BSCNet: Boundary- and Scale-Consistent Mean Teacher for Semi-Supervised Building Change Detection in High-Resolution Remote Sensing Images

Abstract

Pixel-level annotation of bi-temporal high-resolution imagery is costly because annotators must distinguish genuine changes from pseudo-changes caused by illumination, seasonality, shadows, and residual misregistration. From a temporal-symmetry perspective, unchanged regions approximately preserve cross-temporal semantic correspondence, whereas genuine building changes introduce localized symmetry breaking between the two acquisition times. This paper presents BSCNet, a semi-supervised framework for binary building change detection that jointly models boundary-sensitive differences and scene-dependent scale preferences. A shared-weight MixTransformer extracts multi-level bi-temporal features. The Edge-Aware Optimization Module suppresses spatially invariant channel responses, enhances residual spatial cues, and predicts a Sobel-supervised edge map. The Parallel Selective Context Module aggregates depthwise-separable branches with different receptive fields and produces an image-level scale distribution. The Multi-scale Edge-Consistent Mean Teacher framework aligns the final prediction, intermediate edge representation, and scale-selection distribution between an exponential-moving-average teacher and the student. Experiments on WHU-CD and LEVIR-CD under 5%, 10%, and 20% labeled-data settings show consistent improvements over RCL, C2F-SemiCD, and CutMix-CD. With 5% labeled data, BSCNet achieves F1/IoU scores of 88.57%/79.49% on WHU-CD and 88.88%/79.98% on LEVIR-CD. An additional UAV-CD evaluation examines transfer to 0.06 m low-altitude UAV imagery containing both building and land changes; under 5% supervision, BSCNet obtains an F1/IoU of 68.07%/51.60%. Progressive ablations confirm complementary gains from the boundary, scale, and consistency components.

1. Introduction

Remote sensing building change detection identifies newly constructed, demolished, or substantially modified buildings by comparing registered observations acquired at different times. This comparison can also be interpreted through temporal symmetry. After spatial registration, unchanged regions are expected to retain approximately corresponding semantic structures across the two observations, whereas construction, demolition, or structural modification introduces localized temporal asymmetry. This symmetry is not strict pixel-wise equality, because illumination, seasonal conditions, shadows, and minor registration errors may alter local appearance without changing the underlying object semantics. From this perspective, an effective change detector should emphasize structural departures from cross-temporal correspondence while remaining insensitive to appearance variations that do not represent genuine changes. Accurate building-change maps support building-inventory maintenance, urban expansion analysis, post-disaster assessment, and geographic database updating. Classical approaches based on radiometric differencing, change vector analysis, thresholding, clustering, or object-level comparison remain useful when annotations are unavailable, but they are sensitive to radiometric inconsistency, seasonal variation, shadows, and residual misregistration [1,2,3]. Deep models reduce the dependence on handcrafted decision rules by learning hierarchical bi-temporal representations [4,5,6].

Supervised building change detection has progressively shifted from direct image differencing to learnable temporal interaction. DSIFN [7] fuses multi-level bi-temporal features through a deeply supervised decoder; DASNet [8] introduces dual attention and a contrastive objective; DSAMNet [9] combines attention-guided metric learning with deep supervision; and ChangeStar [10] explores single-temporal supervision by synthesizing object changes. These methods demonstrate the importance of feature alignment, comparison, and decoding, but they assume that sufficiently accurate pixel-level change masks are available during training.

This assumption is restrictive for building change detection. Each mask requires the joint interpretation of two dates, and difficult pixels are concentrated near corners, oblique contours, narrow gaps, small detached buildings, and partially occluded structures. The spatial extent of changed buildings also varies considerably: local detail is important for small houses and thin structures, whereas broader context is required to recover large or elongated buildings without interior omissions. With limited annotations, boundary-sensitive representations and receptive-field preferences are difficult to learn reliably.

Semi-supervised learning uses unlabeled image pairs to reduce the annotation burden. Mean Teacher [11] produces temporally smoothed targets with exponential moving averaging, FixMatch [12] combines confidence-filtered pseudo-labels with weak-to-strong consistency, Cross-Consistency Training [13] perturbs intermediate representations, and Cross Pseudo Supervision [14] uses two networks for mutual supervision. In dense prediction, however, agreement on the final probability map does not guarantee stable and aligned internal representations for contour localization or scale selection. Prediction-level agreement alone may therefore leave the teacher and student with inconsistent boundary and contextual responses.

Three coupled requirements therefore motivate this work. First, the network should suppress spatially invariant global responses while preserving local residual cues relevant to change boundaries. This formulation is deliberately distinguished from a strict spatial-frequency decomposition: global average pooling yields a per-channel direct-current response, not a spatial low-pass feature map. Second, the network should model scene-dependent scale preference for image pairs containing different mixtures of building sizes. Third, unlabeled supervision should constrain not only the final prediction but also the edge representation and the scale-selection distribution. Existing semi-supervised change-detection methods address pseudo-label reliability, boundary refinement, multi-scale context, or feature consistency, but few jointly regularize boundary representations and scene-level scale weights within a teacher–student framework.

To address these requirements, we propose BSCNet. The Edge-Aware Optimization Module (EAOM) performs global-response suppression, residual enhancement, and auxiliary edge prediction. The Parallel Selective Context Module (PSCM) processes the boundary-aware difference representation with multiple depthwise-separable branches and produces an image-level scale distribution. The Multi-scale Edge-Consistent Mean Teacher (MECMT) framework enforces consistency in the final prediction, the intermediate edge representation, and the scale-selection distribution. PSCM is applied to boundary-aware bi-temporal difference features, and its compact scale distribution is explicitly aligned between the teacher and student. Unlike a prediction-only Mean Teacher, MECMT constrains two intermediate variables directly related to boundary localization and contextual support.

The main contributions are as follows:

(1)

We formulate EAOM as a global-response suppression and residual-enhancement module, avoiding the inaccurate interpretation of global average pooling as a complete spatial low-frequency decomposition. An auxiliary edge head connects Sobel-derived boundary supervision to the boundary-aware difference representation, while residual edge modulation strengthens boundary responses without discarding the underlying four-scale difference features.

(2)

We develop PSCM to aggregate multiple depthwise-separable receptive-field branches over boundary-aware bi-temporal difference features. The resulting four-dimensional attention vector is explicitly defined as an image-level, scene-dependent scale preference.

(3)

We extend Mean Teacher consistency from the final change map to the intermediate edge representation and scale-selection distribution, allowing unlabeled image pairs to regularize the internal mechanisms associated with contour localization and contextual support.

(4)

Experiments on WHU-CD [15] and LEVIR-CD [16] under 5%, 10%, and 20% labeled-data settings show consistent improvements over RCL, C2F-SemiCD, and CutMix-CD. A supplementary UAV-CD [17] evaluation further examines cross-platform and cross-resolution transfer to 0.06 m low-altitude UAV imagery containing both building and land changes. Progressive and factorized ablations verify the contributions of EAOM, PSCM, residual edge modulation, and the three MECMT consistency terms.

2. Related Work

2.1. Deep Learning-Based Remote Sensing Change Detection

Early deep RSCD models established the bi-temporal encoder–decoder formulation that remains widely used. FC-Siamese [18] compares features extracted by shared convolutional encoders, whereas SNUNet-CD [19] introduces densely connected nested decoding to aggregate multi-level Siamese representations. FCCDN [20] constrains both bi-temporal feature extraction and feature fusion through a dual encoder–decoder architecture, a nonlocal feature pyramid, and self-supervised feature constraints. These CNN-based approaches improve spatial reconstruction and feature reuse, but their temporal interaction is usually implemented through concatenation, differencing, dense fusion, or attention within a predefined local hierarchy.

Transformer-based methods extend this formulation by explicitly modeling long-range temporal relationships. BIT [21] represents each temporal image with compact semantic tokens and reasons about their differences through a transformer, while ChangeFormer [22] combines hierarchical transformer encoders with multi-scale difference modules. ScratchFormer [23] integrates local inductive bias and global interaction without relying on generic pretrained weights, and Changer [24] studies temporal feature interaction through aggregation–distribution, parameter-free feature exchange, and flow-based dual alignment. These methods demonstrate that the design of temporal interaction can be as important as the backbone itself.

Related advances in high-resolution remote-sensing semantic segmentation further illustrate the importance of precise spatial representation. GPINet refines encoder–decoder features with geometric priors [25], while EAAHNet explores Euclidean-affinity-augmented hyperbolic representations for pixel-wise semantic inference [26]. Dual-Domain Decoupled Fusion Network further models complementary spatial and frequency representations through dual-domain feature decoupling and fusion [27]. These studies are not change-detection methods, but they provide complementary evidence that contextual geometry and representation refinement are important when dense predictions must preserve object structure and boundaries.

Efficiency and global-context modeling have subsequently become prominent directions. TinyCD [28] uses a lightweight feature-mixing architecture to reduce computational demand. ChangeMamba [29] introduces spatiotemporal state-space modeling for binary change detection, semantic change detection, and building-damage assessment, offering a linear-complexity alternative for global dependency modeling. In high-resolution remote-sensing semantic segmentation, Frequency Domain-Enhanced Spectral-Spatial Fusion Transformer further combines spectral-spatial interaction with frequency-domain enhancement to strengthen dense representation [30]. Beyond remote sensing, transformer-based architectures have also demonstrated effectiveness in other vision tasks requiring spatiotemporal modeling and multi-scale representation. For instance, Wang et al. [31] introduced a dual-scale transformer with variable bitrate synchronization for neural video compression, where two-scale feature extraction and adaptive rate control are jointly optimized to reduce temporal redundancy. This work highlights the broader potential of multi-scale transformer designs for handling sequential visual data, complementing the change-detection context in which bi-temporal feature interaction and scale-adaptive aggregation are similarly critical. Although these supervised methods provide strong architectural foundations, their boundary responses and contextual preferences are learned primarily from labeled data. BSCNet instead considers how these internal representations can be explicitly regularized when only a small subset of paired images is annotated.

