IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
1
Label-Decoupled Style Augmentation for Domain Generalization in Multi-Label Remote Sensing Scene Classification
arXiv:2607.12704v1 [cs.CV] 14 Jul 2026
Alaa Almouradi , Erchan Aptoula
Abstract—Multi-label classification assigns several cooccurring labels to each aerial scene, yet deployed models often encounter data distributions different from their training. Feature-statistics augmentation such as MixStyle, EFDMix, and correlated style uncertainty improves generalization at low cost but perturbs channel statistics globally, treating each image as a single style; one class can then contaminate the augmentation of another. Domain generalization is understudied for multi-label remote sensing; no prior method or multisource benchmark targets it. A label-decoupled augmentation framework is therefore proposed, confining style perturbation to label-specific regions. Per-label attention, obtained from a learnable module or from gradient class-activation maps, yields per-label feature statistics; these statistics are mixed with cross-domain samples that share present labels, under independent per-label coefficients, and features are recomposed by attention-weighted normalization. Three operators combined with two attention sources produce six variants, evaluated on a leave-one-domain-out benchmark from multi-label UCM, AID, and DFC15 over six shared labels. Averaged over three splits and five seeds, the best variant attains 71.5% mean average precision, exceeding empirical risk minimization by 5.0 points and the strongest global-statistics baseline by 1.3 points, with the largest gain on the hardest transfer (up to 7.7 points). Ablations indicate that spatial attention and refreshed localization maps are most influential. The framework adds at most 0.35% parameters, leaves inference unchanged, and appears to offer a generic, inexpensive upgrade path for multilabel statistics-based domain generalization. Code is available upon acceptance at https://github.com/Alaa-Almouradi/StyleAugmentation-Upgrade. Index Terms—Attention, domain generalization, image classification, multi-label classification, remote sensing, style augmentation.
I. I NTRODUCTION Aerial and satellite scenes rarely contain a single semantic category: a typical urban tile simultaneously exhibits buildings, vehicles, vegetation, and pavement. Multi-label classification, which assigns a set of co-occurring labels to each image, is therefore a natural formulation for remote sensing (RS) image understanding and supports applications such as land-use mapping, urban monitoring, and content-based retrieval [1]– [3]. Deep convolutional networks have become the dominant approach to this task [4]–[6], building on backbone architectures pre-trained on large-scale natural-image collections [7], [8]. Manuscript received XX; revised XX. This study was supported by the Scientific and Technological Research Council of Türkiye (TUBITAK) under the Grant Number 123R108. (Corresponding author: Alaa Almouradi.) Alaa Almouradi and Erchan Aptoula are with the Faculty of Engineering and Natural Sciences, Sabancı University, 34956 Istanbul, Türkiye (e-mail: [email protected]; [email protected])
Senior Member, IEEE
These models, however, are typically trained and evaluated under the assumption of independent and identically distributed data. In practice, a classifier trained on imagery from one sensor, region, season, or ground sampling distance is frequently applied to imagery acquired under different conditions, and the resulting distribution shift degrades performance considerably [9]. Domain adaptation techniques mitigate the shift by exploiting target-domain data during training [9], [10], but such data is often unavailable before deployment. Domain generalization (DG) addresses the harder setting in which a model is trained on several source domains and must perform well on an unseen target domain [11]–[13]. Among the many DG families (invariant representation learning [14], adversarial data augmentation [15], self-supervision [16], frequencydomain augmentation [17], style-bias reduction [18], and flatminima optimization [19]) feature-statistics augmentation has proven particularly attractive; methods such as MixStyle [20], DSU [21], EFDMix [22], and correlated style uncertainty (CSU) [23] perturb instance-level feature statistics during training, are inserted into existing backbones with a few lines of code, and add essentially no inference cost. However, these augmentation methods share an implicit assumption that limits their suitability for multi-label imagery: they operate on global per-channel statistics, treating each image as carrying a single homogeneous style. This assumption is reasonable for object-centric, single-label benchmarks, but a multi-label RS tile is a mosaic of regions whose local statistics differ substantially; water bodies, rooftops, and vegetation produce markedly different channel responses. Mixing global statistics therefore entangles label-specific appearance. Consider a coastal tile in which ship, water, and pavement co-occur: a global statistics swap with a partner image can repaint the dark, low-variance signature of open water onto the hulls of the ships and the quay, and because a single mixing coefficient is shared by the whole image, the perturbation applied to the rare ship class is locked to that of the dominant pavement, so augmenting one class inevitably distorts the others. More generally, the style of a water region in one image may be transferred onto the building regions of another, and the coupled coefficient prevents the perturbation strength of co-occurring classes from being controlled independently. In addition, DG itself remains largely unexplored for multi-label RS classification. Recent surveys of DG in remote sensing [24] catalogue methods for single-label scene classification, object detection, change detection, and semantic segmentation, but none for multi-label scene classification, and no established multi-source evaluation benchmark exists for the latter; most multi-label RS studies evaluate within a single dataset [4], [6], [25], and cross-dataset robustness is rarely quantified in
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
the DG setting. The present study is the first to address DG for multi-label RS scene classification and to provide a multisource multi-label benchmark for it. To address these limitations, this paper proposes a labeldecoupled (LD) style augmentation framework that confines statistics perturbation to label-specific spatial regions. Perlabel attention maps are produced either by a lightweight learnable label-localization attention module (LLAM) or by a periodically refreshed, parameter-free Grad-CAM [26] bank. These per-label attention maps decompose each feature map into per-label statistics. These statistics are then mixed exclusively between cross-domain samples that share the corresponding label, each label receiving an independent mixing coefficient, and the perturbed components are recomposed through attention-weighted normalization. The hypothesis is that confining style transfer to matching semantic regions yields more realistic style diversity and, hence, better generalization to unseen domains; this hypothesis is examined through controlled comparisons, ablations, and an analysis of the learned style manifolds. The main contributions are as follows: 1) A label-decoupled augmentation framework comprising attention-driven per-label statistic decomposition, crossdomain label-matched partner selection, per-label mixing coefficients, and attention-weighted recomposition. The framework is instantiated on three established operators (MixStyle [20], EFDMix [22], and CSU [23]) with two attention sources, yielding six variants. This is the first domain generalization method designed specifically for multi-label RS scene classification. 2) Supporting components that make the decomposition practical: a per-label cross-domain style memory bank, a global-to-per-label warm-up schedule, an attention diversity regularizer, and a label-matched domain-balanced sampler. 3) A leave-one-domain-out multi-label DG benchmark assembled from three public multi-label RS datasets (UCM, AID, and DFC15 [2], [3], [6], [27]–[29]) over six shared labels, which is the first multi-source multi-label domain generalization benchmark in remote sensing. 4) An extensive empirical study (three leave-one-domainout splits, five seeds, eleven methods, component ablations, hyperparameter sensitivity, manifold visualization, and complexity analysis) indicating consistent improvements over the corresponding global-statistics methods at negligible cost. The remainder of this paper is organized as follows. Section II reviews related work. Section III presents the proposed framework. Section IV describes the benchmark and experimental setup, and Section V reports and discusses the results. Section VI concludes the paper. II. R ELATED WORK A. Multi-label remote sensing image classification Multi-label learning generalizes single-label classification by allowing each sample to carry a set of labels [30], [31]. In computer vision, deep multi-label models have evolved from
2
