Deficit-Driven Stratified Acquisition for Building STRATA-S12: An Open SAR-Optical Corpus for Self-Supervised Learning

Christian Sanchez-Leon; Dongming Chen; Dongqi Wang

eess.IVOct 13, 2026

Deficit-Driven Stratified Acquisition for Building STRATA-S12: An Open SAR-Optical Corpus for Self-Supervised Learning

Authors: Christian Sanchez-Leon, Dongming Chen, Dongqi Wang

Organizations: Software College, Northeastern University, Shenyang, 110819, China

Abstract

Building a co-registered SAR-optical corpus for self-supervised Earth Observation is above all an acquisition problem, yet the main multi-modal datasets are collected opportunistically, with no per-stratum quota control and no mechanism against the geographic dropout that non-uniform cloud cover leaves behind. We present Deficit-Driven Stratified Acquisition (DDSA), an algorithm that treats building a co-registered Sentinel-1/Sentinel-2 corpus as a constrained maximum-coverage problem over a hemisphere-corrected (class x season x region) stratification, enforces strict temporal pairing and a multi-stage quality gate, and closes the loop during acquisition with a human-in-the-loop deficit-repair stage. Measured on its own product, the algorithm yields three results a single-pass acquisition cannot report. Deficit-driven ordering makes a partial run a covering one: it holds every one of the 192 strata populated from the first few percent of the budget, where an abundance-first order leaves most strata empty until the end. Whether a failed attempt can be recovered depends on the temporal persistence of its cause, not the kind of defect: geometric failures such as nodata are among the most repairable (eta = 41.4%) while cloud is among the least (eta = 18.8%), because a cloudy location is cloudy on most dates of its season. The repair loop supplies 39.5% of the released pairs at no measurable quality cost, with 89.8% surviving manual review. The quality guarantee has a price: 2,928,370 acquisition attempts produced 740,746 released pairs. We release the corpus, STRATA-S12, with STAC/PDS4-aligned FAIR sidecars, and a pre-registered linear-probe check confirms it is fit for self-supervised pre-training. A metadata-level coverage comparison against SSL4EO-S12 and SEN12MS situates the corpus over a shared frame; systematic cross-corpus and cross-method comparison is left to follow-up work.

Figures & tables

Panel of Figure 1 — Figure 1: DDSA algorithmic flow. The 10 m ESA WorldCover map $L$ and a set of parameters (target $T$ , strata $S,R$ , sensor latitude limits $\lambda$ , temporal tolerance $\tau$ , cloud thresholds $c_{m},c_{r},c_{t}$ , max inter-block tile overlap $r$ ) drive a six-phase pipeline. Phases A–C build, stratify and order an acquisition plan. The block-level bulk acquisition (Phase D) is followed by an iterated human-in-the-loop tile-level recovery (Phase E) that explicitly counteracts cloud-induced geographic dropout. Phase F derives the full corpus $D$ , the reduced benchmark split $D_{\mathrm{lite}}$ , and the per-stratum statistics report $\Sigma$ .

Panel of Figure 2 — Figure 2: Representative Phase-E manual removals, one per failure mode. In each panel the left half is the Sentinel-2 optical tile and the right half the Sentinel-1 SAR composite of the same pair. These are defects the automatic Phase-D gates cannot detect by construction; for the temporal seam (c) the discontinuity appears co-registered in both the optical and the SAR channel, which a single-sensor artefact cannot produce. Dense fire-smoke removals also occur but are not shown. The complete list, with source-scene identifiers, is the Phase-E ledger released with the corpus.

Setting	Value	Role
Land-cover classes	8 of 11 (WorldCover v200)	stratification axis
Continental regions	6	stratification axis
Seasons	4 (hemisphere-corrected)	stratification axis
Season window	3-month meteorological	winter $Y$ = Dec $Y{-}1$ –Feb $Y$
Pairs per season	2 (balanced)	atomic S1–S2 pairs per location
Tile size	$512\times 512$ px @ 10 m	released tile

Table 1: Parameter settings used to build STRATA-S12. Symbols match the phase descriptions above and Algorithms 1 – 2 .