2.2. Semi-Supervised Change Detection and Pseudo-Label Learning

General semi-supervised semantic segmentation provides several mechanisms relevant to RSCD. U2PL [32] separates reliable and unreliable predictions and uses low-confidence pixels as category-wise negative evidence instead of discarding them. UniMatch [33] revisits weak-to-strong consistency by combining image-level augmentation with feature perturbation and dual strong views. These studies suggest that unlabeled pixels should not be treated uniformly: the training role of a prediction can depend on its confidence, perturbation stability, and representation context.

Remote-sensing-specific methods adapt these ideas to bi-temporal imagery. SemiCDNet [34] combines supervised prediction with unsupervised constraints, RCL [35] improves changed or unchanged discrimination through reliable contrastive learning, C2F-SemiCD [36] progressively introduces coarse-to-fine consistency, and CutMix-CD [37] enforces consistency on mixed bi-temporal samples. These methods cover adversarial or consistency-based learning, feature discrimination, progressive teacher supervision, and mixed-sample regularization. Their results establish that unlabeled image pairs can improve change detection, but they also expose the sensitivity of the training process to incorrect or incomplete pseudo-labels.

Recent work therefore places greater emphasis on the unit and criterion used to assess pseudo-label reliability. RISL [38] uses instance-level contextual information instead of relying only on independent pixel confidence, whereas PDLCD [39] constructs class and subclass prototypes to improve intra-class compactness and inter-class separation. A Fourier-based semi-supervised method [40] perturbs cross-temporal appearance in the frequency domain and couples this transformation with dynamic pseudo-label selection, while AdaSemiCD [41] evaluates pseudo-label quality through class rebalancing and confusing-area amplification. These approaches address different failure modes, including isolated confidence errors, ambiguous feature geometry, temporal style differences, and class imbalance.

Additional semi-supervised change-detection studies reinforce this trend. Feature-Prediction Alignment (FPA) aligns feature- and prediction-level consistency [42], ECPS improves pseudo-label supervision through ensemble cross-pseudo learning [43], and PUF explicitly models remote-sensing aleatoric uncertainty to guide multi-view semi-supervised learning [44]. SemiBuildingChange instead augments the available supervision by generating pseudo bi-temporal training data for building change detection [45]. Together, these methods motivate the need to distinguish representation-level consistency from mechanisms that explicitly qualify, correct, or diversify pseudo-label supervision.

Other studies improve the structure of the teacher–student system or introduce stronger priors. A CNN–Mamba framework [46] combines heterogeneous local and global learners with cross-architecture feature fusion. CoreNet [47] performs correlation-guided feature refinement and regional pseudo-label reassignment, AG-SemiCD [48] combines attention-guided feature aggregation with adaptive confidence filtering and teacher updating, and StyleDe-CD [49] disentangles input- and feature-level style variation. SAM-CR [50] uses segmentation-foundation-model masks to reshape high-confidence change regions, while DPSNet [51] couples cross-temporal super-resolution with pixel- and subpatch-level reasoning for cross-resolution image pairs.

Foundation-model guidance has also been investigated as an additional source of supervision. SemiCD-VL [52] synthesizes change labels from a visual-language model, separates VLM-derived and consistency-derived supervision through dual projection heads, and introduces auxiliary semantic and metric constraints. This line of work broadens the sources from which unlabeled data can obtain semantic guidance. In contrast, BSCNet does not require an external foundation model; it focuses on transferring the boundary and scale-selection behavior already present inside the teacher network. MECMT therefore complements pseudo-label selection and external-prior approaches by constraining the internal mechanisms most closely associated with contour localization and contextual support.

2.3. Multi-Scale and Contextual Representation

Objects in high-resolution images can vary from a few pixels to large contiguous regions, making multi-scale representation a central component of dense prediction. Feature Pyramid Networks [53] construct top-down pathways with lateral fusion, PSPNet [54] aggregates pooled context over several spatial partitions, and DeepLab [55] uses parallel atrous convolutions to enlarge the effective receptive field. Selective Kernel Networks [56] further learn input-dependent weights for branches with different kernel sizes. These mechanisms provide complementary strategies for combining spatial detail and broader context.

High-resolution representation learning offers a related perspective. HRNet [57] maintains parallel representations at multiple resolutions and repeatedly exchanges information among them, rather than recovering spatial detail only after a low-resolution bottleneck. Such designs are valuable for position-sensitive prediction, but they do not by themselves determine which receptive-field response should dominate for a particular bi-temporal scene. In semi-supervised learning, this issue is more pronounced because scale-selection parameters may be estimated from relatively few labeled examples.

Remote-sensing semantic-segmentation research has similarly explored efficient global–local aggregation and feature-pyramid reasoning. UNetFormer combines a lightweight encoder with global–local Transformer decoding [58], whereas A2-FPN strengthens multiscale feature aggregation through attention-guided pyramid fusion [59]. Frequency Decoupling Network further separates and coordinates frequency-domain representations to improve dense prediction in high-resolution remote-sensing imagery [60]. Although these methods address single-temporal semantic segmentation, they provide complementary evidence that spatial and frequency representations can improve the modeling of fine structures and contextual information.

PSCM follows the principle of selective receptive-field aggregation but differs from a standard selective-kernel block in two respects. First, it receives a boundary-aware bi-temporal difference representation rather than a generic single-temporal feature. Second, its image-level branch distribution is explicitly transferred from the EMA teacher to the student on unlabeled samples. Because the four weights are global to each image pair, PSCM is described as learning a scene-dependent scale preference rather than selecting a different kernel for every object or pixel. This distinction limits the claim to what the implemented attention variable can support.

2.4. Boundary-Aware Dense Prediction and Change Detection

Boundary modeling has a long history in image processing and dense prediction. Sobel [61] and Canny [62] detect local intensity transitions through gradient-based operators, while HED [63] learns multi-level edge responses with deeply supervised side outputs. More recently, Position-Aware Differential Denoising Transformer uses position-aware differential attention to suppress noise while preserving fine semantic structures in remote-sensing imagery [64]. In semantic segmentation, Gated-SCNN [65] separates shape and region processing, and boundary loss [66] optimizes contour agreement through a distance-based formulation. These methods establish that boundary cues can be incorporated as learnable representations or optimization objectives rather than being used only for visualization.

More recent segmentation studies refine difficult contour locations explicitly. PointRend [67] performs point-based prediction at adaptively selected uncertain locations, whereas SegFix [68] replaces unreliable boundary predictions using directions toward more reliable interior pixels. Although these methods produce sharper segmentation contours, they operate through point refinement or prediction-domain correction and are not specifically designed for bi-temporal consistency.

Boundary modeling has also entered supervised and semi-supervised RSCD. BGSNet [69] jointly optimizes semantic change prediction and boundary delineation in a Siamese multitask architecture. Boundary Refinement Teacher [70] identifies uncertain boundary blocks in teacher predictions and re-delineates them with enlarged local views. BSCNet differs in that EAOM constructs a boundary-sensitive change representation before the final decoder by suppressing the spatially invariant global channel response and enhancing residual bi-temporal differences. The Sobel-derived labeled loss is then complemented by edge-feature consistency on unlabeled samples. Frequency-Guided Denoising Network also combines frequency guidance with denoising to preserve fine semantic structures in high-resolution remote-sensing segmentation [71]. This related observation supports the broader motivation for explicitly protecting local detail, while BSCNet differs by embedding boundary-sensitive modeling inside a bi-temporal semi-supervised change-detection framework.

Thus, boundary supervision is embedded in both feature construction and semi-supervised optimization rather than being confined to a final-mask refinement stage.

3. Materials and Methods

3.1. Overall Framework

Given a pair of bi-temporal high-resolution optical images ${\mathbf{X}}^1,{\mathbf{X}}^2\in {\mathbb{R}}^{3\times H\times W}$, the objective is to predict a pixel-level building-change probability map $\mathbf{P}\in {[0,1]}^{1\times H\times W}$. BSCNet is trained under the Multi-scale Edge-Consistent Mean Teacher (MECMT) framework, in which the teacher and student share the same architecture. A shared-weight MixTransformer encoder [72] extracts four-level bi-temporal features:

$${\left\{{\mathbf{F}}_i^1\right\}}_{i=1}^4=\mathcal{B}\left({\mathbf{X}}^1\right),{\left\{{\mathbf{F}}_i^2\right\}}_{i=1}^4=\mathcal{B}\left({\mathbf{X}}^2\right),$$

(1)

where $\mathcal{B}(·)$ denotes the encoder and ${\mathbf{F}}_i^t\in {\mathbb{R}}^{C_i\times H_i\times W_i}$ denotes the feature at temporal index t and scale i. EAOM constructs a boundary-aware bi-temporal difference representation $\mathbf{Z}$. PSCM then aggregates receptive-field responses with different kernel sizes and outputs the fused feature ${\mathbf{F}}_{\mathrm{out}}$ together with an image-level scale-selection distribution $\mathit{\alpha }$. Finally, the decoder integrates and upsamples the fused representation to obtain $\mathbf{P}$. The overall architecture and semi-supervised training workflow of BSCNet are illustrated in Figure 1.

During semi-supervised training, labeled image pairs provide change-mask and boundary supervision. For unlabeled pairs, the EMA teacher supplies targets for the final prediction, the intermediate edge feature, and the scale-selection distribution. Teacher targets are detached from gradient computation, and identical geometric transformations are applied to the two temporal images of each pair to preserve pixel correspondence.

3.2. Edge-Aware Optimization Module

Global average pooling followed by spatial broadcasting produces a spatially invariant response for each feature channel. It is therefore interpreted as a global or direct-current response rather than as a complete spatial low-frequency feature map. EAOM suppresses this global response and enhances the remaining spatial residual. For the i-th feature level,

$${\mathbf{G}}_i^t={\mathcal{U}}_i\left(\mathrm{GAP}\left({\mathbf{F}}_i^t\right)\right),{\mathbf{R}}_i^t={\mathbf{F}}_i^t-{\mathbf{G}}_i^t,t\in \{1,2\},$$

(2)

where ${\mathcal{U}}_i(·)$ broadcasts the pooled channel response to the spatial size $H_i\times W_i$. Here, ${\mathbf{G}}_i^t$ denotes the spatially invariant global-response map obtained by broadcasting the channel-wise pooled response, whereas ${\mathbf{R}}_i^t$ denotes the residual feature obtained after subtracting this response from ${\mathbf{F}}_i^t$. The residual ${\mathbf{R}}_i^t$ contains local spatial deviations, including object contours, texture responses, and local semantic variation; it is not assumed to be a pure frequency component.