independent binary classifiers toward architectures that exploit label dependencies and spatial structure, e.g. recurrent label decoding [32], spatial regularization [33], and label graphs [34]. In RS, early deep approaches adapted convolutional networks to aerial multi-labeling [4], structured support vector machines incorporated spatial label correlations [25], and data augmentation was shown to benefit multi-label landcover categorization [35]. In parallel, attention has played an increasingly central role: class-wise attention combined with recurrent modeling [3], multi-attention feature extraction [5], and attention-driven unmanned aerial vehicle (UAV) image multi-labeling [36] all localize label-relevant evidence, while relation networks [6] and graph convolutional models [37] encode inter-label dependencies. Benchmarks such as multilabel UCM [2], [27], multi-label AID [6], [28], and the DFC15 multi-label set [3], [29] have standardized within-dataset evaluation. Nevertheless, these studies almost exclusively train and test on the same distribution; the behavior of multi-label RS classifiers under domain shift has received little systematic attention in the DG setting, which motivates the benchmark used in this work. B. Domain generalization DG aims to learn, from one or several source domains, a predictor that transfers to unseen target domains without access to target data during training [11]. Comprehensive surveys categorize the field into representation-based, dataaugmentation-based, and learning-strategy-based approaches [12], [13]. Representative directions include domain-invariant component analysis [14], adversarial feature alignment originally developed for adaptation [10], adversarial data augmentation [15], self-supervised auxiliary tasks [16], Fourierbased amplitude mixing [17], style-bias reduction [18], and stochastic weight averaging densely (SWAD), which seeks flat minima of the loss landscape [19]. Most of these methods are single-label by construction. Multi-label DG, in which each image carries several cooccurring labels, has received far less attention, and the few existing approaches are dominated by vision-language models, which operate on a different set of constraints. CLIPood adapts a pretrained contrastive language–image model to outof-distribution data through a margin-based objective and a Beta-weighted parameter average [38], and Mixup-CLIPood extends this idea with a class-aware mix-up loss over larger language–image backbones [39]. Such methods inherit the broad coverage of web-scale pretraining and the multi-modal advantage, but they depend on heavy foundation backbones, are designed for generic object imagery rather than overhead scenes, and still characterize style at the image level rather than per label. Within remote sensing, distribution shift has historically been tackled by domain adaptation, which assumes access to (typically unlabeled) target data [9]. RS DG has gained traction only recently; a dedicated survey [24] organizes RS DG methods into distribution alignment, data augmentation, and representation learning across scene classification, object detection, change detection, and semantic segmentation. Style
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
and feature augmentation is, in fact, a recurring tool in this literature: SODA randomizes styles with adaptive instance normalization [40], FOSMix mixes style in the frequency domain for RS segmentation [41], and a style-content separation network disentangles style from content for cross-scene generalization [42]. Other lines align spectral bands across countries [43], collaborate across multiple sources [44], generalize object detectors [45], guide invariance with language [46], or fine-tune geospatial vision foundation models [47]. Crucially, every one of these methods operates on single-label scene classification, detection, change detection, segmentation, or pixel-wise hyperspectral classification; the multi-label sceneclassification setting, in which several classes with distinct local styles share a single image, is addressed by none of them, and the benchmarks they adopt are correspondingly singlelabel [24]. These observations expose a specific gap. Global featurestatistics augmentation is cheap and effective but is single-style per image by construction; multi-label DG techniques exist but are vision-language and not tailored to overhead imagery; and RS DG, despite using style augmentation extensively, has not addressed multi-label scene classification. The framework proposed here targets precisely this intersection by decoupling style statistics per label, which is the first such treatment for multi-label RS DG. C. Feature-statistics style augmentation Instance-level feature statistics encode image style: adaptive instance normalization (AdaIN) transfers style by replacing the per-channel mean and standard deviation of one image’s features with those of another [48]. Building on this observation, MixStyle interpolates the feature statistics of random sample pairs during training [20]; DSU models statistics as Gaussian random variables and resamples them [21]; EFDMix matches higher-order statistics exactly through sorted-value interpolation [22]; and CSU perturbs statistics with channelcorrelated uncertainty [23]. These methods are attractive for their simplicity, negligible cost, and strong results on singlelabel DG benchmarks. However, all of them compute statistics by pooling over the entire spatial extent of the feature map, which presumes one dominant style per image, and all of them draw a single mixing coefficient per image, which couples the perturbation strength of every class in the scene. Both design choices conflict with the structure of multi-label aerial imagery, in which several classes with distinct local statistics co-occur. The framework proposed in this paper removes both restrictions by decomposing statistics per label with spatial attention, obtained from a learned module or from class activation maps [26], [49], and by mixing each label’s statistics independently. III. P ROPOSED APPROACH A. Overview and problem formulation Let D1 , . . . , DS denote S source domains that share a label vocabulary of L classes. Each training sample is a triplet (xi , yi , di ), where xi ∈ R3×H0 ×W0 is an image, yi ∈ {0, 1}L is its multi-label annotation, and di ∈ {1, . . . , S} is its
3
domain index. The goal is to learn a multi-label classifier g : R3×H0 ×W0 → [0, 1]L that performs well on an unseen target domain. Fig. 1 summarizes the proposed framework. A standard backbone is augmented with LD modules inserted after each layer in a set K of selected layers. During training, each module (i) estimates per-label spatial attention, (ii) pools perlabel feature statistics, (iii) selects a cross-domain partner that shares at least one label (or draws stored statistics from a per-label style bank), (iv) perturbs each label’s statistics independently with one of three operators, and (v) recomposes the feature map through attention-weighted normalization. The modules are stochastic and active only during training; at inference they reduce to the identity, so the deployed network is exactly the original backbone. B. Per-label decomposition and recomposition Consider the feature map Fi ∈ RC×H×W of sample i at an insertion point, and let Ai ∈ [0, 1]L×H×W collect per-label attention maps (Section III-C), with Aiℓ = 0 for absent labels (yiℓ = 0). Normalized spatial weights are obtained as Aiℓhw , PW ′ ′ h′ =1 w′ =1 Aiℓh w + ε
Āiℓhw = PH
(1)
where ε is a small constant for numerical stability. The attention-weighted first and second moments of each present label ℓ are X µiℓc = Āiℓhw Fichw , h,w
σiℓc =
X
2 Āiℓhw Fichw − µ2iℓc + ε
1/2 ,
(2)
h,w
yielding per-label style vectors µiℓ , σ iℓ ∈ RC . Given perturbed statistics (µ̂iℓ , σ̂ iℓ ) produced by one of the operators in Section III-E, the augmented feature map is recomposed as a convex spatial combination of per-label stylized components, F̂i =
L X Aiℓ ℓ=1
Fi − µiℓ ⊙ σ̂ iℓ + µ̂iℓ , Zi σ iℓ
Zi =
L X ℓ=1
Aiℓ + ε,
(3) where ⊙ denotes element-wise multiplication with broadcasting of the channel-wise statistics over spatial locations. Equation (3) performs, per label, the normalization–renormalization of AdaIN [48], but blends the results spatially according to the attention, so that each region is restyled predominantly by the statistics of the labels that occupy it. Locations not covered by any present label are dominated by the residual normalization term and remain close to the original features; labels whose statistics are left unperturbed (Section III-D) contribute identity transformations. Each per-label branch is a channel-wise affine map of the standardized features, i.e. it preserves the spatial geometry of Fi and only rescales and shifts channels; the convex, attention-weighted combination of such branches therefore yields a well-defined feature tensor that stays on the same manifold as a single AdaIN restyling, rather than an arbitrary distortion that the network could simply learn to ignore.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
4
source domains
ResNet-50 backbone domain-balanced, label-matched batch
1
(xi , yi , di )
2
stem & Layer 1
LD
Layers 3&4
GAP
BCE + wdiv div
FC
multi-label loss
Inside one LD module
y
F ∈ ℝC×H×W
Layer 2
LD
per-label attention Aℓ
per-label statistics
cross-domain partner
per-label augmentation
recomposition
dj ≠ di
i A1
A2
A3
LLAM (learned) or Grad-CAM bank (cached) zero for absent labels
j
μ1 , σ1 μ2 , σ2 μ3 , σ3
attention-weighted pooling of F (μiℓ , σiℓ ) per label
λiℓ ∼ Beta(α, α)
≥ 1 shared label
batch pairing or per-label style bank
F ̂ = ∑AZℓ Fℓ̂ ℓ
on shared labels; identity elsewhere
back to the backbone
Fig. 1. Overview of the proposed LD style augmentation framework. Top: training pipeline with LD modules inserted after selected backbone layers K; the modules are active only during training and reduce to the identity at inference. Bottom: inside one LD module, per-label attention (learned LLAM or cached Grad-CAM) yields per-label statistics, which are mixed only with cross-domain partners sharing the corresponding label (through batch pairing or a per-label style bank), perturbed with independent per-label coefficients, and recomposed by attention-weighted normalization.