Panel of Figure 3 — Figure 3: DDSA system architecture. Three external data sources feed a Docker-containerised pipeline of four stages (pre-computation, acquisition, human-in-the-loop, archival), backed by two SQLite (WAL) state stores that drive scheduling and label-purity lookup. The released corpus is the bottom output.

Tile-level quality gates
	Rejected		Repair loop
Reason	$n$	%	recovered	exhausted	$\eta$ (%)
clouds	444,136	21.4	41,131	177,878	18.8
nodata	256,183	12.3	55,722	78,774	41.4
edge_padding	318	0.0	58	7	89.2
Image-wide corruption (source scene unusable)

Table 2: Acquisition rejection counts by reason, from the rejected_tiles log. Counts are per acquisition attempt, not per unique tile: a location may fail several times before a valid pair is found, so the total exceeds the number of released tiles. The left block is the rejection; the right block is what the repair loop then did with it — how many of those attempts it refilled, how many it retried without success, and the resulting efficiency

\eta

(Equation ( 4 )). Reading the two together is the point of the table:

\eta

varies by a factor of five across reasons, and not in the direction the reason suggests. Pairs removed after acceptance by the manual review do not appear here; they are logged as withdrawn pairs rather than as gate rejections and are counted in the scale-and-effort summary. Attempts not yet retried are excluded from

\eta

and reported there as well. Percentages are over the total number of rejected attempts, so the table is closed over its own rows and sums to 100 up to rounding; the wider count of every attempt the acquisition made (delivered, withdrawn and rejected) is reported separately in the scale-and-effort summary.

Figure 4: Recovery efficiency

\eta

by rejection reason, ordered by

\eta

. Bars are coloured by defect family (atmospheric/archive in one hue, geometric in another) and annotated with the number of retried attempts behind each ratio. The ordering inverts the intuitive reading of the taxonomy: the geometric failures, which look terminal, sit high, while cloud sits at the bottom — what separates them is whether the cause persists in time, not the kind of defect. Thin categories, whose

\eta

rests on few retries, are read against their annotated count rather than taken as stable estimates.

	Pair attempts			Delivered pairs
Scope	recovered	exhausted	$\eta$ (%)	from repair	share (%)
Global	562,747	912,290	38.2	292,432	39.5
Africa	147,611	174,121	45.9	78,309	39.6
Asia	158,733	302,889	34.4	88,156	46.2
Australia/Oceania	29,001	48,679	37.3	16,863	33.3
Europe	122,321	134,892	47.6	55,541	39.7

Table 3: Phase-E repair by scope. The left block counts pair attempts and yields

\eta

(Equation ( 4 )); the right block counts delivered pairs and the share of each scope’s corpus that the loop supplied. The two blocks are in different units and are not additive.

Figure 5: Per-region recovery efficiency

\eta

against cloud burden, the share of a region’s rejections that were cloud-driven. Cloud burden modulates recovery but does not determine it: the cloudier regions tend lower, yet the relation is not monotone — North America breaks it, with a moderate cloud share but the lowest efficiency — so what bounds recovery is regional scene availability and orbital geometry rather than cloud alone. The dashed line is a descriptive trend, not a fit.

Region	STRATA-S12	SSL4EO-S12	SEN12MS
N. America	12.1	25.8	22.1
S. America	9.7	12.0	7.3
Europe	18.9	17.5	24.7
Africa	26.7	14.1	17.5
Asia	25.8	29.7	23.2
Oceania	6.8	0.9	5.2

Table 4: Regional share of each corpus (% of that corpus’s own samples) over the six continental regions of the shared frame. Columns sum to 100 up to rounding.

