Extending Dynamic World Surface Water Mapping to Sentinel-1 with AlphaEarth Embeddings

Rohit Mukherjee; Frederick Policelli; Beth Tellman; TC Chakraborty; Jonathan Giezendanner; Jonathan A. Sullivan; Ning Sun

cs.CVOct 5, 2026

Extending Dynamic World Surface Water Mapping to Sentinel-1 with AlphaEarth Embeddings

Authors: Rohit Mukherjee, Frederick Policelli, Beth Tellman, TC Chakraborty, Jonathan Giezendanner, Jonathan A. Sullivan, Ning Sun

Organizations: Pacific Northwest National Laboratory, Richland, WA 99354 USA

Abstract

Dynamic World (DW) maps land use and land cover globally at 10 m from Sentinel-2 (S2) imagery, but only for cloud-free observations, which limits where and when surface water can be mapped. We use the DW water class as weak supervision for a Sentinel-1 (S1) synthetic aperture radar (SAR) model so that DW-like water maps can be produced for every S1 acquisition. Google's AlphaEarth Foundations (AEF) annual embedding supplies spatial context, while S1 backscatter supplies the acquisition-time observation. On 53 globally distributed scenes with independent annotations of 3 m PlanetScope imagery acquired within 48 h of the S1 overpass, the S1-only model already reaches a pooled water intersection over union (IoU) of 0.77, comparable to 0.75 for the operational OPERA DSWx-S1 product, and adding AEF raises it to 0.85. The fused model improves on the S1-only model on 44 of 53 scenes and exceeds OPERA on 48, and on the independent S1S2-Water benchmark it reaches 0.94, compared with 0.87 for OPERA. Optical land-cover products can thus provide scalable training labels for SAR surface water mapping.

Figures & tables

Panel of Figure 1 — Fig. 1: (a) Weakly supervised S1 and AEF workflow; DW labels are used only in training. (b) Three example scenes from the Global Surface Water Dataset (GSWD) in Romania (SID23), Vietnam (SID46), and Spain (SID17) (10 m inference, threshold 0.30), with per-panel water intersection over union (IoU) against the independent PlanetScope-based GSWD annotations on the common valid pixels. In the Spain scene (SID17), a narrow river, the S1-only model is comparable to the fused model. SID denotes the scene identifier in the PlanetScope reference dataset.

Method / product	Input	Precision	Recall	Dice	Water IoU (pooled)	Water IoU (per-scene mean [95% CI])
S1-only	S1	0.883	0.854	0.869	0.768	0.617 [0.527, 0.702]
S1 + AEF ( $t$ )	S1 $+$ acq.-year AEF	0.891	0.950	0.920	0.851	0.741 [0.669, 0.806]
OPERA DSWx-S1	S1 $+$ static ancillary layers	0.859	0.851	0.855	0.747	0.591 [0.508, 0.670]

TABLE I: Surface Water Mapping Accuracy on 53 Globally Sampled Surface Water Scenes (10 m Inference Scored on the 3 m Reference, Threshold 0.30, Common Valid Pixels; Model Seed 42)

Fig. 2: Per-scene water IoU for 53 scenes (10 m inference, threshold 0.30, common valid pixels). Curves show smoothed distributions, diamonds are means, and bars are bootstrap 95% CIs. Outlined circles, squares, and triangles mark SID23, SID46, and SID17 from Fig. 1 (b). S1+AEF(

t

) exceeds S1-only on 44/53 scenes and OPERA on 48/53 (Wilcoxon

p<0.05

(a) AEF feature width $k$ , full training set
Configuration	Water IoU (per scene)	Recall
S1-only ( $k{=}0$ )	0.624 $\pm$ 0.007	0.867 $\pm$ 0.015
$k{=}1$	0.730 $\pm$ 0.006	0.940 $\pm$ 0.006
$k{=}2$	0.737 $\pm$ 0.002	0.948 $\pm$ 0.002
$k{=}3$	0.741 $\pm$ 0.004	0.950 $\pm$ 0.003
$k{=}4$	0.741 $\pm$ 0.002	0.950 $\pm$ 0.003

TABLE II: Sensitivity Analyses on the 53 Surface Water Scenes (10 m Inference, Threshold 0.30, Common Valid Pixels; Mean

\pm

SD over Three Seeds)

Method	Prec.	Recall	IoU (pooled)	IoU (per scene)	Scenes $>$ OPERA
S1-only	0.806	0.982	0.794 $\pm$ 0.026	0.723 $\pm$ 0.019	6–7
S1 + AEF ( $t$ )	0.946	0.995	0.941 $\pm$ 0.001	0.862 $\pm$ 0.002	12
OPERA DSWx-S1	0.904	0.957	0.869	0.763	–

TABLE III: Accuracy on the 15 S1S2-Water Test Scenes (30 m Grid, Threshold 0.30, Common Valid Pixels; Model Values Are Mean

\pm

SD over Three Seeds)

Scene	S1 date	Land cover	Water frac.	OPERA	S1-only	S1+AEF
23	2019-08-18	Shrubland	0.008	0.832	0.034 $\pm$ 0.006	0.813 $\pm$ 0.012
25	2019-08-27	Tree cover (broadleaved)	0.007	0.400	0.630 $\pm$ 0.005	0.671 $\pm$ 0.016
29	2019-09-02	Herbaceous cover	0.036	0.681	0.641 $\pm$ 0.034	0.791 $\pm$ 0.008
33	2019-09-18	Tree cover (broadleaved)	0.071	0.804	0.800 $\pm$ 0.014	0.884 $\pm$ 0.006
35	2019-09-16	Tree cover (needleleaved)	0.040	0.956	0.787 $\pm$ 0.039	0.920 $\pm$ 0.001
47	2019-01-04	Cropland rainfed	0.081	0.788	0.858 $\pm$ 0.002	0.875 $\pm$ 0.003

TABLE S1: Per-scene water IoU on the 15 S1S2-Water test scenes (30 m grid, common valid pixels). S1-only and S1+AEF: mean

\pm

SD over three seeds, threshold 0.30 as in the main paper.