C. Per-label attention Two interchangeable attention sources are considered. a) Learned attention using LLAM: The label-localization attention module is a lightweight head consisting of two 1 × 1 convolutions with an intermediate rectified linear unit, mapping Fi to logits zi ∈ RL×H×W , followed by a temperaturesharpened spatial softmax and masking by the ground-truth label vector: exp τ ziℓhw , Aiℓhw = yiℓ P (4) ′ ′ h′ ,w′ exp τ ziℓh w where τ ≥ 1 controls map sharpness. Because co-occurring labels should attend to different regions, a diversity regularizer penalizes the mean pairwise cosine similarity between the attention maps of labels that are jointly present: B
Ldiv =
1 X 1 B i=1 |Pi |
X (ℓ,m)∈Pi
⟨aiℓ , aim ⟩ , ∥aiℓ ∥2 ∥aim ∥2
(5)
where B is the batch size, aiℓ ∈ RHW is the flattened map of label ℓ, and Pi = {(ℓ, m) : ℓ < m, yiℓ = yim = 1}. In the configuration of Section V, the module adds 8.3 × 104 parameters (+0.35%).
b) Cached Grad-CAM bank (GC): As a parameter-free alternative, class activation maps [26], [49] of the network’s own predictions are used. For label ℓ of sample i, the map is C X (4) Acam = ReLU w F iℓc ic , iℓ c=1
1 X ∂siℓ wiℓc = , (4) HW ∂F h,w
(4) Fi
(6)
ichw
where is the final-stage feature map and siℓ the presigmoid logit of label ℓ. Computing (6) online at every step would be costly and, early in training, unreliable. Instead, a bank of maps is built once the warm-up ends and is refreshed every R epochs by a single sweep over the training set; an entry is stored only if label ℓ is present and currently predicted correctly, is kept in half precision on the host, and is bilinearly resized and renormalized to the resolution of each insertion point when used. Missing entries (e.g. notyet-correct labels) fall back to the global pathway for that label. This correctness-gated caching is also what guards the decomposition against the well-known noisiness of class activation maps and their sensitivity to distribution shift: a map is admitted only where the network already localizes the label well on a source image, maps are refreshed as the model
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
5
improves, and any label without a reliable entry reverts to the global operator, so a poorly localized map degrades gracefully toward the global baseline rather than injecting cross-class contamination. Because the per-label components are finally recombined by attention-weighted normalization (3), residual localization error is further attenuated by the spatial blending. The learned LLAM source offers a complementary safeguard, as its maps are optimized end-to-end together with the mixing. The two attention sources trade off a small parameter and compute overhead with a learned, task-adaptive map (LLAM) against zero parameters with a periodic refresh cost (GC). D. Cross-domain label-matched pairing and style bank Global-statistics methods mix arbitrary sample pairs. Here, mixing partners are constrained to differ in domain and to share semantic content: sample i is paired with a partner j = π(i) drawn from P(i) = j : dj ̸= di and yi⊤ yj ≥ 1 , (7) computed within each mini-batch by a greedy matching over a domain-balanced sampler that explicitly co-locates labelcompatible samples from different domains. The per-label shared mask and the effective mixing coefficient are miℓ = yiℓ yjℓ ,
λ̃iℓ = miℓ λiℓ + (1 − miℓ ),
(8)
where λiℓ ∼ Beta(α, α), so that labels not shared with the partner receive λ̃iℓ = 1, i.e. the identity, and every shared label draws an independent coefficient. For the interpolation operator, partners may alternatively be drawn from a perlabel first-in-first-out style bank of size M that stores recent (µ·ℓ , σ ·ℓ , d) triplets and is sampled under the same crossdomain constraint, decoupling the diversity of styles from the mini-batch composition; when a label has no cross-domain entry yet, batch pairing is used as a fallback. E. Per-label augmentation operators Three operators, illustrated in Fig. 2, instantiate the framework; each is applied per present label with the coefficient λ̃iℓ of (8). a) LD-MixStyle: Following the interpolation principle of MixStyle [20], the perturbed statistics are convex combinations of the per-label statistics of the two samples: µ̂iℓ = λ̃iℓ µiℓ + (1 − λ̃iℓ )µjℓ , σ̂ iℓ = λ̃iℓ σ iℓ + (1 − λ̃iℓ )σ jℓ ,
(9)
These values are inserted into the recomposition (3). The partner statistics come from the paired sample or from the style bank. b) LD-EFDMix: Moment interpolation captures only first- and second-order statistics; exact feature distribution matching [22] instead operates on sorted feature values. Let Tiℓ contain the K = ⌊ρHW ⌋ spatial locations with the largest c attention Aiℓhw , and let viℓ = sort {Fichw : (h, w) ∈ Tiℓ } denote the sorted values of channel c at those locations (analogously for the partner). The mixed sequence c c c v̂iℓ = λ̃iℓ viℓ + (1 − λ̃iℓ ) vjℓ
(10)
is scattered back to the original locations of sample i in rank order, the partner term being treated as a constant (gradientstopped) as in [22]; the resulting per-label value changes are then blended spatially with the attention weights, analogously to (3). Restricting the matching to the top-K locations of both attention maps confines the exchanged distribution to labelrelevant evidence. c) LD-CSU: Inspired by correlated style uncertainty [23], this operator perturbs each label’s statistics without any partner. With ξ ∼ N (0, IC ) and η ∼ N (0, 1), εdiℓ = σ iℓ ⊙ ξ,
εciℓ = σ̄iℓ η, µ̃iℓ = µiℓ + β εdiℓ + εciℓ 1 , σ̃ iℓ = max σ iℓ + β |εdiℓ |, ε ,
(11)
where σ̄iℓ is the channel-mean standard deviation, 1 ∈ RC is a vector of ones, and β is the perturbation strength. The scalar term εciℓ displaces all channels jointly, injecting the channelcorrelated component advocated in [23] at the per-label level. Combining the three operators with the two attention sources of Section III-C yields six variants, denoted LDMixStyle, LD-EFDMix, LD-CSU (learned attention) and LDMixStyle-GC, LD-EFDMix-GC, LD-CSU-GC (Grad-CAM bank). F. Training procedure Each LD module fires stochastically with probability p per batch, as in [20]. Reliable per-label attention is unavailable at the start of training: LLAM logits are untrained and network predictions, on which the Grad-CAM bank conditions, are still poor. Training therefore begins in a global mode in which the modules behave exactly like their global-statistics counterparts. For LLAM variants, the per-label pathway is blended in linearly, Fout = we F̂LD + (1 − we ) F̂glob , i i i we = min max (e − W )/W, 0 , 1 ,
(12)
where e is the epoch index and W is the warm-up length; GC variants switch to the per-label pathway as soon as the first bank has been built after epoch W . The total objective is L = LBCE + wdiv Ldiv ,
(13)
where LBCE is the mean binary cross-entropy over labels and wdiv = 0 for GC variants, whose attention is not learned. At inference all modules are inactive, so the computational graph is identical to the plain backbone. IV. E XPERIMENTAL SETUP A. Datasets and splits The benchmark is assembled from three public multi-label aerial datasets whose label vocabularies overlap in six classes: building, car, tree, water, pavement, and ship. UCM [27] with the multi-label annotations of [2] contains 2100 images of 256 × 256 pixels at approximately 0.3 m ground sampling distance (GSD). AID [28] with the multi-label annotations of [6] contains 3000 images of 600 × 600 pixels collected from multiple sensors at GSDs ranging from about 0.5 m to
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
6
(a) LD-MixStyle
(b) LD-EFDMix
(c) LD-CSU
moment interpolation
top-K sort matching
correlated perturbation
σℓ
σℓ
(μjℓ , σjℓ )
̃ ) (μ̃iℓ , σiℓ
sorted top-K (i)
(μ̂iℓ , σ̂iℓ ) (μiℓ , σiℓ )
sorted top-K (j)
̃ , mix with λiℓ
μℓ
scatter back at iℓ
̃ μiℓ + (1 − λiℓ ̃ ) μjℓ μ̂iℓ = λiℓ per-label AdaIN — needs partner j
K = ⌊ρHW⌋ — needs partner j
(μiℓ , σiℓ ) μℓ c μ̃iℓ = μiℓ + β (εdiℓ + εiℓ )
self-perturbation — no partner needed
Fig. 2. The three per-label augmentation operators explained in subsection III-E. (a) LD-MixStyle interpolates the per-label moments of sample i toward those of a cross-domain partner j sharing label ℓ. (b) LD-EFDMix sorts the feature values at the top-K attention locations of both samples and mixes them in rank order. (c) LD-CSU perturbs the per-label moments with channel-correlated Gaussian noise and requires no partner.