The residual is processed by a $7\times 7$ depthwise-separable convolution followed by a $1\times 1$ projection. Specifically,

$${\mathrm{DSConv}}_{7\times 7}\left(\mathbf{R}\right)={\mathrm{PWConv}}_{1\times 1}\left({\mathrm{DWConv}}_{7\times 7}\left(\mathbf{R}\right)\right),$$

(3)

Here, ${\mathrm{DWConv}}_{7\times 7}$ and ${\mathrm{PWConv}}_{1\times 1}$ denote depthwise and pointwise convolutions, respectively, and ${\mathrm{DSConv}}_{7\times 7}$ denotes their sequential composition.

$${\mathbf{H}}_i^t={\mathrm{Conv}}_{1\times 1}\left({\mathrm{DSConv}}_{7\times 7}\left({\mathbf{R}}_i^t\right)\right).$$

(4)

The output ${\mathbf{H}}_i^t$ denotes the projected residual feature for temporal image t at feature level i. The residual bi-temporal difference at each scale is then computed as

$${\mathbf{D}}_i=\left|{\mathbf{H}}_i^1-{\mathbf{H}}_i^2\right|.$$

(5)

Accordingly, ${\mathbf{D}}_i$ denotes the absolute bi-temporal residual-difference feature at scale i.

EAOM combines the highest-resolution difference ${\mathbf{D}}_1$ with the deepest, lowest-resolution difference ${\mathbf{D}}_4$ to construct the edge-aware branch:

$$\mathbf{C}=\mathrm{Concat}\left({\mathbf{D}}_1,{\mathcal{U}}_{1\leftarrow 4}\left({\mathbf{D}}_4\right)\right),$$

(6)

where ${\mathcal{U}}_{1\leftarrow 4}(·)$ denotes bilinear upsampling to the spatial resolution of ${\mathbf{D}}_1$. The output $\mathbf{C}$ is the endpoint concatenation of the highest-resolution difference feature and the upsampled deepest difference feature. This endpoint fusion combines fine-resolution contour information with coarse-resolution semantic context for auxiliary edge prediction.

The concatenated feature is fused by a cascaded $3\times 3$ and $1\times 1$ convolution:

$$\mathbf{E}={\mathrm{Conv}}_{1\times 1}\left({\mathrm{Conv}}_{3\times 3}\left(\mathbf{C}\right)\right),$$

(7)

where $\mathbf{E}\in {\mathbb{R}}^{C_e\times H_1\times W_1}$ is a multi-channel edge-aware difference feature produced by the endpoint fusion branch. A separate prediction head generates the one-channel edge probability map

$${\mathbf{P}}_e={\mathcal{U}}_e\left[\sigma \left({\mathrm{Conv}}_{1\times 1}\left(\mathbf{E}\right)\right)\right],$$

(8)

where ${\mathbf{P}}_e$ denotes the one-channel edge probability map and ${\mathcal{U}}_e(·)$ upsamples the prediction to the label resolution.

For a binary change label $\mathbf{Y}$, the Sobel boundary target is generated as

$${\mathbf{Y}}_e=\mathbb{I}\left(|{\mathbf{S}}_x\ast \mathbf{Y}| + |{\mathbf{S}}_y\ast \mathbf{Y}|>0\right),$$

(9)

where ${\mathbf{S}}_x$ and ${\mathbf{S}}_y$ are the horizontal and vertical Sobel kernels, ∗ denotes convolution, and $\mathbb{I}(·)$ converts nonzero gradient responses into a binary boundary mask. The edge head is supervised using

$${\mathcal{L}}_{\mathrm{edge}}={\mathcal{L}}_{\mathrm{Dice}}({\mathbf{P}}_e^l,{\mathbf{Y}}_e).$$

(10)

In parallel with the endpoint edge branch, all four residual difference features are aligned to the spatial resolution of ${\mathbf{D}}_1$ and fused to form the base bi-temporal difference representation:

$${\mathbf{D}}_{\mathrm{base}}={\mathrm{Conv}}_{1\times 1}\left(\mathrm{Concat}\left[{\mathbf{D}}_1,{\mathcal{U}}_{1\leftarrow 2}\left({\mathbf{D}}_2\right),{\mathcal{U}}_{1\leftarrow 3}\left({\mathbf{D}}_3\right),{\mathcal{U}}_{1\leftarrow 4}\left({\mathbf{D}}_4\right)\right]\right),$$

(11)

where ${\mathcal{U}}_{1\leftarrow i}(·)$ denotes bilinear upsampling from the i-th feature level to the spatial resolution of ${\mathbf{D}}_1$, and ${\mathbf{D}}_{\mathrm{base}}\in {\mathbb{R}}^{C_z\times H_1\times W_1}$. Let ${\mathcal{B}}_c\left({\mathbf{P}}_e\right)$ denote spatial resizing and channel broadcasting of the edge probability map to match ${\mathbf{D}}_{\mathrm{base}}$. The EAOM output is defined through residual edge modulation as

$$\mathbf{Z}={\mathbf{D}}_{\mathrm{base}}\odot \left(1+{\mathcal{B}}_c\left({\mathbf{P}}_e\right)\right),$$

(12)

where ⊙ denotes element-wise multiplication. The identity term preserves the original four-scale difference representation, whereas the edge-derived modulation term selectively strengthens responses around predicted change boundaries. Consequently, the modulation enhances boundary-sensitive responses without directly suppressing non-boundary features. The resulting feature $\mathbf{Z}$ is passed to PSCM. The auxiliary edge head remains active during inference because its output is required to construct the modulation map; only the edge-supervision loss is omitted during inference.

3.3. Parallel Selective Context Module

PSCM addresses the mismatch between a fixed receptive field and image pairs with different scene-level building-scale distributions. Let $\mathbf{Z}$ denote the EAOM output. Four parallel depthwise-separable branches are defined as

$${\mathbf{B}}_k={\mathrm{PWConv}}_{1\times 1}^{\left(k\right)}\left({\mathrm{DWConv}}_{k\times k}\left(\mathbf{Z}\right)\right),k\in \{3,5,7,9\}.$$

(13)

Padding of $(k-1)/2$ preserves the spatial resolution, and the pointwise convolution maps all branch outputs to the same channel dimension. Here, ${\mathbf{B}}_k$ denotes the output feature of the receptive-field branch with kernel size k.

The branch responses are summarized into a global descriptor:

$$\mathbf{s}=\mathrm{GAP}\left(\sum _{k\in \{3,5,7,9\}}{\mathbf{B}}_k\right).$$

(14)

The scale-selection MLP contains two fully connected layers and is formulated as

$$\mathbf{h}=\mathrm{ReLU}({\mathbf{W}}_1\mathbf{s}+{\mathbf{b}}_1),\mathit{\alpha }=\mathrm{Softmax}({\mathbf{W}}_2\mathbf{h}+{\mathbf{b}}_2),$$

(15)

Here, $\mathbf{h}$ denotes the hidden scale descriptor; ${\mathbf{W}}_1$, ${\mathbf{W}}_2$, ${\mathbf{b}}_1$, and ${\mathbf{b}}_2$ are the learnable weights and biases of the two-layer scale-selection MLP; and $\mathit{\alpha }$ is the resulting image-level branch-weight distribution. The hidden dimension is $d=max(\lfloor C_z/16\rfloor ,32)$, and

$$\mathit{\alpha }=[{\alpha }_3,{\alpha }_5,{\alpha }_7,{\alpha }_9],\sum _k{\alpha }_k=1.$$

(16)

Each image pair therefore receives four scalar branch weights. PSCM consequently models an image-level or scene-dependent scale preference rather than assigning a separate receptive field to every object or pixel. The weighted multi-scale feature is

$${\mathbf{F}}_{\mathrm{ms}}=\sum _{k\in \{3,5,7,9\}}{\alpha }_k{\mathbf{B}}_k.$$

(17)

The lightweight spatial correction unit first pools the feature along the channel dimension:

$${\mathbf{Q}}_{\mathrm{avg}}={\mathrm{AvgPool}}_c\left({\mathbf{F}}_{\mathrm{ms}}\right),{\mathbf{Q}}_{\max }={\mathrm{MaxPool}}_c\left({\mathbf{F}}_{\mathrm{ms}}\right),$$

(18)

then predicts a one-channel spatial attention map. The outputs ${\mathbf{Q}}_{\mathrm{avg}}$ and ${\mathbf{Q}}_{\max }$ are the channel-wise average-pooled and max-pooled spatial descriptor maps, respectively.

$${\mathbf{A}}_{\mathrm{sp}}=\sigma \left({\mathrm{Conv}}_{7\times 7}\left(\mathrm{Concat}[{\mathbf{Q}}_{\mathrm{avg}},{\mathbf{Q}}_{\max }]\right)\right).$$

(19)

The final PSCM feature is obtained by residual spatial recalibration:

$${\mathbf{F}}_{\mathrm{out}}={\mathbf{F}}_{\mathrm{ms}}\odot {\mathbf{A}}_{\mathrm{sp}}+{\mathbf{F}}_{\mathrm{ms}}.$$

(20)

PSCM shares the general principle of selective multi-branch aggregation with selective-kernel designs, but it is applied to boundary-aware bi-temporal difference features. More importantly, its scale-selection distribution is retained as an explicit intermediate variable and regularized between the EMA teacher and student in MECMT.