Descriptor	DDSA $D$	DDSA $D$ (no repair)	SSL4EO-S12	SEN12MS
Hemisphere ratio (N:S)	2.92	2.74	4.99	4.09
Land-cover classes covered	8	8	7	8
Normalised coverage entropy	0.904	0.912	0.871	0.864
Tropical-band fraction (%)	32.7	33.6	18.9	22.7

Table 5: Coverage distribution of the DDSA corpus against opportunistic corpora, computed from public scene metadata over the shared (class

\times

season

\times

region) stratification. Higher normalised entropy and more classes covered indicate a broader spread across strata. The hemisphere ratio is read with care: land area is itself Northern-skewed, so part of any imbalance is geographic, but DDSA’s ratio is driven by its per-stratum quotas rather than by scene availability, so a strong skew in an opportunistic corpus points to cloud-induced dropout beyond that baseline.

Figure 6: Coverage of STRATA-S12 against SSL4EO-S12 and SEN12MS over the shared (class $\times$ season $\times$ region) strata. Each top panel is one corpus; cell colour is the share of that corpus’s samples falling in a (region, season) stratum, so the panels are comparable across corpora that differ in size by orders of magnitude, and the seasonal axis is read left-to-right within a panel. The lower-left panel is the $D-D(\text{no repair})$ difference: its near-zero cells show that the Phase-E repair loop adds volume roughly in proportion to the existing distribution rather than redistributing it, so the stratum balance comes from the deficit-driven ordering of Phase C upstream (cf. Table 5 and Section 4.8 ).

Figure 7: Global sample density of the three corpora, binned on a $4^{\circ}$ grid and normalised to each corpus so the panels compare distribution rather than size. The colour scale is logarithmic and shared: densities span several orders of magnitude between an empty cell and a dense continental one, and on a linear scale everything but the hottest cells collapses into the background. The three regimes are visibly different: SSL4EO-S12 concentrates its densest cells along the northern mid-latitudes and thins below the equator, the signature of a frame that follows population; SEN12MS is sparse everywhere, being two orders of magnitude smaller; STRATA-S12 spreads across the worked continents with its densest cells in equatorial Africa, the Amazon and southern Asia. The empty regions of the bottom-left panel are countries this acquisition has not reached yet, not strata it failed to fill (Section 4.9 ).

Figure 8: Marginal share of each corpus over the three stratification axes. Bars are percentages of each corpus’s own samples, so corpora differing in size by more than an order of magnitude remain comparable. The three marginals are shown in place of the full

K={192}

grid, which is unreadable at this width and answers the same question less directly: which continent, which land cover and which season each acquisition regime over- or under-represents. The season panel is drawn as a profile because the axis is an ordered cycle, with the uniform share marked so departures are read against a reference rather than against a truncated axis.

Budget	Ordering	Strata covered	Min-fill	Norm. entropy
5%	deficit-greedy (Phase C)	192	7	0.996
	abundance-first	2	0	0.130
	random	192	0	0.904
25%	deficit-greedy (Phase C)	192	7	0.982
	abundance-first	12	0	0.463
	random	192	2	0.904

Table 6: Coverage of the delivered pool under three acquisition orderings, at three budgets, over the shared

K={192}

strata. Min-fill is the count in the poorest stratum. Deficit-greedy is the Phase C rule; abundance-first emulates an availability-driven download. The orderings re-order the same delivered pairs, so they coincide at the full budget; the separation at a partial budget is what the ordering contributes.

Figure 9: Coverage versus acquisition budget for the three orderings over the delivered pool. Left: min-fill, the number of pairs in the poorest stratum — deficit-greedy holds a floor from the first few per cent of the budget, while abundance-first leaves the poorest stratum empty until the run is almost complete. Right: normalised coverage entropy. The orderings coincide at the full budget because they re-order the same pool; the gap at a partial budget is the contribution of the ordering (Section 4.9 ).