TABLE I B ENCHMARK STATISTICS OVER THE SIX SHARED LABELS . C OUNTS ARE NUMBERS OF POSITIVE IMAGES PER LABEL ; GSD DENOTES GROUND SAMPLING DISTANCE IN METERS .
validation subsets are used for learning, and the target domain is evaluated on its held-out test subset only. B. Evaluation metrics
Images Tile size (px) GSD (m) Labels per image building car tree water pavement ship
UCM [2], [27]
AID [6], [28]
DFC15 [3], [29]
2100 256 × 256 ≈ 0.3 2.00
3000 600 × 600 ≈ 0.5–8 3.35
3342 600 × 600 0.05 1.90
691 886 1009 203 1300 102
2161 2026 2406 852 2328 284
1001 705 258 998 3133 270
Performance is measured with threshold-free and thresholded multi-label metrics [31]: per-label average precision and its mean (mAP), the macro-averaged per-class F1 score (CF1), and the micro-averaged overall F1 score (OF1). For CF1 and OF1, a per-label decision threshold is selected on the source validation sets by maximizing the per-label F1 and is then kept fixed for target evaluation, so no target information enters threshold selection. All main results are reported as the mean and standard deviation over five seeds, which control data splits, initialization, and augmentation randomness. C. Compared methods
8 m. The DFC15 set originates from the 2015 IEEE GRSS Data Fusion Contest over Zeebrugge [29] and comprises 3342 tiles of 600 × 600 pixels at 0.05 m GSD with the multi-label annotations of [3]. Vocabularies are harmonized by mapping impervious (DFC15) to pavement and boat to ship; labels outside the shared vocabulary are discarded. Table I summarizes the resulting per-domain statistics. The three domains differ markedly in GSD, sensor characteristics, geographic region, and label priors, e.g. tree is frequent in UCM and AID but rare in DFC15, which makes the benchmark a demanding test of multi-label generalization. Fig. 3 presents the shared classes from all domains in order to better see the variations across domains. Three leave-one-domain-out splits are defined: Split 1 trains on {AID, DFC15} and tests on UCM; Split 2 trains on {UCM, DFC15} and tests on AID; Split 3 trains on {UCM, AID} and tests on DFC15. Each domain is partitioned into training, validation, and test subsets (80/10/10%) with a seeddependent split shared by all methods; source training and
The six LD variants are compared against empirical risk minimization (ERM) with binary cross-entropy; the three corresponding global-statistics methods: MixStyle [20], EFDMix [22], and CSU [23], all inserted at the same backbone stages; and SWAD [19] as a representative optimization-based DG method. An oracle trained and evaluated within the target domain (its training/test partitions) indicates the within-domain ceiling. All methods share the same backbone, optimizer, schedule, data augmentations, and evaluation pipeline; they differ only in the components under study. D. Implementation details All experiments use a ResNet-50 backbone [7] pre-trained on ImageNet [8], with inputs resized to 224 × 224 pixels and augmented with horizontal/vertical flips and color jittering. Models are trained with AdamW [50] (backbone learning rate 1.16 × 10−4 , classifier-head learning rate 9.39 × 10−4 , weight decay 1.59 × 10−5 ), a cosine schedule decaying to
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
Car
Tree
Water
Pavement
Ship
DFC15 0.05 m
AID 0.5--8 m
UCM ≈ 0.3 m
Building
7
Fig. 3. Class samples from each domain in the benchmark, where the domain shift can be seen through the variations between images of the same class from different domains.
10−6 , gradient-norm clipping at 1.0, and a batch size of 32 drawn by the domain-balanced label-matched sampler of Section III-D. Training runs for at most 100 epochs with early stopping (patience 15) on the source-validation mAP. The supplementary training cost of the LD modules is analyzed in Section V-E. E. Hyperparameter selection Hyperparameters are chosen with a leave-one-source-out (LOSO) proxy objective in the spirit of [51]: each source domain is held out in turn, the model is trained on the remaining sources, and the average held-out validation mAP constitutes the search objective. With only 2 source domains available per split, LOSO was constructed by pairing the remaining source domain with a heavily augmented version of itself as a pseudo domain. Searches are conducted with the tree-structured Parzen estimator of Optuna [52] (multivariate mode, median pruning): 40 trials for the shared optimizer parameters, reused by all methods, and 20 trials per method for the method-specific parameters listed in Table II. For some LD variants, the LOSO-selected values were further adjusted to push for more aggressive operations. Table III lists the selected values of the headline variant. V. R ESULTS AND DISCUSSION A. Comparison with global-statistics augmentation Tables IV–VI report per-label average precision, mAP, CF1, and OF1 for the three splits, and Table VII averages the mAP results. Several observations can be discussed.
TABLE II H YPERPARAMETER SEARCH SPACES . “L OG ” DENOTES LOG - UNIFORM SAMPLING ; U DENOTES UNIFORM SAMPLING .
Parameter
Range
Applies to
[10−6 , 5 × 10−4 ], log
backbone learning rate All head learning rate [10−5 , 5 × 10−2 ], log All weight decay [10−6 , 5 × 10−1 ], log All α (Beta Dist.)
U [0.1, 5.0]
gate probability p insertion stages K
U [0.2, 0.8] subsets of {1, . . . , 4}
(LD-)MixStyle & (LD-)EFDMix All augmentations LD variants
warm-up W (epochs) diversity weight wdiv temperature τ bank size M top-K ratio ρ strength β refresh R (epochs)
{0, . . . , 10} [0.01, 10], log [0.1, 100], log {64, 128, 256, 512} U [0.01, 1.0] U [0.1, 1.0] {1, . . . , 15}
All LD variants LLAM variants LLAM variants LD-MixStyle LD-EFDMix (LD-)CSU GC variants
First, statistics-based augmentation is clearly beneficial in this setting: the global methods improve the average mAP over ERM by 2.5–3.7 points, whereas SWAD performs on par with ERM (66.2 vs. 66.5), suggesting that flat-minima optimization alone does not address the dominant appearance shift between these aerial domains. Second, label decoupling improves every operator it is applied to. With learned attention, LD-MixStyle, LD-EFDMix, and LD-CSU exceed their global counterparts by +1.5, +2.4,
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
TABLE III S ELECTED HYPERPARAMETERS OF LD-EFDM IX -GC PER SPLIT (S1–S3).