3.4. Multi-Scale Edge-Consistent Mean Teacher Framework

MECMT extends Mean Teacher consistency from the final change map to the intermediate edge feature and scale-selection distribution. For labeled image pairs, the student is supervised using binary cross-entropy, Dice loss, and the Sobel-derived edge target:

$${\mathcal{L}}_{\sup }={\mathcal{L}}_{\mathrm{BCE}}({\mathbf{P}}_s^l,\mathbf{Y})+{\mathcal{L}}_{\mathrm{Dice}}({\mathbf{P}}_s^l,\mathbf{Y})+{\mathcal{L}}_{\mathrm{Dice}}({\mathbf{P}}_{e,s}^l,{\mathbf{Y}}_e).$$

(21)

Throughout this subsection, superscripts l and u denote labeled and unlabeled samples, respectively, whereas subscripts s and t denote student and teacher outputs. Accordingly, ${\mathbf{P}}_s^l$ is the student change-probability map for a labeled pair, and ${\mathbf{P}}_{e,s}^l$ is the corresponding student edge-probability map. The Dice loss is

$${\mathcal{L}}_{\mathrm{Dice}}(\mathbf{P},\mathbf{Y})=1-{\displaystyle \frac{2{\sum }_j{p}_j{y}_j+ϵ}{{\sum }_j{p}_j+{\sum }_j{y}_j+ϵ}},$$

(22)

and the binary cross-entropy loss is

$${\mathcal{L}}_{\mathrm{BCE}}(\mathbf{P},\mathbf{Y})=-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{j=1}^N}\left[y_{j}log{p}_j+(1-y_j)log(1-p_j)\right].$$

(23)

In the Dice and binary cross-entropy formulations, $p_j$ and $y_j$ denote the predicted change probability and binary target of pixel j, respectively; N is the number of evaluated pixels; and $ϵ$ is a small positive constant used for numerical stability.

The teacher and student have identical structures. The teacher is initialized from the student and updated after each training iteration using exponential moving average:

$${\mathit{\theta }}_T^{\left(r\right)}=\eta {\mathit{\theta }}_T^{(r-1)}+(1-\eta ){\mathit{\theta }}_S^{\left(r\right)},$$

(24)

where r denotes the iteration and $\eta =0.999$. The vectors ${\mathit{\theta }}_S^{\left(r\right)}$ and ${\mathit{\theta }}_T^{\left(r\right)}$ denote the student and teacher parameters at iteration r, respectively. Teacher outputs are detached from gradient computation.

To preserve bi-temporal registration, the same geometric transformation is synchronously applied to both images in each pair. The weak augmentation consists of random horizontal flipping, vertical flipping, and ${90}^{\circ }$ rotation. The strong student view additionally uses color jitter with strength 0.2 and Gaussian noise with standard deviation sampled from $[0,0.05]$, while the teacher receives the weakly augmented view. When geometric views differ, the teacher outputs are transformed into the student coordinate system before consistency losses are evaluated.

For an unlabeled pair, the teacher change probability is converted into a binary pseudo-label:

$${\widehat{\mathbf{Y}}}_t^u=\mathbb{I}({\mathbf{P}}_t^u\ge 0.5).$$

(25)

The prediction consistency loss is

$${\mathcal{L}}_{\mathrm{pc}}={\mathcal{L}}_{\mathrm{BCE}}({\mathbf{P}}_s^u,{\widehat{\mathbf{Y}}}_t^u).$$

(26)

The teacher edge representation is used as a stop-gradient target for the corresponding student representation:

$${\mathcal{L}}_{\mathrm{ec}}=\frac{1}{|\mathsf{\Omega }|} {‖{\mathbf{E}}_s^u-\mathrm{sg}\left({\mathbf{E}}_t^u\right)‖}_2^2,$$

(27)

where $\mathrm{sg}(·)$ denotes stop-gradient and $\left|\mathsf{\Omega }\right|$ is the number of feature elements. The scale-selection consistency is

$${\mathcal{L}}_{\mathrm{sc}}=D_{\mathrm{KL}}({\mathit{\alpha }}_t^u\Vert {\mathit{\alpha }}_s^u)=\sum _k{\alpha }_{t,k}^{u}log{\displaystyle \frac{{\alpha }_{t,k}^u+ϵ}{{\alpha }_{s,k}^u+ϵ}}.$$

(28)

Here, ${\mathbf{E}}_s^u$ and ${\mathbf{E}}_t^u$ denote the student and teacher edge-feature tensors for an unlabeled pair, respectively, whereas ${\mathit{\alpha }}_s^u$ and ${\mathit{\alpha }}_t^u$ denote their corresponding image-level scale-selection distributions.

The overall consistency coefficient uses the standard exponential ramp-up

$$\lambda \left(r\right)={\lambda }_{max}exp\left[-5{\left(1-min\left({\displaystyle \frac{r}{R}},1\right)\right)}^2\right],$$

(29)

where ${\lambda }_{max}=1.0$ and $R=30$ epochs. The total objective is

$$\mathcal{L}={\mathcal{L}}_{\sup }+\lambda \left(r\right)\left({\omega }_{\mathrm{pc}}{\mathcal{L}}_{\mathrm{pc}}+{\omega }_{\mathrm{ec}}{\mathcal{L}}_{\mathrm{ec}}+{\omega }_{\mathrm{sc}}{\mathcal{L}}_{\mathrm{sc}}\right),$$

(30)

where ${\omega }_{\mathrm{pc}}$, ${\omega }_{\mathrm{ec}}$, and ${\omega }_{\mathrm{sc}}$ independently control the contributions of prediction, edge-feature, and scale-selection consistency, respectively. The ramp-up coefficient limits the influence of teacher-generated targets during the early training stage.

Overall, MECMT allows unlabeled image pairs to constrain not only the final prediction but also the internal edge and scale representations used to generate it. This formulation directly connects the semi-supervised objective to the two principal architectural mechanisms of BSCNet.

4. Results

4.1. Experimental Setup

4.1.1. Datasets

Experiments are conducted on two widely used high-resolution building change detection datasets, namely WHU-CD and LEVIR-CD. The two datasets differ substantially in spatial resolution, scene composition, building density, and change scale, providing distinct evaluation conditions for the proposed method.

WHU-CD contains two aerial images acquired over Christchurch, New Zealand, in 2012 and 2016. The images have a spatial resolution of approximately 0.2 m and cover an area of about 20.5 km². The main changes are caused by building construction and demolition. Compared with general land-cover change detection datasets, WHU-CD contains a large number of densely distributed residential buildings, large industrial buildings, irregular building contours, and narrow gaps between adjacent instances. These characteristics make accurate boundary localization particularly important. We adopt the official WHU-CD partition, which contains 6096 training pairs, 762 validation pairs, and 762 testing pairs. All samples are non-overlapping $256\times 256$ bi-temporal patches, yielding 7620 image pairs in total.

LEVIR-CD consists of 637 pairs of high-resolution Google Earth images collected from different cities in the United States. Each image has a size of $1024\times 1024$ pixels and a spatial resolution of approximately 0.5 m. The dataset contains building changes with substantial differences in size, shape, orientation, and surrounding background. In particular, it includes scattered small houses, densely arranged residential buildings, elongated industrial buildings, and large commercial structures. We follow the standard LEVIR-CD partition and crop the images into non-overlapping $256\times 256$ patches, obtaining 7120 training pairs, 1024 validation pairs, and 2048 testing pairs.

To evaluate cross-platform and cross-resolution generalization, we additionally use UAV-CD [17,73]. UAV-CD contains 2660 pairs of $768\times 768$ low-altitude UAV images with a spatial resolution of approximately 0.06 m. The dataset was expanded from UAV building-change imagery by increasing sample diversity and incorporating land changes in addition to building changes. Compared with WHU-CD and LEVIR-CD, UAV-CD therefore combines a substantially finer ground resolution, a different acquisition platform, and a broader mixture of changed objects. These characteristics make it a suitable stress test for EAOM boundary modeling and PSCM scale aggregation under high-detail UAV scenes.

Figure 2, Figure 3, and Figure 4 show representative samples from WHU-CD, LEVIR-CD, and UAV-CD, respectively. In each example, the first and second rows correspond to the bi-temporal images, while the third row gives the pixel-level change annotation. White pixels denote changed regions and black pixels denote unchanged background regions. The three datasets exhibit clear differences in spatial resolution, scene composition, object scale, and change type. In particular, UAV-CD contains finer local structures and a broader mixture of building and land changes, providing a complementary visual illustration of the cross-platform and cross-resolution generalization setting considered in this study.

4.1.2. Semi-Supervised Data Partition

To evaluate performance under limited annotation, three labeled-data ratios are considered, namely 5%, 10%, and 20%. For each setting, labeled samples are selected only from the official WHU-CD training split or the standard LEVIR-CD training split, and the remaining training samples are treated as unlabeled data. The validation and testing sets are kept unchanged across all label ratios, ensuring that the evaluation data are identical for the three semi-supervised settings.

The detailed sample numbers are reported in

Table 1. Numbers of labeled, unlabeled, validation, and testing samples under different annotation ratios.

Dataset	Label Ratio	Labeled Train	Unlabeled Train	Validation	Test
WHU-CD	5%	304	5792	762	762
	10%	609	5487	762	762
	20%	1219	4877	762	762
LEVIR-CD	5%	356	6764	1024	2048
	10%	712	6408	1024	2048
	20%	1424	5696	1024	2048

. Under the most challenging 5% setting, only 304 labeled samples are available on WHU-CD and 356 labeled samples are available on LEVIR-CD. Such a low-label regime places strong demands on the reliability of pseudo-labels and intermediate-feature consistency.

For the supplementary UAV-CD experiment, we construct a fixed source-pair split of 2128/266/266 pairs for training/validation/testing (8:1:1). The split is performed at the original $768\times 768$ image-pair level before patch extraction to prevent adjacent patches from the same source pair from entering different subsets. Each source pair is then divided into a $3\times 3$ grid of non-overlapping $256\times 256$ patches, yielding 19,152 training, 2394 validation, and 2394 testing patches. Within the training source pairs, approximately 5%, 10%, and 20% are selected as labeled data, corresponding to 107, 213, and 426 labeled source pairs (963, 1917, and 3834 patches), while the remaining 2021, 1915, and 1702 source pairs are treated as unlabeled data. The same fixed validation and test source pairs are used for all three label ratios.

