultancy report. The methodology is the same published science used by the hydrology research community; the value is in operationalising it, maintaining the time series as new imagery arrives, and presenting results in formats that non-specialist decision-makers can interrogate. The Tonga crop-estimation programme demonstrated a similar principle in a different domain: routine satellite analytics, applied consistently, produce institutional knowledge that a single field survey cannot.
Minimum detectable change depends on lake size and shoreline geometry. For a lake with a gently shelving shoreline, a 30 m pixel represents a meaningful width of exposed lakebed; the area change corresponding to one pixel of recession around a 10 km perimeter is roughly 30 hectares. For a steep-sided lake, the same pixel-width recession represents far less area loss but may correspond to a larger volume change. Analysts should document the shoreline slope assumption when reporting detection limits.
Typical figures
| Spatial resolution (water delineation) | 30 m (Landsat DSWE/MNDWI); 10-20 m (Sentinel-2 MSI) |
| Revisit interval | 8 days effective (Landsat 8+9 combined); 5 days (Sentinel-2A+2B at mid-latitudes); 27 days (Sentinel-3 SRAL); 91 days (ICESat-2) |
| Archive depth | Landsat: 1982 to present (Collection 2); Sentinel-2: 2015 to present; Sentinel-3: 2016 to present; radar altimetry heritage missions: early 1990s to present |
| Water-surface elevation precision | 3-4 cm (Sentinel-3 SRAL, lakes >~1 km); a few centimetres (ICESat-2 ATL13 on calm water) |
| Key spectral bands | Green (~0.56 µm) and SWIR (~1.6 µm) for MNDWI; NIR for NDWI; Ku-band radar for Sentinel-3 altimetry; 532 nm photon-counting lidar for ICESat-2 |
| Minimum lake size (reliable delineation) | ~1 ha at Sentinel-2 resolution; ~1-2 km width for Sentinel-3 altimetry ground-track retrievals |
| Minimum detectable area change | Approximately 1-2 pixels (~0.09-0.18 ha at Sentinel-2; ~0.9-1.8 ha at Landsat), depending on shoreline contrast and compositing density |
| Latency (open data) | Landsat Collection 2: typically 1-3 days after acquisition; Sentinel-2: 1-3 days via Copernicus Data Space; Sentinel-3 lake levels: near-real-time via Copernicus Global Land Service |
| Delivery formats | GeoTIFF water masks, GeoPackage or Shapefile shoreline polygons, CSV or NetCDF area/elevation time series, PDF trend report |
Analytics Satellize can run
| Annual lake surface-area time series | MNDWI or DSWE thresholding on same-season annual composites from Landsat Collection 2; gap-filling with Sentinel-2 post-2015 | CSV time series with per-year area estimates and data-density flags; GeoTIFF composite stack |
| Trend rate and breakpoint report | Mann-Kendall trend test and Sen slope estimator on annual area series; BFAST or structural-break detection for step changes | PDF report with trend magnitude (ha/yr), statistical significance, and identified breakpoints with candidate attribution |
| Georeferenced shoreline polygon archive | Water-mask vectorisation at key dates (baseline, mid-period, recent); shoreline change envelope calculation | GeoPackage with dated shoreline polygons and recession-distance raster |
| Lake volume-change estimate | Hypsometric curve derived from combined Sentinel-3 or ICESat-2 elevation and Landsat/Sentinel-2 area observations; volume anomaly integration | NetCDF or CSV volume-anomaly time series with stated uncertainty bounds |
| Cloud-gap data-quality assessment | Per-year count of cloud-free observations within the compositing window; flagging of years below a minimum threshold | Data-density map and table included in all deliverables; advisory on which trend years carry elevated uncertainty |
| Seasonal versus secular decomposition | Harmonic regression or STL decomposition applied to full monthly area series to isolate trend component from seasonal cycle | Decomposition plot and tabulated seasonal amplitude versus trend magnitude |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.