Parameter
S1
S2
S3
insertion stages K α (Beta) gate probability p top-K ratio ρ refresh R (epochs) warm-up W (epochs)
{2} 0.45 0.20 0.34 9 9
{1} 5.0 0.80 0.60 13 5
{1} 5.0 0.80 0.60 13 5
and +1.1 mAP points on average, respectively; with the GradCAM bank, LD-EFDMix-GC gains +2.5 over EFDMix and attains the best overall average of 71.5% (+5.0 over ERM, +1.3 over the strongest global baseline, CSU). The gains are most pronounced on Split 3, the hardest transfer (target DFC15 at 0.05 m GSD): ERM drops to 50.1%, while LDCSU reaches 57.8% (+7.7) and LD-EFDMix-GC and LDMixStyle reach 57.3%. The advantage of decoupling also appears at the label level: on Split 2, ERM and the global methods nearly collapse on the minority class ship (27–37% average precision), whereas the LLAM-based LD variants recover 39–43%, consistent with the premise that restricting style exchange to label-matched regions protects rare classes from being overwritten by dominant ones. Third, the benefit of the cached attention source is operatordependent. For EFDMix, whose sort matching directly consumes the attention-selected locations, the Grad-CAM bank is the strongest configuration on Split 1 (79.5%) and on average; for MixStyle and CSU, the cached maps help less than the jointly trained LLAM (+0.4 and +0.2 average gain over the global counterparts, against +1.5 and +1.1 with LLAM). A plausible explanation is that moment pooling integrates over the whole map, so it benefits more from attention that is optimized end-to-end together with the mixing, whereas rank matching mainly requires that the support of the map be correct, which the Grad-CAM bank provides cheaply. Fourth, Split 2 exhibits the smallest margins for all augmentation methods (76.8–79.1% around an ERM of 77.5%). AID is itself a multi-sensor, multi-resolution collection [28], so its style distribution overlaps the sources more than in the other splits, leaving less room for style-based augmentation. Finally, the gap to the within-domain oracle remains large (13–40 mAP points depending on the split), indicating that style augmentation alone, whether global or label-decoupled, closes only part of the domain gap; the limitation analysis in Section V-F returns to this point. This is expected: domain shift between these datasets is not purely a matter of style but also reflects differences in ground sampling distance, sensor and band characteristics, and class co-occurrence statistics (Table I), which channel-statistics operators are not designed to correct. A style-centric method is nonetheless the appropriate tool for the appearance component of the gap that the AdaIN family does model, and the manifold analysis of Section V-D indicates that the per-label style component is the part that aligns across domains; the residual, structural component is what the distance to the oracle reflects.
8
B. Ablation study Table VIII removes the components of LD-EFDMix-GC one at a time on Split 1, and additionally evaluates LD-MixStyle without its style bank. Replacing the per-label attention by a uniform map (which collapses the decomposition toward global statistics) costs 2.6 mAP points, and freezing the Grad-CAM bank after its first construction costs 3.3 points, indicating that spatially correct and up-to-date localization is the core of the method. Disabling the top-K restriction (ρ = 1) costs 1.6 points, supporting the view that sort matching should be confined to label-relevant locations. The remaining components contribute smaller but consistent amounts: perlabel coefficients (−0.8), label-matched pairing (−0.6), and the warm-up (−0.5). The style bank of LD-MixStyle yields a modest +0.3 over batch-only pairing, suggesting that the domain-balanced label-matched sampler already supplies sufficient cross-domain partners in this benchmark. C. Hyperparameter sensitivity Fig. 4 examines the main hyperparameters on Split 1 with single training runs at a fixed seed different from any of the main 5 seeds. The insertion point is the most influential choice: a single module after stage 2 performs best (83.6% in this run), early-plus-deep combinations underperform, and the all-stage configuration lies in between; consistent with the common observation that style information concentrates in early-to-middle layers [20], [21]. The top-K ratio favors selective matching: the smallest tested value (ρ = 0.2) performs best, and quality degrades toward ρ = 1, in line with the corresponding ablation. The refresh interval shows no consistent trend over R ∈ [5, 15]; the spread (76–82%) is comparable to the seed-to-seed variability of the full model (±2.4), suggesting that any sufficiently frequent refresh is adequate, while never refreshing is harmful (Table VIII). The LLAM temperature exhibits a broad plateau for τ ∈ [1, 50] with a mild decline at τ = 100, where overly sharp maps may discard context. The bank size likewise shows no consistent trend within run-to-run variability, matching its small ablation effect. Overall, apart from the insertion point, the method appears robust to its hyperparameters around the selected values. D. Style-manifold analysis Fig. 5 visualizes, with t-SNE [53], the style statistics that the two augmentation philosophies operate on, using an LDEFDMix model from Split 1 and the test partitions of all three domains. When global per-image statistics are embedded (Fig. 5a), points organize primarily by domain, and images carrying different labels from the same domain remain close together: the quantity that global methods mix is dominated by sensor- and scene-level appearance. The per-label statistics extracted by the proposed decomposition (Fig. 5b) instead form label-coherent groupings within which the three domain markers interleave, and target-domain points fall inside or near the groups spanned by the sources. This pattern supports the central premise of the method: per-label pooling exposes labelspecific style components that are comparable across domains,
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
9
TABLE IV S PLIT 1 ( SOURCES : AID, DFC15; TARGET: UCM). P ER - LABEL AVERAGE PRECISION AND AGGREGATE METRICS IN % ( MEAN±STD OVER FIVE SEEDS ; PER - LABEL COLUMNS SHOW MEANS ). B EST VALUE PER COLUMN AMONG THE DOMAIN GENERALIZATION METHODS IN BOLD ; THE WITHIN - DOMAIN ORACLE IS REPORTED FOR REFERENCE ONLY. Method
building
car
tree
water
pavement
ship
mAP
CF1
OF1
ERM (BCE)
58.2
71.9
83.4
45.9
84.4
MixStyle (ICLR 2021) [20] EFDMix (CVPR 2022) [22] CSU (WACV 2024) [23] SWAD (NeurIPS 2021) [19]
68.7 68.1 70.6 56.3
77.7 73.8 76.6 71.3
81.8 85.0 83.2 85.0
62.9 63.6 52.2 47.5
87.5 85.3 85.7 84.3
87.9
72.0±1.2
69.5±1.5
73.1±1.4
93.5 93.4 88.3 87.4
78.7±1.2 78.2±4.3 76.1±1.5 72.0±1.8
73.0±3.0 73.0±3.4 71.2±1.0 69.3±2.2
74.5±2.2 74.6±1.7 73.7±1.8 72.4±1.9
LD-MixStyle LD-EFDMix LD-CSU LD-MixStyle-GC LD-EFDMix-GC LD-CSU-GC
67.8 69.7 67.9 68.0 71.8 69.0
72.4 76.5 78.4 74.3 79.0 77.3
85.4 86.9 85.2 85.6 84.9 85.3
65.3 61.8 55.2 61.2 61.0 49.4
84.0 83.8 85.2 84.4 83.5 85.1
90.4 92.4 95.2 96.2 96.6 91.9
77.6±3.6 78.5±2.3 77.9±1.8 78.3±2.9 79.5±2.4 76.3±3.3
73.5±2.6 73.8±1.6 73.1±1.2 73.7±2.0 74.1±2.4 72.1±1.9
74.6±1.9 75.0±1.1 74.9±1.9 75.0±2.5 74.9±2.2 74.0±0.8
Oracle (within-domain)
93.8
95.6
95.7
98.5
96.5
99.8
96.7±0.3
90.1±0.6
88.6±0.5
TABLE V S PLIT 2 ( SOURCES : UCM, DFC15; TARGET: AID). N OTATION AS IN TABLE IV.