4.1.3. Comparison Methods

BSCNet is compared with ten representative semi-supervised change-detection methods. RCL [35] improves changed/unchanged discrimination through reliable contrastive learning. C2F-SemiCD [36] progressively introduces coarse-to-fine consistency, while CutMix-CD [37] applies change-aware mixed-sample consistency. RISL [38] evaluates pseudo-label reliability at the instance level, and PDLCD [39] introduces class and subclass prototypes to strengthen feature discrimination. AdaSemiCD [41] combines pseudo-label qualification, adaptive region fusion, and adaptive EMA updating. Boundary Refinement Teacher (BRT) [70] explicitly re-delineates uncertain teacher boundaries. SemiCD-VL [52] introduces visual-language-model guidance as an additional source of supervision. CoreNet [47] performs correlation-guided feature refinement and regional pseudo-label reassignment, whereas AG-SemiCD [48] combines attention-guided feature aggregation with adaptive confidence filtering.

Together, these methods cover contrastive learning, coarse-to-fine consistency, sample mixing, instance- and prototype-level pseudo-label evaluation, boundary refinement, foundation-model guidance, regional reassignment, and adaptive confidence filtering.

For the UAV-CD generalization experiment, RCL, C2F-SemiCD, and CutMix-CD are retained as protocol-aligned quantitative baselines. In addition, the supplied same-sample UAV-CD prediction masks are used for a separate qualitative comparison with representative semi-supervised change-detection methods, including PDLCD, CoreNet, and AG-SemiCD. This visual comparison is analyzed independently from the WHU-CD/LEVIR-CD quantitative ranking so that qualitative boundary behavior is not interpreted as evidence of universal numerical superiority. Quantitative metrics for PDLCD, CoreNet, and AG-SemiCD on UAV-CD are not reported because these methods are compared only through the supplied same-sample prediction masks; their official implementations were not available for retraining under the fixed source-pair-level split used in this study, so a protocol-aligned quantitative comparison could not be conducted without risking an unfair evaluation.

For the protocol-aligned controlled comparisons on WHU-CD and LEVIR-CD, RCL, C2F-SemiCD, CutMix-CD, and BSCNet use the same dataset partitions, labeled-data ratios, input resolution, and evaluation metrics described above. The remaining recent methods in Tables 3 and 4 are included as literature-reported results under their published experimental protocols and are therefore not included in the run-to-run statistical comparison. The supplementary UAV-CD experiment uses the source-pair-level split and patch-extraction protocol specified separately in Section 4.3; the protocol-aligned UAV-CD baselines use the same labeled/unlabeled partition and evaluation set as BSCNet.

4.1.4. Implementation Details

All experiments are conducted on an NVIDIA A40 GPU with 48 GB memory under Ubuntu 18.04.6 LTS. The implementation uses Python 3.8 and PyTorch 1.10.0 with CUDA 11.1. Following the official WHU-CD and standard LEVIR-CD protocols, all network inputs are $256\times 256$ bi-temporal image patches. For UAV-CD, the original $768\times 768$ source pairs are first assigned to train/validation/test subsets and are then cropped into nine non-overlapping $256\times 256$ patches per pair, preserving the same BSCNet input size without resizing the UAV imagery.

BSCNet is optimized using Adam with an initial learning rate of $5\times {10}^{-5}$ and a weight decay of $1\times {10}^{-3}$. The network is trained for 80 epochs, and the EMA decay coefficient is set to 0.999. The overall consistency coefficient follows the exponential ramp-up defined in the methodology and increases from approximately $exp(-5)$ to 1.0. The prediction-, edge-, and scale-consistency terms are controlled by independent coefficients. The adopted configuration uses ${\omega }_{\mathrm{pc}}=1.0$, ${\omega }_{\mathrm{ec}}=0.1$, and ${\omega }_{\mathrm{sc}}=0.01$; the sensitivity of the latter two coefficients is examined in Section 4.4.5.

Each mini-batch contains four labeled and four unlabeled bi-temporal pairs, giving a total batch size of 8. The two temporal images of each pair receive identical geometric transformations to maintain registration. The weak augmentation uses horizontal and vertical flipping and ${90}^{\circ }$ rotation; the strong student view additionally applies color jitter with strength 0.2 and Gaussian noise with standard deviation sampled from $[0,0.05]$.

For model selection, the EMA teacher checkpoint with the highest validation F1 score is used for final evaluation. The principal training parameters are summarized in

Table 2. Main training parameters of BSCNet.

Parameter	Value
Optimizer	Adam
Initial learning rate	$5\times {10}^{-5}$
Weight decay	$1\times {10}^{-3}$
Total batch size	8 (4 labeled + 4 unlabeled)
Training epochs	80
Teacher update	EMA ($\eta =0.999$)
Pseudo-label threshold	0.5
Ramp-up duration	30 epochs
Overall consistency weight	Exponential ramp-up ($exp(-5)\to 1.0$)
Consistency-term weights	$1.0/0.1/0.01$ for ${\omega }_{\mathrm{pc}}/{\omega }_{\mathrm{ec}}/{\omega }_{\mathrm{sc}}$
Evaluation model	Best-validation EMA teacher

. Only one network is retained for inference; the student–teacher duplication is required during training but does not double the reported inference FLOPs. To assess run-to-run variability under limited supervision, the protocol-aligned RCL, C2F-SemiCD, CutMix-CD, and BSCNet configurations are evaluated over five independent runs. The labeled-data ratio, validation set, and test set are kept unchanged, while the random seed controlling network initialization, data ordering, and stochastic augmentation is varied across runs. The same five seeds are used for all four methods at each labeled-data ratio.

Table 3. Quantitative comparison on the WHU-CD test set. F1 and IoU are reported in percentages. For RCL, C2F-SemiCD, CutMix-CD, and BSCNet, values are reported as mean ± standard deviation over five independent runs; the remaining methods are shown as literature-reported point estimates.

Method	5% Labels		10% Labels		20% Labels
	F1	IoU	F1	IoU	F1	IoU
RCL [35]	77.69 ± 0.52	63.52 ± 0.69	83.61 ± 0.39	71.84 ± 0.58	86.08 ± 0.31	75.56 ± 0.48
C2F-SemiCD [36]	84.55 ± 0.35	73.24 ± 0.52	88.79 ± 0.26	79.84 ± 0.42	90.16 ± 0.21	82.08 ± 0.35
CutMix-CD [37]	87.56 ± 0.30	77.87 ± 0.47	88.71 ± 0.25	79.71 ± 0.40	88.83 ± 0.22	79.90 ± 0.36
RISL [38]	89.80	81.48	90.46	82.59	91.14	83.72
PDLCD [39]	91.14	83.72	92.82	86.60	93.65	88.05
AdaSemiCD [41]	80.81	67.80	82.90	70.80	85.32	74.40
BRT [70]	90.17	82.10	91.01	83.50	91.54	84.40
SemiCD-VL [52]	89.99	81.80	90.83	83.20	91.77	84.80
CoreNet [47]	91.48	84.30	92.24	85.60	92.88	86.70
AG-SemiCD [48]	91.66	84.60	92.47	86.00	93.11	87.10
BSCNet	88.57 ± 0.28	79.49 ± 0.45	89.53 ± 0.23	81.05 ± 0.38	90.47 ± 0.19	82.59 ± 0.32

and 

Table 4. Quantitative comparison on the LEVIR-CD test set. F1 and IoU are reported in percentages. For RCL, C2F-SemiCD, CutMix-CD, and BSCNet, values are reported as mean ± standard deviation over five independent runs; the remaining methods are shown as literature-reported point estimates.

Method	5% Labels		10% Labels		20% Labels
	F1	IoU	F1	IoU	F1	IoU
RCL [35]	83.78 ± 0.41	72.09 ± 0.61	85.87 ± 0.34	75.23 ± 0.52	86.77 ± 0.29	76.64 ± 0.45
C2F-SemiCD [36]	88.44 ± 0.27	79.28 ± 0.44	89.67 ± 0.23	81.28 ± 0.38	90.72 ± 0.20	83.02 ± 0.33
CutMix-CD [37]	87.74 ± 0.26	78.15 ± 0.42	88.91 ± 0.23	80.03 ± 0.37	89.53 ± 0.20	81.05 ± 0.32
RISL [38]	90.06	81.92	90.45	82.56	90.51	82.67
PDLCD [39]	90.89	83.30	91.16	83.75	91.48	84.30
AdaSemiCD [41]	87.45	77.70	88.52	79.40	89.07	80.30
BRT [70]	90.71	83.00	91.19	83.80	91.60	84.50
SemiCD-VL [52]	90.05	81.90	90.47	82.60	90.53	82.70
CoreNet [47]	91.30	84.00	91.66	84.60	91.95	85.10
AG-SemiCD [48]	91.60	84.50	91.95	85.10	92.24	85.60
BSCNet	88.88 ± 0.24	79.98 ± 0.39	89.92 ± 0.20	81.68 ± 0.33	90.90 ± 0.17	83.31 ± 0.29

report the mean and standard deviation of F1 and IoU for these controlled runs; methods whose values are taken directly from their original publications remain reported as point estimates.

For practical efficiency profiling, all compared methods are evaluated on the same NVIDIA A40 GPU using $256\times 256$ bi-temporal inputs. Training-time and peak-memory measurements use the same total batch size of 8 as the main experiments, whereas inference latency is measured with a batch size of 1 after a warm-up stage. CUDA synchronization is performed before and after each timed inference to avoid asynchronous timing bias. Training time is reported per epoch on the WHU-CD training split, and peak GPU memory is recorded during training.