Fig. S1: Four S1S2-Water scenes showing S1 backscatter, reference water, and predictions from S1-only, S1+AEF, and OPERA. Models use seed 42 and threshold 0.50. IoU is scored at 30 m on common valid pixels.

GSWD
Scoring	S1-only	S1+AEF	OPERA	S1+AEF $-$ OPERA	Scenes $>$ OPERA
Main paper (3 m)	0.768	0.851	0.747	0.104	48/53
(i) Both at 30 m	0.767	0.855	0.749	0.106	47/53
(ii) Binary 30 m model	0.761	0.841	0.747	0.095	48/53
(iii) Pure 30 m blocks	0.837	0.916	0.825	0.091	47/53
(iv) OPERA grid a	0.764	0.854	0.751	0.103	42/46

TABLE S2: Resolution-Matched Comparison with OPERA DSWx-S1, Pooled IoU (Seed 42, Threshold 0.30, Common Valid Pixels of Each Variant)

Explore similar work

Beyond Backscatter: AlphaEarth Land-Cover Priors for Rapid SAR Flood Segmentation Across Foundation Backbones
Jun 28, 2026Sanjay Thasma, Yu-Hsuan Ho, Ali MostafaviFlood Detection Land Cover
Department of Computer Science and Engineering, Texas A&M University, College Station, TX · Urban Resilience.AI Lab, Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX
Learning Regional Snow Water Equivalent and Snow Height Variations from Sentinel-1 InSAR Acquisitions
Sep 28, 2026Luca Barco, Lorenzo Innocenti, Bianca Bartoli +3Sentinel-2 Sea-Surface Temperature Data
Politecnico di Torino, Dipartimento di Automatica e Informatica, Torino, Italy · Fondazione LINKS, AI Data and Space, Italy
High-Resolution Flood Mapping With Sentinel-1 and Sentinel-2 via Misalignment-Robust Cross-Sensor Learning and Generative Despeckling
Jun 29, 2026David Ma, Jeremy Feinstein, Shreya Pandit +2Flood Detection Sentinel-2
Argonne National Laboratory, Lemont, IL 60439 USA · Ohlone College, Fremont, CA 94539 USA

Jun 28, 2026, cs.CVJ/K move · Enter open · S save

Beyond Backscatter: AlphaEarth Land-Cover Priors for Rapid SAR Flood Segmentation Across Foundation Backbones

Sanjay Thasma, Yu-Hsuan Ho, Ali Mostafavi

Department of Computer Science and Engineering, Texas A&M University, College Station, TX · Urban Resilience.AI Lab, Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX

Flood Detection Land Cover Prior Knowledge Integration Dinov3 Beyond Retrieval

Open paper AskPDF arXiv

Rapid flood mapping is critical for emergency response, yet optical imagery is often unusable during major flooding and single-temporal SAR is ambiguous, since new inundation, permanent water, and other smooth surfaces produce similar backscatter. This study evaluates whether stable land-context priors can improve post-event SAR flood segmentation when a registered, seasonally matched pre-event acquisition is unavailable. Using the CONUS (Continental United States) subset of ImpactMesh-Flood, we compare four backbones spanning distinct pretraining regimes-a from-scratch CNN UNet, an ImageNet-pretrained UNet, the SAR-pretrained TerraMind Vision Transformer, and the optical-satellite-pretrained DINOv3 Vision Transformer-in SAR-only, SAR+DEM, and SAR+AlphaEarth configurations under an identical fusion design, training protocol, and event-stratified split. Models are selected on a validation flood event and evaluated separately on two held-out events, Hurricane Florence and the Louisiana floods, with three-seed reporting for auxiliary configurations. Both auxiliary priors improve over the observed SAR-only baselines across all backbones and test events. AlphaEarth exceeds DEM on the harder Florence event for every backbone and achieves the best Florence IoU, while DEM is competitive on Louisiana and produces the best result there. The seed analysis reveals a trade-off: DEM is more stable across initializations, whereas AlphaEarth offers higher peak performance and higher recall on the harder event. Cross-event differences track flood-class prevalence and similarity to the training distribution, underscoring the need for per-event evaluation. We reframe single-temporal SAR flood segmentation as an alignment between radar observations and stable land-surface priors, where learned and physical context offer complementary pathways to more reliable rapid flood mapping.

Abstract

Explore similar work

Beyond Backscatter: AlphaEarth Land-Cover Priors for Rapid SAR Flood Segmentation Across Foundation Backbones

Learning Regional Snow Water Equivalent and Snow Height Variations from Sentinel-1 InSAR Acquisitions

High-Resolution Flood Mapping With Sentinel-1 and Sentinel-2 via Misalignment-Robust Cross-Sensor Learning and Generative Despeckling