Mapping 40,000 waterbodies from satellite data

A Sentinel-2 pixel in the 10 m bands covers 100 square metres. A small pond of 0.1 hectares therefore spans about ten pixels, and on Landsat's 30 m grid it is barely one. That single fact shapes almost every decision in a waterbody mapping programme: which sensor to use, which index to compute, how to deal with cloud and season, and how far you can trust an automated result without sending someone into the field to check it.

The project in brief

Over two years, Kaiinos delivered a regional water programme that identified 40,000 surface waterbodies, including lakes, reservoirs and ponds. The programme needed more than a map. It needed a register: one record per waterbody, with a geometry, an identifier and attributes that other government departments could rely on and reuse in their own systems.

We built three components. The first was machine-learning based satellite data analysis to extract water from imagery across the whole region. The second was a set of mobile apps for field validation, so survey teams could confirm, correct or reject what the models found. The third was a web GIS for enterprise integration, which made the register available to other government systems rather than leaving it as a file on one team's server.

Why a waterbody register matters

Small surface waterbodies do a lot of quiet work. They store wet-season runoff for irrigation, recharge shallow aquifers, buffer local floods and supply livestock. Agriculture is the largest water user on the planet: UN-Water reports that agriculture accounts for 72% of all freshwater withdrawals, against 16% for municipalities and 12% for industry. In regions that depend on rain-fed farming, a reservoir or pond can decide whether a second crop is possible.

Without a register, these assets are hard to protect. Encroachment is the obvious risk: a reservoir bed that is dry for part of the year can be filled, ploughed or built on, and without a dated baseline there is no evidence that water was ever there. Drought planning has the same problem: you cannot prioritise desilting, inlet repair or recharge works without knowing how many waterbodies exist, where they are and how their extent behaves through the year. A register with stable IDs turns a scattered asset base into something that can be counted, monitored and managed. Deciding where new storage should go is the mirror problem, which we cover for check dams in our forest monitoring case study.

The spectral basis of water detection

Clear water absorbs strongly in the near infrared (NIR) and shortwave infrared (SWIR), while reflecting relatively more in the green band. Vegetation does the opposite in the NIR, and dry soil reflects strongly in both. Normalised difference indices exploit this contrast. The Normalised Difference Water Index of McFeeters (1996) is computed as (Green minus NIR) divided by (Green plus NIR). Water pixels return positive values, while vegetation and soil tend to return values near zero or below.

NDWI has a known weakness. Xu's 2006 paper in the International Journal of Remote Sensing notes that water enhanced with NDWI is often mixed with built-up land noise, which overestimates water area. Xu's Modified NDWI (MNDWI) replaces the NIR band with a middle infrared band, computed as (Green minus SWIR) divided by (Green plus SWIR). Built-up surfaces, soil and vegetation then return negative values, and open water separates more cleanly. On Sentinel-2 this means using B3 with B11, which requires resampling because B11 is a 20 m band.

Sensors and their trade-offs

No single sensor solves the problem. Optical sensors give the spectral information water indices need but cannot see through cloud. Radar sees through cloud but carries no spectral signature, so production pipelines combine them.

Sentinel-2 is usually the primary source for small waterbodies because of its 10 m green and NIR bands. Landsat's longer archive establishes what existed before recent change. Sentinel-1 fills the wet season, when waterbodies are fullest and optical scenes are most often clouded out. Calm open water reflects the pulse away from the sensor and appears dark, while rough land scatters energy back. Wind-roughened water, radar shadow behind hills and double-bounce from emergent vegetation all confuse this, so VV and VH backscatter work best as classifier inputs rather than as a stand-alone water map.

Thresholding and where it breaks

The simplest extraction is a threshold: every pixel with NDWI or MNDWI above a cut-off is water. A fixed value of zero is a common starting point, and automatic methods such as Otsu's threshold adapt the cut-off to each scene's histogram. Thresholding is fast and transparent, and works well on large, clear lakes. The problems are the exceptions, which across tens of thousands of features are not rare.