4.1.5. Evaluation Metrics

Precision, Recall, F1 score, and Intersection over Union (IoU) are used as the principal accuracy metrics. Let TP, FP, and FN denote the numbers of true-positive, false-positive, and false-negative changed pixels, respectively. Precision and Recall are defined as

$$\mathrm{Precision}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}},\mathrm{Recall}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}.$$

(31)

Precision measures the reliability of pixels predicted as changes, whereas Recall measures the proportion of actual changed pixels successfully detected. The F1 score is their harmonic mean:

$$\mathrm{F}1={\displaystyle \frac{2\times \mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}={\displaystyle \frac{2\mathrm{TP}}{2\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}}.$$

(32)

IoU directly evaluates the overlap between the predicted and ground-truth change regions:

$$\mathrm{IoU}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}+\mathrm{FN}}}.$$

(33)

The F1/Dice and IoU/Jaccard formulations are closely related to classical set-similarity coefficients [74,75].

Because changed pixels generally occupy only a limited proportion of remote sensing images, overall pixel accuracy may be dominated by the unchanged class. F1 and IoU are therefore more informative for evaluating the balance between false alarms and missed detections. In addition, the number of parameters and floating-point operations are reported to analyze model complexity.

4.2. Comparison with Existing Methods

4.2.1. Quantitative Results on WHU-CD

Table 3 reports the quantitative comparison on WHU-CD. BSCNet consistently outperforms RCL, C2F-SemiCD, and CutMix-CD under all three annotation ratios. With 5% labeled data, it obtains 88.57% F1 and 79.49% IoU, exceeding CutMix-CD by 1.01 and 1.62 percentage points, respectively. At 10% labeled data, BSCNet improves the strongest of these three baselines by 0.74 percentage points in F1 and 1.21 percentage points in IoU. At 20% labeled data, the corresponding gains over C2F-SemiCD are 0.31 and 0.51 percentage points.

Several recent methods achieve higher absolute scores. AG-SemiCD obtains the highest F1 and IoU under the 5% and 10% settings, whereas PDLCD performs best under the 20% setting. CoreNet also provides strong results across the three annotation ratios. These comparisons position BSCNet as a competitive method that improves established semi-supervised baselines, while the strongest recent pseudo-label refinement and feature-reassignment approaches retain an advantage in overall accuracy.

The five-run statistics further show that the protocol-aligned results are stable across independent training runs. For BSCNet, the F1 standard deviation decreases from 0.28 to 0.19 percentage points as the labeled ratio increases from 5% to 20%, while the corresponding IoU standard deviation decreases from 0.45 to 0.32 points. The larger variability under 5% supervision is expected because the model relies more strongly on pseudo-label learning when fewer annotated samples are available.

4.2.2. Qualitative Results on WHU-CD

Figure 5 presents qualitative comparisons under the 5% labeled-data setting. White and black denote correctly predicted changed and background regions, respectively. Red regions represent false-positive predictions, while green regions represent false-negative predictions.

RCL reduces several large-area errors but still produces fragmented omissions inside buildings and irregular responses near oblique contours. C2F-SemiCD improves object completeness through coarse-to-fine teacher supervision, although thin false-negative strips remain along several boundaries. CutMix-CD produces compact predictions and relatively high precision, but it occasionally omits narrow structures or parts of large changed buildings. BSCNet preserves the main building regions more completely and yields cleaner boundaries in the displayed examples.

In densely distributed residential scenes, BSCNet better preserves gaps between adjacent buildings and reduces interior fragmentation. For large industrial buildings, the residual edge modulation retains the base difference feature while strengthening contour responses, avoiding the interior suppression that can occur with direct edge masking. These observations are consistent with the quantitative improvements of BSCNet over RCL, C2F-SemiCD, and CutMix-CD.

To further examine the behavior of BSCNet across different levels of bi-temporal change conspicuity, we additionally select six representative WHU-CD test samples covering substantial, moderate, and low changes, with two samples for each level. The grouping is based on the visual conspicuity and structural characteristics of the differences between T1 and T2 rather than solely on the number of changed pixels. The corresponding T1 image, T2 image, ground-truth change mask, and BSCNet prediction are shown in Figure 6.

For the substantial-change samples, the differences between T1 and T2 are visually pronounced, and BSCNet successfully recovers the dominant changed structures, although localized false-positive and false-negative pixels remain near several object boundaries. In the moderate-change cases, the temporal differences become more localized or structurally complicated. The principal changed regions are still identified, while noticeable false-negative responses occur around some elongated structures and irregular boundaries.

The low-change samples contain comparatively subtle bi-temporal differences that are more difficult to distinguish visually. BSCNet nevertheless detects the principal changed regions in both examples, with the remaining errors mainly concentrated near fine object boundaries. These results indicate that BSCNet is not restricted to highly conspicuous changes, while also showing that subtle and boundary-dominated changes remain more challenging than large, visually evident structural changes.

4.2.3. Quantitative Results on LEVIR-CD

Table 4 reports the quantitative comparison on LEVIR-CD. BSCNet outperforms RCL, C2F-SemiCD, and CutMix-CD at all three labeled ratios. With 5% labeled data, BSCNet reaches 88.88% F1 and 79.98% IoU, improving C2F-SemiCD by 0.44 and 0.70 percentage points. At 10% labeled data, the gains are 0.25 percentage points in F1 and 0.40 percentage points in IoU. Under 20% supervision, BSCNet obtains 90.90% F1 and 83.31% IoU, exceeding C2F-SemiCD by 0.18 and 0.29 percentage points.

AG-SemiCD achieves the highest F1 and IoU at all three annotation ratios, followed by CoreNet and PDLCD. RISL and SemiCD-VL also perform strongly under low-label supervision. BRT is particularly relevant to BSCNet because both methods address boundary quality, although they intervene at different stages: BRT refines teacher pseudo-label boundaries, whereas BSCNet regularizes a boundary-aware internal representation and applies residual edge modulation. Overall, BSCNet remains competitive and consistently improves the established comparison methods, but it does not achieve the highest absolute accuracy among all recent approaches.

A similar stability pattern is observed on LEVIR-CD. The F1 standard deviation of BSCNet is 0.24, 0.20, and 0.17 percentage points under 5%, 10%, and 20% supervision, respectively, while the corresponding IoU standard deviations are 0.39, 0.33, and 0.29 points. These values indicate modest run-to-run variability and a gradual reduction in stochastic sensitivity as the amount of labeled supervision increases.

4.2.4. Qualitative Results on LEVIR-CD

Figure 7 compares the protocol-aligned methods on LEVIR-CD under the 10% labeled-data setting. The selected samples contain small residential buildings, large industrial structures, elongated objects, and densely distributed changes.

RCL still misses parts of several small buildings and produces fragmented responses near long boundaries. C2F-SemiCD and CutMix-CD substantially improve object completeness, but thin false-positive or false-negative strips remain around some small and elongated objects. BSCNet produces more complete building regions in the displayed examples and better preserves the spacing between neighboring changes.

For large and elongated structures, the combination of residual edge modulation and multi-branch context aggregation reduces interior omissions while retaining long external contours. The displayed examples provide qualitative evidence of the boundary and object-completeness improvements produced by BSCNet.

4.2.5. Across-Dataset Trends and Label-Efficiency Analysis

The results reveal three common trends. First, relative to RCL, C2F-SemiCD, and CutMix-CD, BSCNet provides its largest IoU margin under the 5% setting: 1.62 percentage points on WHU-CD and 0.70 percentage points on LEVIR-CD. This pattern indicates that the combination of boundary/scale modeling and intermediate consistency is particularly beneficial when direct annotations are scarce.

Second, the performance gain over these established baselines decreases as the labeled ratio increases. From 5% to 20% labeled data, BSCNet improves from 79.49% to 82.59% IoU on WHU-CD and from 79.98% to 83.31% on LEVIR-CD, while its margin over C2F-SemiCD narrows. Additional annotations therefore reduce, but do not eliminate, the benefit of the proposed semi-supervised regularization.

Third, recent instance-level, prototype-based, foundation-guided, and region- reassignment methods form a stronger comparison group. PDLCD, CoreNet, and AG-SemiCD achieve higher overall accuracy in several settings, whereas BSCNet provides consistent gains over RCL, C2F-SemiCD, and CutMix-CD. The following ablation experiments therefore focus on whether the proposed boundary, scale, and consistency mechanisms account for these gains.

4.3. Cross-Platform and Cross-Resolution Generalization on UAV-CD

4.3.1. Quantitative Generalization Results

UAV-CD provides a substantially different evaluation condition from WHU-CD and LEVIR-CD. Its 0.06 m low-altitude UAV imagery contains much finer local texture, stronger sensitivity to shadows and viewpoint variation, and both building and land changes [17,73]. We therefore use it to test whether the boundary-sensitive and scale-aware mechanisms of BSCNet remain useful when the acquisition platform, ground resolution, and change distribution differ from the two core building benchmarks.

Table 5. Semi-supervised generalization results on UAV-CD under different labeled-data ratios. F1 and IoU are reported in percentages. Results are reported under the same 5%, 10%, and 20% labeled-data settings used in the core evaluation.

Method	5% Labels		10% Labels		20% Labels
	F1	IoU	F1	IoU	F1	IoU
RCL [35]	59.94	42.80	62.40	45.35	64.50	47.60
C2F-SemiCD [36]	65.73	48.95	67.64	51.10	69.45	53.20
CutMix-CD [37]	66.40	49.70	68.29	51.85	70.21	54.10
BSCNet	68.07	51.60	70.00	53.85	71.71	55.90

reports the semi-supervised quantitative results on UAV-CD under the 5%, 10%, and 20% labeled-data settings.

Under 5% supervision, BSCNet obtains an F1/IoU of 68.07%/51.60%, improving the strongest protocol-aligned baseline, CutMix-CD, by 1.67 percentage points in F1 and 1.90 percentage points in IoU. The IoU gains over C2F-SemiCD and RCL are 2.65 and 8.80 percentage points, respectively. At 10% supervision, BSCNet reaches 70.00% F1 and 53.85% IoU, maintaining a 2.00-point IoU advantage over CutMix-CD. Under the 20% setting, the F1/IoU increases to 71.71%/55.90%, remaining 1.80 IoU points above CutMix-CD.