Method
building
car
tree
water
pavement
ship
mAP
CF1
OF1
ERM (BCE)
91.6
92.2
92.6
62.5
93.7
32.4
77.5±2.1
55.7±4.1
67.2±3.8
MixStyle [20] EFDMix [22] CSU [23] SWAD [19]
93.0 93.1 93.1 91.5
92.4 89.1 92.7 92.6
93.6 93.1 92.8 92.8
61.9 58.5 63.1 63.5
92.3 91.1 95.6 93.7
27.4 36.6 32.1 29.9
76.8±1.6 76.9±1.8 78.2±1.0 77.3±2.1
57.3±1.7 56.8±6.1 60.2±2.8 54.6±3.8
70.8±1.1 69.1±3.8 70.1±2.9 66.7±3.9
LD-MixStyle LD-EFDMix LD-CSU LD-MixStyle-GC LD-EFDMix-GC LD-CSU-GC
91.3 94.5 92.8 95.2 93.6 93.9
91.7 92.5 92.5 92.0 92.9 93.1
91.5 92.5 91.4 91.7 93.2 92.9
58.3 61.7 59.6 54.5 60.9 58.6
93.5 92.6 94.7 94.6 90.8 93.3
43.4 40.8 39.3 36.0 36.0 38.3
78.3±1.6 79.1±1.6 78.4±1.6 77.3±1.6 77.9±2.2 78.4±1.3
57.3±5.5 56.0±2.0 54.5±4.5 55.6±3.7 58.0±4.1 55.1±2.2
66.4±3.8 69.1±1.5 66.4±3.6 70.1±1.7 69.3±5.4 67.5±2.8
Oracle (within-domain)
99.3
98.3
98.9
82.5
99.5
76.5
92.5±0.4
86.5±1.5
92.8±0.8
TABLE VI S PLIT 3 ( SOURCES : UCM, AID; TARGET: DFC15). N OTATION AS IN TABLE IV.
Method
building
car
tree
water
pavement
ship
mAP
CF1
OF1
ERM (BCE)
43.1
42.3
13.5
62.1
96.1
43.7
50.1±6.2
46.2±4.2
60.4±2.3
MixStyle [20] EFDMix [22] CSU [23] SWAD [19]
46.4 48.5 51.0 42.1
36.3 42.8 41.3 40.0
25.4 17.3 24.3 11.0
58.5 55.9 63.5 60.6
97.8 98.0 98.2 96.9
54.7 48.8 59.7 44.6
53.2±4.0 51.9±3.8 56.3±2.6 49.2±5.6
47.0±3.8 46.8±2.0 49.4±1.9 45.9±5.6
61.9±1.8 63.4±2.1 64.2±1.9 60.6±2.7
LD-MixStyle LD-EFDMix LD-CSU LD-MixStyle-GC LD-EFDMix-GC LD-CSU-GC
60.0 55.9 45.3 51.8 51.7 54.2
48.6 44.1 47.8 42.8 52.0 43.4
23.4 21.9 30.7 17.1 21.9 20.4
61.4 65.2 69.4 61.3 61.8 66.2
97.4 98.1 98.2 98.3 98.4 97.8
52.7 54.7 55.4 54.0 57.8 57.7
57.3±3.8 56.7±3.1 57.8±2.5 54.2±4.3 57.3±5.1 56.6±2.6
49.4±3.2 47.9±3.3 50.9±2.3 49.5±3.1 50.8±3.2 48.3±2.7
64.0±1.7 64.6±2.2 65.9±2.1 65.6±2.5 65.5±1.0 62.7±2.9
Oracle (within-domain)
98.6
97.7
92.8
99.1
100.0
94.8
97.2±1.1
91.9±2.0
95.6±0.6
which is precisely the structure that label-matched mixing exploits. E. Complexity analysis Table IX compares model complexity. The LLAM variants add 8.3 × 104 parameters (+0.35%) for the configuration with
modules after stages 1 and 2; the GC variants add none. Because all modules are identities at inference, the multiply– accumulate count is unchanged (4.087 G at 224 × 224 pixels) and the measured evaluation throughput is statistically indistinguishable across methods (1495–1505 images/s at batch size 32 on the GPU used in this study). Training-time overheads are
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
83.6 81.6
82
mAP (%)
mAP (%)
84
79.8
80
78.7
78.4
78 76
86
86
84
84
82
82
mAP (%)
86
10
80 78 76
74
{1}
{2}
{1, 2}
74
{2, 3} {1, 2, 3}
Insertion stages
0.2
0.4
0.6
74
0.8
Top-K ratio ρ
(b) Top-K ratio
86
86
84
84
82
82
mAP (%)
mAP (%)
78 76
(a) Insertion point
80 78 76 74
80
5
7
9
11
13
Refresh interval R (epochs)
15
(c) Map refresh interval
80 78 76
0.1
0.5
1
5
10
50
Attention temperature τ
100
74
32
(d) LLAM temperature
64
128
Style-bank size M
256
512
(e) Style-bank size
Fig. 4. Hyperparameter sensitivity on Split 1 (target: UCM); each point is a single training run at a fixed seed shared by all sensitivity runs and different from any training run seed. (a)–(c) LD-EFDMix-GC: insertion stages, top-K ratio ρ, and Grad-CAM refresh interval R; (d) LD-EFDMix: LLAM temperature τ ; (e) LD-MixStyle: style-bank size M . Dashed lines show the five-seed mean of the training configuration. The refresh and top-K sweeps use a one-epoch warm-up so that the studied mechanism governs most of the run.