Setting	Value
Backbone	vit_small_patch16
Input channels	6
Mask ratio	0.7
Effective batch size	256
Base learning rate	1.5e-04
Weight decay	0.05

Table 7: MAE pre-training configuration (frozen for all seeds).

Figure 10: MAE pre-training on the representative subset of STRATA-S12: masked-reconstruction loss (top) and learning rate (bottom) over 100 epochs. The dashed line marks the end of the learning-rate warmup.

Metric	Pre-trained on $D$	Random init
micro-mAP (primary)	0.5552 $\pm$ 0.0022	0.5158 $\pm$ 0.0068
macro-mAP (descriptor)	0.3366 $\pm$ 0.0004	0.2920 $\pm$ 0.0046

Table 8: Frozen linear-probe performance on reBEN (BigEarthNet v2.0, 19-class multi-label), mean

\pm

standard deviation over 3 seeds. The pre-registered criterion is met when

\mathrm{mean}_{\text{pre}}-\mathrm{std}_{\text{pre}}

exceeds

\mathrm{mean}_{\text{rnd}}+\mathrm{std}_{\text{rnd}}

on micro-mAP.

Figure 11: Frozen linear-probe scores on reBEN (BigEarthNet v2.0), pre-trained on

D

versus random initialisation. Bars are the mean over 3 seeds and the black dots the individual seed values; error bars are one standard deviation. Both micro-mAP (the primary metric of the pre-registered criterion) and macro-mAP are shown so that neither is read in isolation.

Figure 12: Per-class average precision on reBEN, pre-trained on

D

versus random initialisation, averaged over 3 seeds. Each row is one of the 19 classes; the grey and green dots are the two arms and the connecting bar is the per-class gain from pre-training. Classes are ordered by that gain. The pre-trained encoder is ahead on every class, with the widest gains on the compositionally rich land covers and the narrowest on the classes either arm already separates well or that are too rare for a linear probe to recover.

Stratified Sampling

See all 10 papers

Aug 20, 2026

Far from the Crowd: Scalable Self-Supervised Learning via Geographic Isolation

Self-supervised pretraining on remote sensing imagery typically treats all samples as equally informative, despite large variability in geographic and visual structure. We propose a curriculum learning strategy for self-supervised Earth observation that ranks samples by geographic isolation, a label-free proxy derived entirely from geolocation metadata already present in geospatial datasets, requiring no image decoding, no model feedback, and no manual annotation. Unlike visual complexity proxies, it scales as O(D log D) with dataset size D and is well-defined for both contrastive and reconstructive objectives. We integrate the proposed measure into MoCoV2 and MAE pretraining and evaluate across three downstream tasks from CopernicusBench (BigEarthNet, DFC-2020, LCZ). Our curriculum reaches baseline final-epoch performance using as few as 20% of the training budget (MAE) and at most 40% (MoCo) of the training budget, and improves final downstream performance by up to +5 mAP on BigEarthNet, with gains of 1-5 points across benchmarks, matching visual-complexity curricula while reducing pre-computation cost by more than 140x (4 s vs. 568 s on SSL4EO). A CKA and effective-rank analysis further reveals that curriculum-trained encoders develop higher-dimensional, more uniformly utilized embedding spaces throughout training.

Daniele Rege Cambrin, Francesco Rossi, Mattia Varile

AIKO, Corso Stati Uniti 45, Turin, Italy

Self-Supervised Pre-Training Geospatial Representation Learning Curriculum Learning

Ask PDF

May 27, 2026

A Deep Learning Iterative Framework for Sentinel-1 Stripmap Enhancement Based on Azimuth Doppler Decomposition