The absolute UAV-CD scores are markedly lower than those on WHU-CD and LEVIR-CD, which is consistent with the higher-detail and more heterogeneous UAV setting. As an external fully supervised context rather than a semi-supervised baseline, AEGL-Net reports 75.18% F1 and 60.23% IoU on UAV-CD [73]. The gap to this fully supervised result indicates that the additional UAV experiment should be interpreted as evidence of transfer under limited labels rather than as saturation of the dataset. Overall, the results suggest that EAOM and PSCM remain beneficial under substantial resolution and platform changes, while the smaller absolute accuracy also exposes the need for stronger pseudo-label control in complex UAV scenes.

4.3.2. Qualitative Comparison on UAV-CD

For PDLCD, CoreNet, and AG-SemiCD, only the supplied same-sample prediction masks are available; therefore, these three methods are included for qualitative illustration only and are not listed in Table 5, which reports only the protocol-aligned quantitative baselines. Figure 8 presents a same-sample qualitative comparison on UAV-CD among representative semi-supervised change-detection methods, including RCL, C2F-SemiCD, CutMix-CD, PDLCD, CoreNet, AG-SemiCD, and the proposed BSCNet. This comparison enables direct visual examination of boundary continuity, change-region completeness, false alarms, and missed detections under identical UAV-CD samples. White and black pixels denote correctly classified changed and background regions, respectively, whereas red and green pixels indicate false-positive and false-negative regions.

The compared methods exhibit distinct error patterns across the six UAV scenes. In samples (b), (d), and (f), BSCNet confines most residual errors to relatively narrow bands around long or oblique change boundaries and preserves the dominant changed regions without introducing extensive interior errors. This behavior is particularly evident in sample (f), where BSCNet produces a substantially thinner false-positive band than several comparison methods. These observations are consistent with the intended role of residual edge modulation, which preserves the base multi-scale difference representation while strengthening boundary-sensitive responses.

Nevertheless, the visualization also reveals remaining limitations of BSCNet. In samples (a), (c), and (e), noticeable false-negative or false-positive regions remain around irregular changes and fine-scale boundaries. These cases indicate that the substantially finer spatial resolution of UAV-CD does not eliminate ambiguity caused by complex object geometry, local appearance variation, and heterogeneous land changes. Therefore, the qualitative comparison is used to illustrate the boundary-continuity and change-completeness characteristics targeted by BSCNet rather than to claim universal superiority over the compared methods. This interpretation is also consistent with the WHU-CD and LEVIR-CD quantitative results, where several recent methods achieve higher overall F1 and IoU.

4.4. Ablation Study

The ablation study includes a progressive component analysis and four factorized experiments.

Table 6. Progressive ablation results on WHU-CD and LEVIR-CD. EAOM, PSCM, and MECMT denote the Edge-Aware Optimization Module, Parallel Selective Context Module, and Multi-scale Edge-Consistent Mean Teacher framework, respectively. The symbol “✓” indicates that the component is enabled, and “–” indicates that it is disabled.

Dataset	Label Ratio	EAOM	PSCM	MECMT	F1	Pre	IoU	Rec
WHU-CD	5%	–	–	–	81.44	81.04	68.69	81.83
		✓	–	–	84.59	82.22	73.29	87.10
		✓	✓	–	86.68	85.40	76.49	87.99
		✓	✓	✓	88.57	86.08	79.49	91.22
	10%	–	–	–	83.64	82.73	71.89	84.58
		✓	–	–	86.05	83.24	75.52	89.07
		✓	✓	–	88.20	86.60	78.89	89.86
		✓	✓	✓	89.53	86.86	81.05	92.38
LEVIR-CD	5%	–	–	–	82.80	82.33	70.64	83.27
		✓	–	–	85.12	83.32	74.10	87.01
		✓	✓	–	87.46	86.07	77.71	88.88
		✓	✓	✓	88.88	87.62	79.98	90.17
	10%	–	–	–	84.63	84.32	73.35	84.94
		✓	–	–	87.01	85.20	77.01	88.90
		✓	✓	–	89.18	88.05	80.47	90.33
		✓	✓	✓	89.92	88.27	81.68	91.62

evaluates the cumulative contributions of EAOM, PSCM, and MECMT, while

Table 7. Ablation of the MECMT consistency terms under 5% supervision. PC, EC, and SC denote prediction consistency, edge-feature consistency, and scale-selection consistency, respectively.

Configuration	PC	EC	SC	WHU-CD		LEVIR-CD
				F1	IoU	F1	IoU
Supervised-only full architecture	–	–	–	86.68	76.49	87.46	77.71
Prediction-only Mean Teacher	✓	–	–	87.53	77.82	88.04	78.63
PC + edge consistency	✓	✓	–	88.06	78.67	88.42	79.25
PC + scale consistency	✓	–	✓	87.84	78.31	88.31	79.06
Full MECMT	✓	✓	✓	88.57	79.49	88.88	79.98

,

Table 8. Comparison of edge-supervision and edge-modulation designs under 5% supervision.

Variant	WHU-CD		LEVIR-CD
	F1	IoU	F1	IoU
No edge branch	86.92	76.87	87.46	77.71
Auxiliary edge supervision only	87.63	77.99	88.05	78.65
Direct modulation: ${\mathbf{D}}_{\mathrm{base}}\odot {\mathcal{B}}_c\left({\mathbf{P}}_e\right)$	87.24	77.36	87.81	78.27
Residual modulation: ${\mathbf{D}}_{\mathrm{base}}\odot (1+{\mathcal{B}}_c\left({\mathbf{P}}_e\right))$	88.57	79.49	88.88	79.98

,

Table 9. Comparison of scale-aggregation strategies under 5% supervision.

Scale Strategy	WHU-CD		LEVIR-CD
	F1	IoU	F1	IoU
Single $3\times 3$ branch	87.55	77.86	88.13	78.78
Single $5\times 5$ branch	87.78	78.22	88.29	79.03
Single $7\times 7$ branch	87.85	78.34	88.41	79.23
Single $9\times 9$ branch	87.69	78.08	88.24	78.96
Uniform average of four branches	88.08	78.70	88.55	79.45
Learned PSCM weighting	88.57	79.49	88.88	79.98

and

Table 10. Sensitivity analysis of the edge- and scale-consistency weights under 5% supervision.

${\mathit{\omega }}_{\mathbf{ec}}$	${\mathit{\omega }}_{\mathbf{sc}}$	WHU-CD IoU	LEVIR-CD IoU
0	0	77.82	78.63
0.05	0.01	78.88	79.31
0.10	0.005	79.31	79.82
0.10	0.01	79.49	79.98
0.10	0.02	79.18	79.64
0.20	0.01	79.22	79.71

isolate the consistency terms, edge-modulation strategy, scale-aggregation strategy, and consistency-loss weights.

4.4.1. Progressive Contribution of EAOM, PSCM, and MECMT

The three components are progressively added to the baseline model. The experiments are conducted on both datasets under 5% and 10% annotation ratios, as summarized in Table 6.

Adding EAOM yields average gains of 2.57 percentage points in F1 and 3.84 percentage points in IoU over the four dataset–ratio combinations. The largest change is in Recall, supporting the role of boundary-sensitive residual enhancement in recovering omitted changed pixels. Adding PSCM provides a further average gain of 2.19 percentage points in F1 and 3.41 percentage points in IoU. Its comparatively strong Precision improvement indicates that multi-scale contextual aggregation helps suppress inaccurate responses. MECMT then adds 1.34 percentage points in F1 and 2.16 percentage points in IoU on average, showing that unlabeled consistency provides additional gains after the two architectural modules have been introduced.

4.4.2. Factorized Analysis of the MECMT Consistency Terms

Table 7 separates prediction consistency (PC), edge-feature consistency (EC), and scale-selection consistency (SC) under the 5% setting. The full EAOM+PSCM architecture trained only with labeled data is used as the starting point.

Prediction-only Mean Teacher improves IoU by 1.33 percentage points on WHU-CD and 0.92 percentage points on LEVIR-CD over the supervised-only full architecture. Adding EC or SC individually produces further gains, with EC contributing slightly more under the 5% labeled-data setting. Combining EC and SC gives the highest result on both datasets, indicating that contour-level and scene-level scale consistency provide complementary regularization.

4.4.3. Effect of Residual Edge Modulation

The EAOM edge head can be used only as an auxiliary task, as a direct multiplicative mask, or as the proposed residual modulation. Table 8 distinguishes these alternatives.

Auxiliary edge supervision improves the boundary representation even without explicit modulation. Direct multiplication is less effective because low edge probabilities attenuate the base feature within building interiors. Residual modulation avoids this failure mode: the identity term preserves the four-scale difference representation, while the edge-derived term strengthens responses around predicted boundaries. The resulting gains directly validate the residual formulation used in EAOM.

4.4.4. Fixed and Adaptive Multi-Scale Aggregation

Table 9 compares individual receptive-field branches, uniform multi-branch averaging, and the learned PSCM distribution.

Among the single-branch variants, the $7\times 7$ branch achieves the highest F1 and IoU on both datasets. Uniform averaging nevertheless improves on every single-branch configuration, confirming the value of complementary receptive fields. Learned PSCM weighting provides a further gain of 0.79 percentage points in IoU on WHU-CD and 0.53 percentage points on LEVIR-CD over uniform averaging. This comparison supports the contribution of image-level adaptive weighting without interpreting it as object- or pixel-level scale selection.

4.4.5. Sensitivity to Consistency-Loss Weights

The prediction-consistency coefficient is fixed at ${\omega }_{\mathrm{pc}}=1.0$. Table 10 examines representative values of ${\omega }_{\mathrm{ec}}$ and ${\omega }_{\mathrm{sc}}$ under 5% supervision.

The results show a broad optimum around ${\omega }_{\mathrm{ec}}=0.1$ and ${\omega }_{\mathrm{sc}}=0.01$. Removing both intermediate terms reduces the objective to prediction-only Mean Teacher, whereas excessively increasing either coefficient weakens performance because feature-level or distribution-level agreement begins to dominate the final change objective. The comparatively stable results around the selected setting indicate that BSCNet is not dependent on a single sharply tuned coefficient pair.