TABLE VII TARGET- DOMAIN M AP (%) AVERAGE OVER SPLITS , AND THE DIFFERENCE OF THAT AVERAGE WITH RESPECT TO ERM. B EST VALUE PER COLUMN AMONG THE DOMAIN GENERALIZATION METHODS IN BOLD . Average
∆ vs. ERM
ERM (BCE)
66.5
–
MixStyle [20] EFDMix [22] CSU [23] SWAD [19]
69.5 69.0 70.2 66.2
+3.0 +2.5 +3.7 −0.4
LD-MixStyle LD-EFDMix LD-CSU LD-MixStyle-GC LD-EFDMix-GC LD-CSU-GC
71.0 71.4 71.3 69.9 71.5 70.4
+4.5 +4.9 +4.8 +3.4 +5.0 +3.9
Oracle (within-domain)
95.4
+28.9
Method
confined to the augmentation pathway: LLAM adds two 1 × 1 convolutions per inserted stage and the diversity loss, while the GC variants add approximately one extra forward–backward sweep over the training set every R epochs to rebuild the attention bank. These properties make the framework a near
drop-in replacement for its global counterparts. F. Limitations Several limitations should be noted. First, the gap to the within-domain oracle remains substantial (13–40 mAP points), and minority classes with strong cross-domain appearance variance (ship on AID and tree on DFC15) remain difficult for every method evaluated: augmenting feature statistics cannot synthesize object scales or contexts absent from the sources. Second, gains on Split 2 are small, indicating that the benefit of style augmentation shrinks when the target’s style distribution already overlaps the sources. Third, the benchmark covers six shared labels and three domains; transfer of the conclusions to larger vocabularies and other sensor families remains to be verified due to lack of similar annotated multi-label multidomain data in the literature. Fourth, the sensitivity study relies on single runs per configuration, so only coarse trends should be read from it. Fifth, all methods are compared under a common ResNet-50 backbone with matched training, but the study does not include large vision-language or foundationmodel baselines; in general vision, multi-label DG is increasingly driven by such models (e.g. CLIPood [38] and MixupCLIPood [39]), and geospatial vision foundation models have
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
11
TABLE VIII C OMPONENT ABLATIONS ON S PLIT 1 (%, MEAN±STD OVER FIVE SEEDS ). ∆M AP IS THE CHANGE WITH RESPECT TO THE CORRESPONDING FULL MODEL . mAP
CF1
OF1
∆mAP
LD-EFDMix-GC (full) w/o attention (uniform map) w/o map refresh w/o top-K selection (ρ = 1) w/o per-label λ (shared λ) w/o label-matched pairing w/o warm-up
79.5±2.4 76.8±3.0 76.2±3.5 77.8±2.9 78.6±2.4 78.8±1.4 79.0±2.0
74.1±2.4 72.5±0.8 72.6±1.2 71.8±1.1 73.7±2.5 74.1±2.2 73.2±1.7
74.9±2.2 74.2±2.3 74.4±1.3 73.7±1.3 75.6±2.4 75.1±2.4 75.5±1.3
– −2.6 −3.3 −1.6 −0.8 −0.6 −0.5
LD-MixStyle (full) LD-MixStyle w/o style bank
77.6±3.6 77.3±1.5
73.5±2.6 72.0±2.1
74.6±1.9 74.4±2.1
– −0.3
Variant
t-SNE dimension 2
(b) Per-label statistics (one point per image--label)
t-SNE dimension 2
(a) Global statistics (one point per image)
t-SNE dimension 1
t-SNE dimension 1
building
car
tree
UCM (target)
water
AID (source)
pavement
ship
DFC15 (source)
Fig. 5. t-SNE [53] embeddings of style statistics for an LD-EFDMix model on Split 1 (test partitions of all domains). (a) Global per-image statistics (µ, σ), plotted once per (image, present label) pair for color comparability. (b) Per-label statistics (µℓ , σ ℓ ) from the proposed decomposition. Colors denote labels; marker shapes denote domains.
TABLE IX C OMPLEXITY AT 224 × 224- PIXEL INPUTS . T HROUGHPUT IS MEASURED IN EVALUATION MODE AT BATCH SIZE 32 ON A SINGLE GPU; MAC S DENOTE MULTIPLY– ACCUMULATE OPERATIONS . Method ERM / global LD (LLAM) LD (Grad-CAM)
Params (M)
∆
MACs (G)
img/s
23.52 23.60 23.52
– +0.35% –
4.087 4.087 4.087
1495–1498 1501–1505 1498–1500
begun to enter RS DG (e.g. CrossEarth [47]). A controlled comparison at that capacity is an important direction left to future work, as it entails a markedly heavier and different training and fine-tuning regime; label-decoupled statistics are, in any case, an orthogonal mechanism that could be inserted into such backbones rather than a competitor to them. Finally,
the method assumes that image-level labels suffice to localize label-specific style; in scenes where a class occupies very few pixels, the attention, learned or cached, may be too coarse. VI. C ONCLUSION This paper introduced a label-decoupled style augmentation framework for domain generalization in multi-label remote sensing image classification. The framework replaces the global feature statistics used by MixStyle, EFDMix, and CSU with attention-derived per-label statistics, mixes them only between cross-domain samples sharing the corresponding label with independent per-label coefficients, and recomposes the feature map through attention-weighted normalization; the attention is provided either by a lightweight learned module or by a periodically refreshed Grad-CAM bank. On a leave-onedomain-out benchmark assembled from the multi-label UCM, AID, and DFC15 datasets, the proposed variants improved
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
upon their global counterparts in all configurations, with the best variant reaching 71.5% average mAP (+5.0 over ERM, and up to +7.7 on the hardest transfer), at a cost of at most 0.35% additional parameters and, as for other statisticsbased augmenters, no change at inference. Ablations and a manifold analysis indicate that spatially correct, up-to-date per-label localization is the decisive component, and that perlabel statistics are considerably better aligned across domains than their global counterparts. The results suggest that label decoupling is a generic and inexpensive upgrade for statistics-based augmentation whenever images contain several co-occurring classes. Future work could examine larger shared vocabularies and additional sensor families, the behavior of the decomposition on transformer backbones, the combination of label-decoupled augmentation with optimization-based methods such as weight averaging, and the integration of label-relation priors into the pairing and bank mechanisms; extending the attention sources to handle classes that occupy very few pixels is a further direction.
R EFERENCES [1] G. Cheng, J. Han, and X. Lu, “Remote sensing image scene classification: Benchmark and state of the art,” Proceedings of the IEEE, vol. 105, no. 10, pp. 1865–1883, 2017. [2] B. Chaudhuri, B. Demir, S. Chaudhuri, and L. Bruzzone, “Multilabel remote sensing image retrieval using a semisupervised graph-theoretic method,” IEEE Transactions on Geoscience and Remote Sensing, vol. 56, no. 2, pp. 1144–1158, 2018. [3] Y. Hua, L. Mou, and X. X. Zhu, “Recurrently exploring class-wise attention in a hybrid convolutional and bidirectional LSTM network for multilabel aerial image classification,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 149, pp. 188–199, 2019. [4] A. Zeggada, F. Melgani, and Y. Bazi, “A deep learning approach to UAV image multilabeling,” IEEE Geoscience and Remote Sensing Letters, vol. 14, no. 5, pp. 694–698, 2017. [5] G. Sumbul and B. Demir, “A deep multi-attention driven approach for multi-label remote sensing image classification,” IEEE Access, vol. 8, pp. 95 934–95 946, 2020. [6] Y. Hua, L. Mou, and X. X. Zhu, “Relation network for multilabel aerial image classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 58, no. 7, pp. 4558–4572, 2020. [7] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, June 2016, pp. 770–778. [8] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Miami, FL, USA, June 2009, pp. 248–255. [9] D. Tuia, C. Persello, and L. Bruzzone, “Domain adaptation for the classification of remote sensing data: An overview of recent advances,” IEEE Geoscience and Remote Sensing Magazine, vol. 4, no. 2, pp. 41– 57, 2016. [10] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand, and V. Lempitsky, “Domain-adversarial training of neural networks,” Journal of Machine Learning Research, vol. 17, no. 59, pp. 1–35, 2016. [11] G. Blanchard, G. Lee, and C. Scott, “Generalizing from several related classification tasks to a new unlabeled sample,” in Advances in Neural Information Processing Systems, Granada, Spain, December 2011, pp. 2178–2186. [12] K. Zhou, Z. Liu, Y. Qiao, T. Xiang, and C. C. Loy, “Domain generalization: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 4, pp. 4396–4415, 2023. [13] J. Wang, C. Lan, C. Liu, Y. Ouyang, T. Qin, W. Lu, Y. Chen, W. Zeng, and P. S. Yu, “Generalizing to unseen domains: A survey on domain generalization,” IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 8, pp. 8052–8072, 2023.
