SWOT Satellite Tracks Small Lakes With Twice as Many Observations

A customized processing workflow using lower-level satellite radar data produced accurate water-level records for 37 small Danish lakes while retaining about twice as many valid observations as the standard operational product. Researchers evaluated the Surface Water and Ocean Topography satellite’s High Rate Pixel Cloud, or HR PIXC, version D, against daily gauge measurements collected from July 2023 to May 2026. Across the lakes, the method achieved a median root-mean-square error of 5.57 centimetres and a Pearson correlation coefficient of 0.91 relative to in-situ records. It retained a median of 125 valid timestamps per lake, compared with 60 for filtered LakeSP data, allowing it to capture more rapid hydrological changes over the nearly three-year period.

The study focused on the relative movement of a lake’s surface rather than its absolute elevation. Each lake ranged from 0.25 to 40 square kilometres and was monitored against gauges referenced to Denmark’s DVR90 height system. Before comparison, the investigators removed the constant median difference between each gauge and its corresponding SWOT series, because the two systems use different vertical references and the study did not investigate the origin of those offsets. Gauge uncertainty depended on the sensor: pressure instruments had a stated uncertainty of 15 millimetres, while ultrasound instruments had uncertainties ranging from 5 to 30 millimetres. The evaluation therefore tested how well SWOT followed lake-level changes over time rather than whether it reproduced a gauge’s exact height.

SWOT differs from conventional nadir altimeters because its Ka-band radar interferometer can form two-dimensional images of inland water. The instrument observes within two swaths, each 50 kilometres wide, separated by a 20-kilometre nadir gap, and its High Rate PIXC product retains measurements beyond the defined swath boundaries. Denmark lies near 54.7 to 57.2 degrees north, where SWOT’s 21-day science orbit provides an average of 3.3 passes per cycle across lakes larger than 0.25 square kilometres. After accounting for 42 inaccessible acquisitions among 417, the effective temporal resolution was about 9.5 days. When the daily gauge series were downsampled to that interval, their median correlation with the original records was 0.96, explaining a median 92 percent of the observed variability.

To build a lake-level series, the researchers combined all relevant PIXC tiles for each satellite cycle and pass, then extracted observations inside static lake polygons. They used points classified as open water and rejected measurements falling outside the 99.7 percent confidence interval of elevations within the same acquisition. A median, chosen for its resistance to outliers, summarized the surviving points into one elevation per lake and timestamp. Acquisitions with fewer than 200 points were discarded because exploratory testing found that such cases could produce outliers. The final selection strategy also excluded points marked by SWOT’s geolocation quality flag. These choices reduced errors without requiring spatial weighting beyond the lake mask.

The resulting performance was close to that of the standard LakeSP product, but with an important temporal advantage. Filtered LakeSP produced a median root-mean-square error of 4.92 centimetres and a median correlation of 0.94, modestly better than the custom workflow. Its mean number of retained timestamps was 62.8 per lake, compared with 118.9 for the PIXC approach. The extra observations were especially valuable in Tivolisøen, a lake of only 0.269 square kilometres whose water level varied by more than 150 centimetres. The custom series captured peaks between November 2023 and February 2024, March and May 2024, and January and March 2026 that were absent from the filtered LakeSP record. At Søby Sø, by contrast, the seasonal signal was so small that it fell within the satellite’s error budget and produced low correlation.

Lake size alone did not determine performance. The smallest gauged lake, Tivolisøen, achieved a root-mean-square error of 6.82 centimetres and a correlation of 0.99 because its level changed substantially, while Mossø, covering 16.569 square kilometres, also showed low error and strong agreement with the gauge. The workflow used only the open-water class, potentially disadvantaging narrow or small lakes that contain many shoreline pixels, yet it remained effective at the sizes and hydrological conditions studied. A mask-sensitivity experiment expanded and contracted lake boundaries by 20 metres. Median error moved from 5.57 to 5.18 centimetres with expansion and 5.02 centimetres with contraction, while median correlation rose slightly from 0.91 to 0.92 and retained timestamps remained nearly unchanged.

The analysis also uncovered signs of a time-varying residual roll error in inland SWOT observations. A roll error would tilt apparent elevations across the satellite swath, producing differences between lakes on opposite sides of the ground track. The investigators therefore tested whether the offset-corrected difference between SWOT and gauges varied with each lake’s perpendicular distance from that track. An acquisition was classified as showing a significant roll component when the correlation had a p-value of 0.05 or lower and an absolute Pearson coefficient of at least 0.7. Among 160 acquisitions containing at least 10 observed lakes, 29, or about 18 percent, met those criteria. Examples showed slopes that changed in magnitude and even reversed direction between cycles, which argues against a fixed reference-surface error.

This tilt did not markedly change the median error of an individual pass, but it widened the spread of measurements, especially near the far-range edge of the swath. Across the qualifying acquisitions, the median absolute value of the central gauge–SWOT difference was 1.79 centimetres. The median one-standard-deviation spread was 5.43 centimetres and reached 12.60 centimetres at the 95th percentile. The study was designed to establish the error’s presence, not to measure or correct its absolute amplitude, so its quantitative treatment remains unresolved. The authors note that the effect is most consequential for lakes with small seasonal changes and for measurements of water-surface slope, including rivers oriented perpendicular to a SWOT ground track.

The results demonstrate a trade-off between the convenience of an operational product and the temporal detail available from lower-level data. LakeSP offered slightly better aggregate accuracy and could be queried in seconds, whereas processing the PIXC observations took about one to two minutes per pass over Denmark. Nevertheless, the custom workflow preserved roughly double the valid timestamps and captured higher-frequency variation without materially reducing overall accuracy. Its fixed filters are best suited to conditions similar to those in Denmark, where lake boundaries are relatively stable; highly dynamic lakes may require a date-matched dynamic mask. Steep terrain can also block radar observations, while ice, wind and elongated mountainous lakes may reduce usable data. Within those limits, the study shows that SWOT’s two-dimensional sampling can make small, previously inaccessible lakes practical to monitor at high temporal frequency while exposing a residual instrument effect that users must consider.

Subject of Research: Evaluation of SWOT satellite water-level time series for small Danish lakes and identification of residual roll errors

Article Title: Evaluation of SWOT HR PIXC version D water level time series of small lakes

Article References: Köhn, S. J., & Nielsen, K. (2026). Evaluation of SWOT HR PIXC version D water level time series of small lakes. Hydrology and Earth System Sciences, 30(20), 6443-6458. https://doi.org/10.5194/hess-30-6443-2026

Image Credits: AI Generated

DOI: 10.5194/hess-30-6443-2026

Keywords: SWOT satellite, small lakes, water-level time series, satellite altimetry, radar interferometry, Danish lakes, in-situ gauges, remote sensing, roll errors, Evaluation, SWOT, PIXC