4.4.6. Qualitative Ablation Analysis

Figure 9 and Figure 10 visualize the progressive effects of the three principal components. The baseline produces false-positive and false-negative regions around building contours, within large changed objects, and between densely arranged structures.

On WHU-CD, EAOM removes many contour-related errors, while PSCM improves the completeness of large changed regions. The complete MECMT configuration further reduces scattered errors and produces more continuous boundaries. A similar trend is visible on LEVIR-CD, where the complete model better preserves elongated structures and reduces local omissions.

The progressive quantitative and qualitative results characterize the cumulative contributions of EAOM, PSCM, and MECMT. The factorized experiments further validate residual edge modulation, adaptive scale weighting, and the independent consistency terms.

4.5. Computational Complexity and Practical Efficiency

Table 11. Computational complexity and practical efficiency under the unified $256\times 256$ bi-temporal input configuration.

Method	FLOPs (G)	Params (M)	Training Time (s/epoch)	Peak Memory (GB)	Inference Latency (ms/pair)
RCL [35]	73.23	46.85	198	14.3	22.6
C2F-SemiCD [36]	62.10	16.17	176	11.5	20.1
CutMix-CD [37]	72.91	46.85	205	14.7	23.0
BSCNet	45.72	47.13	184	13.2	17.8

extends the original FLOP/parameter comparison with practical engineering indicators measured under the unified $256\times 256$ bi-temporal input configuration. In addition to model size and arithmetic complexity, the evaluation considers training time per epoch, peak training GPU memory, and single-pair inference latency.

BSCNet retains the lowest inference FLOPs and latency among the four controlled configurations, requiring 45.72 G FLOPs and approximately 17.8 ms per bi-temporal pair. Relative to C2F-SemiCD, the FLOP reduction is approximately 26.38%, while the latency is lower by approximately 11.4%. Relative to CutMix-CD, the FLOP reduction is approximately 37.29%, accompanied by a latency reduction of approximately 22.6%.

Parameter count and FLOPs characterize different aspects of computational complexity. The parameter count measures the number of trainable weights stored by the model, whereas FLOPs depend on how frequently these weights are applied and on the spatial resolution of the corresponding feature maps. BSCNet therefore can contain 47.13 M parameters—substantially more than the 16.17 M parameters of C2F-SemiCD—while still requiring fewer arithmetic operations. In particular, EAOM and PSCM rely extensively on depthwise-separable convolutions over hierarchical feature representations, which limits the spatial computational cost of the additional boundary- and scale-modeling operations. BSCNet should therefore be characterized as inference-computation-efficient rather than parameter-light.

The training-side profile is more nuanced. BSCNet requires approximately 184 s per epoch and 13.2 GB peak GPU memory under the stated configuration, which is higher than C2F-SemiCD because MECMT maintains both student and EMA-teacher branches and evaluates additional edge- and scale-consistency terms during training. Nevertheless, the teacher duplication is removed at inference, and the auxiliary edge head is retained only as part of the single BSCNet forward path required for residual modulation. These results indicate that the main efficiency advantage of BSCNet lies in inference computation rather than minimum training cost or parameter count.

5. Discussion

The quantitative results, progressive component ablations, and qualitative comparisons support three main observations about BSCNet. First, boundary modeling is relevant to high-resolution building change detection. Errors in the baseline and comparison methods are frequently concentrated around oblique edges, narrow gaps, building corners, and elongated structures. The results are consistent with EAOM reducing some of these errors through residual-response enhancement and explicit edge supervision.

Second, the appropriate receptive field varies considerably across remote sensing scenes. Small residential buildings require fine local details, whereas large industrial buildings require broader spatial context to avoid incomplete predictions. The improvements observed after introducing PSCM suggest that scene-dependent aggregation of multi-scale responses is beneficial relative to the preceding configuration.

Third, unlabeled data can supervise internal representations in addition to the final pseudo-label map. The additional gains from MECMT suggest that teacher–student agreement at the edge-feature and scale-selection levels provides useful information for training. This is especially beneficial in the 5% setting, where direct annotations are insufficient to reliably learn detailed boundaries and scale distributions.

Taken together, these observations provide a temporal-symmetry interpretation of BSCNet. Unchanged areas tend to preserve cross-temporal semantic and structural correspondence, whereas genuine changes generate localized departures from this approximate symmetry. EAOM emphasizes such departures through residual bi-temporal difference modeling and boundary-aware modulation, while PSCM supplies multi-scale contextual support for asymmetric structures with different spatial extents. MECMT further regularizes the final prediction, edge representation, and scale-selection distribution between the teacher and student, helping the model maintain stable internal responses when unlabeled pairs contain appearance perturbations that do not necessarily correspond to real structural change. This interpretation is conceptual rather than a claim that BSCNet imposes a formal geometric-symmetry or equivariance constraint.

Some residual errors remain visible in the qualitative results. They mainly occur around extremely thin structures, highly irregular corners, tightly adjacent buildings, and regions with strong appearance differences between the two acquisition dates. The progressive experiments show a Recall-oriented improvement that reduces false negatives, although the comparison tables also show that several recent methods achieve higher overall F1 and IoU.

The remaining performance gap can be attributed to three methodological differences. First, BSCNet converts the teacher probability map into binary pseudo-labels using a fixed threshold of 0.5. Recent methods such as AG-SemiCD, CoreNet, PDLCD, PUF, and KAN-UG [76] place stronger emphasis on adaptive confidence filtering, regional pseudo-label reassignment, prototype-level feature discrimination, or uncertainty-aware pseudo-label utilization. These mechanisms directly target erroneous or ambiguous teacher predictions, whereas BSCNet primarily regularizes boundary-sensitive and scale-related internal representations. As a result, incorrect teacher pseudo-labels can still enter the prediction-consistency loss of BSCNet.

Second, PSCM produces four scalar branch weights for each image pair and consequently models scene-level rather than spatially varying scale preference. Although the factorized experiments show that learned weighting is more effective than a single receptive-field branch or uniform averaging, the current formulation cannot independently select different receptive fields for small and large changed objects occurring simultaneously at different locations. Third, MECMT aligns the final prediction, edge feature, and scale-selection distribution but does not explicitly reassign or correct erroneous teacher regions before they are used as supervision. These characteristics explain why BSCNet can improve established baselines while remaining below recent methods whose primary contribution is stronger pseudo-label correction or discriminative restructuring.

The UAV-CD generalization experiment provides a complementary view of these limitations. BSCNet maintains an advantage over the protocol-aligned RCL, C2F-SemiCD, and CutMix-CD baselines under 0.06 m low-altitude UAV imagery, and the same-sample visualization shows comparatively continuous boundaries in several difficult cases. However, the absolute UAV-CD accuracy is substantially lower than on WHU-CD and LEVIR-CD. The different acquisition platform, much finer ground resolution, high-frequency texture, local viewpoint variation, shadows, and the inclusion of land changes beyond buildings increase the ambiguity of teacher predictions. This pattern supports the transferability of boundary and scale consistency while also showing that cross-platform generalization requires stronger uncertainty-aware pseudo-label control and region-specific correction. Future work will investigate adaptive confidence estimation, regional pseudo-label refinement, and spatially varying scale selection while retaining the current boundary-aware formulation.

6. Conclusions

This paper presented BSCNet for semi-supervised binary building change detection in high-resolution optical remote sensing images. EAOM suppresses spatially invariant global channel responses and enhances residual bi-temporal differences, while an auxiliary Sobel-supervised edge head encourages boundary-sensitive feature learning. PSCM aggregates four depthwise-separable receptive-field branches and produces an image-level scale-selection distribution. MECMT extends teacher–student consistency from the final prediction to the intermediate edge feature and scale distribution.

On WHU-CD and LEVIR-CD, BSCNet consistently outperforms RCL, C2F-SemiCD, and CutMix-CD under 5%, 10%, and 20% labeled-data settings, while remaining competitive with recent semi-supervised methods. Under 5% supervision, it obtains 88.57% F1 and 79.49% IoU on WHU-CD, and 88.88% F1 and 79.98% IoU on LEVIR-CD. The progressive ablation results show that EAOM, PSCM, and MECMT provide complementary improvements, and the factorized experiments verify the effectiveness of residual edge modulation, adaptive scale aggregation, and the independent MECMT consistency terms. These results also support viewing building change detection as the localization of temporal asymmetry within otherwise corresponding bi-temporal structures. Under this interpretation, EAOM emphasizes localized symmetry-breaking cues near changed structures, PSCM provides scale-adaptive contextual support for these cues, and MECMT encourages consistent prediction-, edge-, and scale-level representations under limited supervision. The symmetry connection therefore provides a complementary conceptual view of the bi-temporal change-detection problem without altering the implemented architecture or its experimental claims.

A supplementary UAV-CD evaluation further examines transfer to low-altitude UAV imagery with approximately 0.06 m spatial resolution and a mixture of building and land changes. Under 5%, 10%, and 20% supervision, BSCNet obtains F1/IoU scores of 68.07%/51.60%, 70.00%/53.85%, and 71.71%/55.90%, respectively, consistently exceeding the protocol-aligned RCL, C2F-SemiCD, and CutMix-CD baselines. The accompanying same-sample qualitative comparison with PDLCD, CoreNet, and AG-SemiCD further illustrates the boundary behavior of BSCNet in high-detail UAV scenes.

The core evaluation remains focused on binary building changes in WHU-CD and LEVIR-CD, while UAV-CD provides a supplementary cross-platform and cross-resolution test under much finer UAV imagery. The four PSCM weights represent scene-level rather than spatially varying scale preferences, and extremely thin structures, tightly adjacent objects, strong bi-temporal appearance differences, UAV-specific texture and viewpoint effects, and heterogeneous land changes remain challenging. Future work will investigate spatially adaptive scale selection, uncertainty-aware pseudo-label control, regional pseudo-label correction, and broader evaluation on multi-class and cross-sensor change-detection tasks.