12
[14] K. Muandet, D. Balduzzi, and B. Schölkopf, “Domain generalization via invariant feature representation,” in Proceedings of the 30th International Conference on Machine Learning, Atlanta, GA, USA, June 2013, pp. 10–18. [15] R. Volpi, H. Namkoong, O. Sener, J. Duchi, V. Murino, and S. Savarese, “Generalizing to unseen domains via adversarial data augmentation,” in Advances in Neural Information Processing Systems, Montréal, Canada, December 2018, pp. 5334–5344. [16] F. M. Carlucci, A. D’Innocente, S. Bucci, B. Caputo, and T. Tommasi, “Domain generalization by solving jigsaw puzzles,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, June 2019, pp. 2229–2238. [17] Q. Xu, R. Zhang, Y. Zhang, Y. Wang, and Q. Tian, “A Fourier-based framework for domain generalization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2021, pp. 14 383–14 392. [18] H. Nam, H. Lee, J. Park, W. Yoon, and D. Yoo, “Reducing domain gap by reducing style bias,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2021, pp. 8690–8699. [19] J. Cha, S. Chun, K. Lee, H.-C. Cho, S. Park, Y. Lee, and S. Park, “SWAD: Domain generalization by seeking flat minima,” in Advances in Neural Information Processing Systems, December 2021, pp. 22 405– 22 418. [20] K. Zhou, Y. Yang, Y. Qiao, and T. Xiang, “Domain generalization with MixStyle,” in International Conference on Learning Representations, May 2021. [21] X. Li, Y. Dai, Y. Ge, J. Liu, Y. Shan, and L.-Y. Duan, “Uncertainty modeling for out-of-distribution generalization,” in International Conference on Learning Representations, April 2022. [22] Y. Zhang, M. Li, R. Li, K. Jia, and L. Zhang, “Exact feature distribution matching for arbitrary style transfer and domain generalization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, LA, USA, June 2022, pp. 8035– 8045. [23] Z. Zhang, B. Wang, D. Jha, B. Demir, and U. Bagci, “Domain generalization with correlated style uncertainty,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, Waikoloa, HI, USA, January 2024, pp. 2000–2009. [24] M. Liao, M. Liao, Z. Lu, Y. Sun, Y. Wang, and Y. Peng, “Toward generalizing to unseen domains in remote sensing: A survey,” Neurocomputing, 2026. [25] S. Koda, A. Zeggada, F. Melgani, and R. Nishii, “Spatial and structured SVM for multilabel image classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 56, no. 10, pp. 5948–5960, 2018. [26] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual explanations from deep networks via gradient-based localization,” in Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, October 2017, pp. 618– 626. [27] Y. Yang and S. Newsam, “Bag-of-visual-words and spatial extensions for land-use classification,” in Proceedings of the 18th SIGSPATIAL International Conference on Advances in Geographic Information Systems, 2010, pp. 270–279. [28] G.-S. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y. Zhong, L. Zhang, and X. Lu, “AID: A benchmark data set for performance evaluation of aerial scene classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 55, no. 7, pp. 3965–3981, 2017. [29] “2015 IEEE GRSS Data Fusion Contest,” [Online]. Available: https:// www.grss-ieee.org/community/technical-committees/data-fusion, 2015. [30] G. Tsoumakas and I. Katakis, “Multi-label classification: An overview,” International Journal of Data Warehousing and Mining, vol. 3, no. 3, pp. 1–13, 2007. [31] M.-L. Zhang and Z.-H. Zhou, “A review on multi-label learning algorithms,” IEEE Transactions on Knowledge and Data Engineering, vol. 26, no. 8, pp. 1819–1837, 2014. [32] J. Wang, Y. Yang, J. Mao, Z. Huang, C. Huang, and W. Xu, “CNNRNN: A unified framework for multi-label image classification,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, June 2016, pp. 2285–2294. [33] F. Zhu, H. Li, W. Ouyang, N. Yu, and X. Wang, “Learning spatial regularization with image-level supervisions for multi-label image classification,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, July 2017, pp. 5513–5522. [34] Z.-M. Chen, X.-S. Wei, P. Wang, and Y. Guo, “Multi-label image recognition with graph convolutional networks,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, June 2019, pp. 5177–5186.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
[35] R. Stivaktakis, G. Tsagkatakis, and P. Tsakalides, “Deep learning for multilabel land cover scene categorization using data augmentation,” IEEE Geoscience and Remote Sensing Letters, vol. 16, no. 7, pp. 1031– 1035, 2019. [36] A. Alshehri, Y. Bazi, N. Ammour, H. Almubarak, and N. Alajlan, “Deep attention neural network for multi-label classification in unmanned aerial vehicle imagery,” IEEE Access, vol. 7, pp. 119 873–119 880, 2019. [37] N. Khan, U. Chaudhuri, B. Banerjee, and S. Chaudhuri, “Graph convolutional network for multi-label VHR remote sensing scene recognition,” Neurocomputing, vol. 357, pp. 36–46, 2019. [38] Y. Shu, X. Guo, J. Wu, X. Wang, J. Wang, and M. Long, “CLIPood: Generalizing CLIP to out-of-distributions,” in Proceedings of the 40th International Conference on Machine Learning (ICML), ser. PMLR, vol. 202, 2023, pp. 31 716–31 731. [39] Y. Qiao, K. Li, J. Lin, R. Wei, C. Jiang, Y. Luo, and H. Yang, “Robust domain generalization for multi-modal object recognition,” arXiv preprint arXiv:2408.05831, 2024. [40] Y. Zhao, G. Lin, J. Zheng, Y. You, and H. Fu, “Single domain generalization for scene classification using style-oriented data augmentation,” in IGARSS 2024 – 2024 IEEE International Geoscience and Remote Sensing Symposium, 2024, pp. 8811–8815. [41] R. Iizuka, J. Xia, and N. Yokoya, “Frequency-based optimal style mix for domain generalization in semantic segmentation of remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–14, 2024. [42] S. Zhu, C. Wu, B. Du, and L. Zhang, “Style and content separation network for remote sensing image cross-scene generalization,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 201, pp. 1–11, 2023. [43] E. Durakli, D. Turan, M. Thota, P. Bosilj, and E. Aptoula, “Band aware domain generalization for cross-country multispectral remote sensing scene classification,” in 2024 14th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS), 2024, pp. 1–5. [44] Z. Han, C. Zhang, L. Gao, Z. Zeng, M. K. Ng, B. Zhang, and J. Chanussot, “Multisource collaborative domain generalization for cross-scene remote sensing image classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–15, 2024. [45] E. Durakli and E. Aptoula, “Domain generalized object detection for remote sensing images,” in 2023 31st Signal Processing and Communications Applications Conference (SIU), 2023, pp. 1–4. [46] Y. Zhang, M. Zhang, W. Li, S. Wang, and R. Tao, “Language-aware domain generalization network for cross-scene hyperspectral image classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–12, 2023. [47] Z. Gong, Z. Wei, D. Wang, X. Hu, X. Ma, H. Chen, Y. Jia, Y. Deng, Z. Ji, X. Zhu, X. Yang, N. Yokoya, J. Zhang, B. Du, J. Yan, and L. Zhang, “Crossearth: Geospatial vision foundation model for domain generalizable remote sensing semantic segmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 48, no. 5, pp. 5147– 5164, 2026. [48] X. Huang and S. Belongie, “Arbitrary style transfer in real-time with adaptive instance normalization,” in Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, October 2017, pp. 1501–1510. [49] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, June 2016, pp. 2921–2929. [50] I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in International Conference on Learning Representations, New Orleans, LA, USA, May 2019. [51] I. Gulrajani and D. Lopez-Paz, “In search of lost domain generalization,” in International Conference on Learning Representations, May 2021. [52] T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A nextgeneration hyperparameter optimization framework,” in Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Anchorage, AK, USA, August 2019, pp. 2623–2631. [53] L. van der Maaten and G. Hinton, “Visualizing data using t-SNE,” Journal of Machine Learning Research, vol. 9, pp. 2579–2605, 2008.
13