Can AlphaEarth embeddings extend dynamic world water mapping to Sentinel-1?

Reproducing Dynamic World water from radar alone

Dynamic World classifies 10 m Sentinel-2 images in near real time, but only where clouds and cloud shadows allow it, which limits valid coverage especially in the tropics. Sentinel-1 C-band SAR observes through clouds day and night, yet smooth non-water surfaces, radar shadow, and other low-backscatter terrain can be mistaken for open water. The paper asks whether a model can reproduce Dynamic World water maps from Sentinel-1 on cloudy days, complementing the optical product and increasing observation frequency.

The approach uses the Dynamic World water class as weak supervision, meaning the labels come from an existing optical land-cover product rather than manual annotation. The authors train an S1 water model on those pseudo-labels alone, then test whether adding Google's AlphaEarth Foundations embeddings as spatial context improves it. AEF compresses a year of optical, radar, lidar, and other Earth observations into a 64-dimensional embedding for each 10 m pixel, encoding persistent land cover and, indirectly, water occurrence.

S1 supplies the moment, AEF supplies the context

The two inputs are not interchangeable. Sentinel-1 backscatter supplies the acquisition-time observation, while the AEF embedding is fixed within a calendar year and cannot by itself represent within-year changes in water extent. Its role is to supply spatial context for interpreting that observation, letting the model learn which context helps rather than applying fixed rules.

This matters because operational SAR products such as OPERA DSWx-S1 reduce false detections over radar-dark land using static ancillary layers, including land cover, long-term water occurrence, and height above nearest drainage. AEF serves a similar purpose but is learned jointly with S1 from the training labels. The authors note HAND could complement their model as an input, while learned embeddings could inform rule-based masks.

Fifty-three scenes scored against independent 3 m labels

The main evaluation uses 53 globally distributed scenes with independent annotations of 3 m PlanetScope imagery acquired within 48 hours of the S1 overpass. S1–PlanetScope offsets have a median of 19.9 hours, a maximum of 44 hours, and are at most 24 hours for 30 scenes. Reference values of 0 and 1 denote non-water and water and are valid for scoring; any other value is excluded.

Metrics are aggregated two ways, and the distinction matters. Pooled metrics aggregate confusion counts across scenes, whereas per-scene IoU gives each scene equal weight. The paper reports pooled precision, recall, and Dice, plus water IoU in both forms, with bootstrap confidence intervals for the per-scene mean and paired Wilcoxon signed-rank tests for within-scene differences. The primary comparison is OPERA DSWx-S1's 30 m B02 BWTR layer for the same S1 acquisition, where water includes inundated vegetation and HAND-excluded, layover/shadow, ocean, and no-data pixels are invalid.

Weak labels alone reach OPERA-level accuracy

Trained on Dynamic World pseudo-labels alone, without manual annotation, the S1-only model reaches a pooled water IoU of 0.77 on the 53 scenes, comparable to 0.75 for the operational OPERA DSWx-S1 product. At threshold 0.50 the pooled IoU is 0.77 for S1-only, 0.86 for S1+AEF, and 0.85 for the prior-year substitution.

The authors describe this as the S1-only model matching or slightly exceeding OPERA on this benchmark. It is a notable result because the model was trained without hand-labeled SAR water data, suggesting optical land-cover products can provide scalable training labels for SAR surface water mapping. The comparison is against OPERA's 30 m product resampled to the 3 m reference grid, and the paper notes published OPERA validation scores only entirely water or non-water 30 m pixels, whereas this evaluation includes all valid pixels, treating boundaries more strictly.

Annual context raises IoU and dampens input sensitivity

Adding AEF raises pooled water IoU from 0.77 to 0.85. The fused model improves on the S1-only model on 44 of 53 scenes and exceeds OPERA on 48 of 53, with Wilcoxon p < 0.05. On the independent S1S2-Water benchmark it reaches 0.94, compared with 0.87 for OPERA.

The gain is not only in the mean. Across three training seeds, per-scene IoU varies less for the fused model than for S1-only, with standard deviation 0.001 versus 0.007. Using the S1 images distributed with S1S2-Water instead of the authors' Earth Engine S1 lowers pooled S1-only IoU from 0.82 to 0.72, but fused IoU only from 0.94 to 0.93. Running the network on GSWD inputs resampled to 3 m instead of the 10 m training grid lowers S1-only IoU from 0.77 to 0.57, while the fused model stays at 0.85. Because S1 preprocessing differs between processing chains, the authors present this stability as practically useful.

AEF fixes missed water in one benchmark and false positives in the other

The two benchmarks expose different S1-only errors, and AEF reduces both. On the 53-scene GSWD benchmark, AEF mainly recovers missed water, most of it in WorldCover permanent water bodies. Overlaid on WorldCover 2021, the false negatives corrected by AEF, 1.38 million pixels, lie mainly in permanent water bodies (77%) and mangroves (13%); the corrected false positives, 0.84 million pixels, lie mainly on bare or sparse vegetation (59%) and tree cover (20%). The authors state these associations describe where errors occur, not their cause.

On S1S2-Water, whose scenes cover whole landscapes with little water, S1-only errors are mainly false positives, and AEF raises precision from 0.81 to 0.95. The gain is concentrated in scenes where S1 alone fails. AEF raises scene IoU by more than 0.05 on 24 scenes, changes it by at most 0.05 on 26, and lowers it by more than 0.05 on three.

Annual embeddings favor persistent water and miss narrow channels

Several limitations qualify how far these results transfer. Annual context can favor areas of persistent water, where most false positives added by AEF lie; 62% of the added false positives, 0.81 million pixels, are in permanent water bodies. The model was not designed or evaluated for post-flood inundation, and it can reproduce systematic Dynamic World label errors.

The evaluation split is tile-level rather than basin-held-out, and most training tiles, 4678 of 5278, are centered on river reaches. At 10 m, predictions have limited accuracy on narrow channels; on scene SID26, fused IoU is 0.268 at 10 m versus 0.525 with 3 m inputs. Acquisition-year AEF can also include observations after the S1 acquisition. On S1S2-Water, the fused model was neither trained nor tuned on that benchmark, yet at native resolution and threshold 0.50 its mean per-scene IoU of 0.88 is close to the 0.85 reported for the best S1 model trained on that benchmark, though scene sets and averaging differ.

The decision-relevant split is scene type, not the average

The headline numbers hide the paper's most useful distinction. AEF's benefit is not uniform: it raises scene IoU by more than 0.05 on 24 of 53 scenes, leaves 26 essentially unchanged, and lowers three. On GSWD the corrected errors are dominated by missed permanent water; on S1S2-Water the gain comes from suppressing false positives in landscapes with little water, lifting precision from 0.81 to 0.95.

For a reader deciding whether this approach fits a task, the relevant question is which error dominates. Where the concern is missed water in persistent water bodies and mangroves, AEF helps most. Where the concern is false positives over radar-dark land in mostly dry landscapes, AEF also helps, but through precision rather than recall. Where the target is narrow river channels at 10 m, or post-flood inundation, the evidence does not support the same expectation. The prior-year substitution retains 94% of the pooled gain over S1-only, which suggests the context does not need to be perfectly time-matched, but the model still cannot represent within-year water change from AEF alone.

Sources used in this answer