Synthetic Aperture Radar (SAR) imagery enables all-weather, day-and-night Earth observation; however, it remains difficult to interpret due to speckle noise and other intrinsic imaging artifacts. Sentinel-1 (S1) constitutes one of the most widely used spaceborne SAR missions, offering systematic global coverage, high temporal resolution, dual-polarization imaging, and free data availability. Among S1 modes, Stripmap (SM) provides the highest resolution, yet speckle noise and spatial constraints often hinder applications requiring finer spatial detail. This motivates the need for effective image enhancement strategies. In this work, we propose a self-supervised enhancement framework for S1 SM imagery based on azimuth subaperture decomposition. The method exploits the physical consistency between subaperture reconstructions and the corresponding full-aperture image to generate paired training data without external sensors, simulated ground truth, or multi-temporal stacks. The proposed framework integrates single- and multi-frame learning and incorporates an iterative inference scheme that progressively refines image quality. Experiments on real S1 SM data show that the proposed approach consistently outperforms the widely adopted self-supervised deep learning baseline MERLIN, in terms of PSNR and SSIM, while MERLIN attains higher ENL, highlighting a trade-off between structural fidelity and speckle smoothing. Overall, the results demonstrate that subaperture-based supervision provides a physically grounded, reproducible, and operationally viable approach for SAR image enhancement using S1 data. It is worth noting that the proposed approach can be extended to other SAR platforms, polarizations, and acquisition modes.

Juan Francisco Amieva, Christian Ayala, Roberto Del Prete +1

Tracasa Instrumental S.L. · European Space Agency · Public University of Navarre

Image Enhancement Self-Supervised Learning Remote Sensing Synthetic Aperture Radar Image Denoising

Ask PDF

Jun 18, 2026

SARLO-80: Worldwide Slant SAR Language Optic Dataset 80cm

Multimodal foundation models have advanced rapidly thanks to large optical benchmarks, but comparable resources for synthetic aperture radar (SAR) remain limited. Existing SAR--optical datasets largely rely on low-resolution, intensity-only Ground Range Detected~(GRD) products and do not preserve complex-valued SAR measurements or native acquisition geometry, which restricts physically grounded multimodal learning. In particular, large-scale public datasets combining very-high-resolution (VHR) SAR SLC, aligned optical imagery, and natural-language descriptions are still lacking. We present a VHR SAR--optical--text dataset built from open-access Umbra spotlight acquisitions distributed as Sensor Independent Complex Data (SICD). From around 2,500 worldwide scenes (VV/HH, 20cm--2m native resolution), we standardize all SAR data to an 80cm slant-range grid via band-limited FFT resampling and tile the imagery into 1024 by 1024 patches. For each SAR patch, we retrieve a high-resolution optical tile and warp it into the SAR grid using local coordinate correspondences for local pixel-level alignment. We further generate three caption variants (SHORT/MID/LONG) per sample to support vision--language training and evaluation. Our dataset contains 119,566 triplets (complex and amplitude slant-range SAR patch, aligned optical patch, natural-language description) covering 257 locations across 72 countries and a broad range of land types and infrastructures. We release fixed train/validation/test splits and the full preprocessing and baseline code to enable reproducible benchmarks for multimodal alignment on cross-modal retrieval and conditional generation in native SAR geometry. The dataset is publicly available on the Hugging Face Hub at https://huggingface.co/datasets/ONERA/SARLO-80.

Solène Debuysère, Nicolas Trouvé, Nathan Letheule +2

1DEMR-ONERA – The French Aerospace Lab, Université Paris-Saclay, Palaiseau, France · 2DTIS-ONERA – The French Aerospace Lab, Université Paris-Saclay, Palaiseau, France · 3Hugging Face, London

Image-Text Retrieval Cross-Modal Retrieval Synthetic Aperture Radar Remote Sensing Benchmarks

Ask PDF

Abstract

Stratified Sampling

Explore similar work

Far from the Crowd: Scalable Self-Supervised Learning via Geographic Isolation

A Deep Learning Iterative Framework for Sentinel-1 Stripmap Enhancement Based on Azimuth Doppler Decomposition

SARLO-80: Worldwide Slant SAR Language Optic Dataset 80cm