Terrain and building shadows have low reflectance across all bands and can mimic water. Dark asphalt and some roofing materials confuse NDWI, which is the issue MNDWI was designed to reduce. Turbid water carrying suspended sediment reflects more in the red and NIR, pushing index values down towards land. Floating and emergent vegetation, common in shallow waterbodies, gives a vegetation signature over what is functionally a waterbody. Each failure mode produces either false positives or missed water, and no single threshold fixes all of them.

Machine learning classifiers

Machine learning replaces a single rule with a model trained on labelled examples. A random forest classifier takes a stack of features per pixel: raw band reflectances, NDWI, MNDWI, NDVI, SAR backscatter, slope from a DEM and texture measures. It learns combinations that a threshold cannot express, for example that a low-reflectance pixel on a steep slope is more likely shadow than water. Random forests train quickly, tolerate noisy labels and give feature importance scores that help explain results.

Convolutional networks go further by using spatial context. U-Net, introduced by Ronneberger, Fischer and Brox for biomedical segmentation, uses an encoder-decoder architecture with skip connections that preserve fine boundaries, which suits water edges well. A U-Net sees the shape of an embankment, the pattern of an inlet channel and the texture of floating vegetation, not just the spectrum of one pixel. Model quality is measured with intersection over union (IoU), F1, precision and recall on held-out labelled tiles.

Seasonality and multi-date composites

A waterbody is not a fixed object. A reservoir can be full in the months after the rains, shrink through the dry season and be completely dry before the next rains. A map from a single date captures one moment, which is why programmes classify many dates and summarise them. The JRC Global Surface Water Explorer, built from Landsat imagery under the Copernicus Programme, shows this approach at global scale.

According to its documentation, the JRC dataset is mapped at 30 m and provides layers for water occurrence, occurrence change intensity, seasonality, recurrence and transitions. A register needs similar summaries per feature. The maximum observed extent defines the waterbody's boundary for protection against encroachment. The minimum extent and the frequency of observed water describe how permanent it is. Building these from multi-date composites, with cloud masking per scene, gives a stable outline that does not depend on the date a single image happened to be captured.

From raster to register with stable IDs

Classifiers output rasters; a register needs vectors. The conversion has several steps: apply a morphological filter to remove isolated pixels, polygonise connected water regions, simplify geometry to remove pixel staircase edges, and reproject into the coordinate reference system the downstream systems expect. A minimum mapping unit, expressed as a pixel count or area, removes fragments too small to classify reliably.

The harder problem is identity. The same reservoir appears in dozens of dates, split into fragments at low water and merged with a neighbour at high water. Deduplication matches polygons across dates by overlap, assigns each physical waterbody one record and gives it an identifier that persists when the geometry is updated. Stable IDs are what make a register useful: field reports, repair works and encroachment cases can all reference the same record for years, even as the outline is refined.

Field validation and integration

Satellite classification produces candidates, not facts, which is why the programme included mobile apps for field validation. Apps of this kind let survey teams navigate to a mapped waterbody, confirm that it exists, correct its type or boundary and record attributes that imagery cannot provide, such as local name, condition of the embankment or inlet, and use. Every record captured with a GPS fix and a photograph becomes evidence attached to the register entry.

The same field records feed back into the models as labelled training data, which helps most with shadows, turbid water and vegetated waterbodies. Apps for this work need offline maps, because many waterbodies sit outside reliable mobile coverage.

The web GIS let other government systems consume the register directly rather than receive file exports. The OGC Web Feature Service standard gives feature-level access to geometries and attributes through operations such as GetFeature, and OGC notes that its newer OGC API - Features offers similar functionality as a modern web API.

Publishing through standards keeps one authoritative copy: when a field team corrects a boundary, every connected system sees the update.

Limits worth stating plainly

Pixel size sets a hard floor. At 10 m GSD, a 0.1 hectare pond covers about ten pixels, many of them on the edge. At 30 m, it is roughly one pixel, and detection becomes unreliable. Mixed pixels, where a single pixel contains both water and land, blur boundaries and bias area estimates. Narrow features such as channels and canals are especially affected. Higher resolution commercial imagery helps with the smallest features but adds licensing cost and usually offers a shorter archive.

Other limits are physical. Waterbodies under tree canopy are partly invisible to optical sensors. Seasonal waterbodies that were dry on every cloud-free date in the analysis window can be missed entirely, which is why historical archives and field input matter. Stating these limits up front tells programme managers where field effort should go. Similar trade-offs appear across our agriculture work, where small parcels push the same resolution limits.

Surface water as terrain for robots

A waterbody register is also a base map for autonomous machines. Autonomous surface vessels survey bathymetry with single-beam or multibeam echo sounders. The maximum extent polygon becomes the geofence, and survey lines are generated inside it at a spacing set by the sounder's swath and the water depth. The latest Sentinel-1 or Sentinel-2 extent shows whether the waterbody holds enough water to launch at all. The result is a bed surface model. Combined with the shoreline at a known water level, it gives a stage-volume curve, so engineers can see how much storage siltation has taken.

Drones inspecting embankments for seepage, breaches and vegetation need the embankment centreline as a flight line and a DSM to hold a constant height above the crest. Water quality sampling boats need routes that stay inside the current shoreline and reach representative points. In every case, results are filed against the stable ID, so a repeat survey is compared with the same record.

Change triggers close the loop. A sharp fall in observed water frequency, or a land-cover change inside the maximum extent, can flag possible encroachment or embankment failure and become the next drone mission.

Our Smartout platform maintains operational digital twins that combine bathymetry, satellite and drone data for environments such as fisheries and maritime sites. The next layer we are building towards is a robot perception engine that syncs robot-ready twins to machines and takes their sensor feeds back. Mission planning and fleet coordination are longer-term directions, described on our robotics solutions page.

Frequently asked questions

What is the difference between NDWI and MNDWI?

NDWI (McFeeters, 1996) uses the green and near infrared bands. MNDWI (Xu, 2006) replaces near infrared with a shortwave infrared band, which suppresses the built-up land noise that NDWI often confuses with water. On Sentinel-2, NDWI uses B3 and B8 at 10 m, while MNDWI uses B3 and B11, a 20 m band that must be resampled.

Can satellites detect waterbodies through cloud?

Optical sensors such as Sentinel-2 and Landsat cannot. Synthetic aperture radar such as Sentinel-1 can, because it emits its own microwave signal and works day or night through cloud. Calm water appears dark in backscatter. Radar is usually combined with optical indices as classifier inputs, since wind, terrain shadow and vegetation can confuse it.

What is the smallest waterbody satellite data can map?

It depends on ground sample distance. At Sentinel-2's 10 m resolution, a 0.1 hectare pond covers about ten pixels and can usually be detected, though its edges contain mixed pixels. At Landsat's 30 m, the same pond is roughly one pixel. Higher resolution commercial imagery helps for the smallest features, and field validation remains essential.

Why does a waterbody register need stable IDs?

Water extent changes through the year, so the same waterbody can appear as several fragments or merge with a neighbour depending on the date. A stable ID gives each physical waterbody one persistent record, so field reports, repair works, encroachment cases and other systems refer to the same feature for years, even as its boundary is refined.

How is a waterbody register shared with other systems?

Through a web GIS that publishes the register using open standards such as the OGC Web Feature Service or OGC API - Features. Other systems can then query features by ID or area directly, without exporting files. One authoritative copy is maintained, and corrections from field validation reach every connected system.

Start with the map machines need

Whether you are managing water assets or planning autonomous surveys on them, the register is the foundation. We can walk through the pipeline, sensor choices and integration options for your region. Book a technical call with our team.