Transformer-based architectures have increasingly gained traction in remote sensing tasks due to their ability to model long-range dependencies and contextual relationships in spatial data (Hong et al., 2021; Li et al., 2025; Chen et al., 2024). Traditional convolutional neural networks (CNNs), while effective at capturing local patterns, often struggle with learning global representations, which are critical in analyzing high-resolution remote sensing imagery (Roy et al., 2022). Vision Transformers (ViTs), initially proposed for natural image classification, have been adapted for remote sensing tasks, demonstrating competitive or superior performance compared to CNN counterparts (Khan et al., 2020; Tanaka et al., 2023; He et al., 2024). One major adaptation involves the incorporation of hierarchical structures and locality inductive biases into transformer models to address the high computational cost and lack of inherent translation equivariance (Zhu et al., 2020). Other approaches like TransUNet integrate transformer encoders with CNN-based decoders, combining the global context modeling of transformers with the detailed spatial resolution capabilities of CNNs (Chen L. et al., 2021). In the context of land cover classification, transformers have shown notable performance semantic segmentation tasks, where precise delineation of land types is required. Remote sensing datasets often encompass diverse and complex landscapes, making the modeling of global relationships essential for accurate classification (Ashtiani et al., 2021). Multi-scale and multi-modal transformers have been developed to leverage information from various spectral bands and resolutions, further improving classification accuracy (Masana et al., 2020). Furthermore, hybrid models that combine CNNs with transformers have been introduced to mitigate the limitations of pure transformer models in spatial feature extraction. These models often use CNNs to extract initial low-level features, followed by transformers to model the interrelations across spatial patches. This synergy has led to improved performance in tasks such as change detection, object detection, and land use classification. Research has also explored domain-specific adaptations, such as employing transformers for hyperspectral image classification, where the spectral dimension introduces additional complexity (Sheykhmousa et al., 2020). Transformers’ ability to handle sequential data makes them particularly suited for capturing spectral-spatial correlations. The CoastVisionNet builds upon this trajectory by embedding a transformer backbone tailored for coastal land cover segmentation, suggesting the potential benefits of leveraging transformer architectures in domains characterized by complex spatial dynamics and heterogeneous features. The choice of transformers aligns with the growing trend of adopting attention-based models in remote sensing, particularly where global context and feature interactions significantly impact classification outcomes (Zheng et al., 2022).

Attention mechanisms have revolutionized deep learning-based image segmentation by enhancing a model’s ability to focus on relevant spatial and channel-wise features. In semantic segmentation, accurately classifying each pixel in an image necessitates the discrimination of subtle contextual differences across regions, a task well-suited for attention-enhanced architectures (Mascarenhas and l Agarwal, 2021). Spatial attention mechanisms guide the model to emphasize significant regions in an image, effectively acting as a soft spatial mask. This is particularly useful in land cover classification, where certain areas, such as water bodies or vegetation, may occupy only a small portion of the image yet are crucial for accurate segmentation. Spatial attention enhances feature maps by weighting the importance of each spatial location based on its relevance to the task (Zhang et al., 2022). Channel attention, on the other hand, focuses on reweighting the importance of each feature channel. In convolutional neural networks, different channels encode different semantic information. Channel attention modules, such as the Squeeze-and-Excitation (SE) block, dynamically adjust the contribution of each channel, enabling the model to prioritize more informative features. This has been particularly useful in tasks requiring fine-grained recognition and classification (Dai and Gao, 2021). More advanced architectures combine spatial and channel attention to simultaneously refine spatial and semantic features. Dual attention mechanisms, like those used in the Dual Attention Network (DANet), allow for modeling both spatial dependencies and channel interrelationships, enhancing segmentation accuracy across complex scenes. Other techniques include attention gates in encoder-decoder frameworks, which selectively propagate information through the network hierarchy, improving feature localization (Taori et al., 2020). The integration of attention mechanisms with transformers further amplifies their benefits. Transformers inherently utilize self-attention, which models all pairwise interactions between elements, offering a holistic view of the input (Peng et al., 2022). However, integrating explicit spatial and channel attention modules allows for finer control over the learned features and enhances interpretability. In the context of CoastVisionNet, the incorporation of integrated spatial-channel attention modules is crucial. Coastal regions are characterized by high spatial heterogeneity and diverse land cover types, such as beaches, mangroves, urban zones, and agricultural fields. A combined attention approach enables the model to focus on salient spatial patterns and important feature channels that distinguish these categories. This dual attention strategy enhances the discriminative power of the network, leading to more accurate and context-aware segmentation results (Bazi et al., 2021).

Coastal land cover classification is a critical component of environmental monitoring, urban planning, and disaster management (Hong et al., 2021). These regions are characterized by dynamic landscapes influenced by natural and anthropogenic factors, necessitating robust methods for accurate classification (Dong et al., 2022). Traditional classification approaches relied on pixel-based methods using spectral indices and machine learning algorithms such as support vector machines (SVMs) and random forests, often limited by their inability to capture spatial context (Chen C.-F. et al., 2021). With the advent of deep learning, convolutional neural networks (CNNs) have become the dominant approach for land cover classification, offering superior performance through hierarchical feature extraction (Maurício et al., 2023). However, these models often require large annotated datasets and may struggle with classifying small or irregularly shaped objects typical of coastal environments. Recent developments have introduced multi-scale and multi-temporal methods to address the temporal and spatial variability in coastal regions (Liu et al., 2024). These methods leverage time-series data to capture seasonal changes and long-term trends, enhancing the model’s ability to distinguish between classes that exhibit similar spectral signatures but differ temporally. Incorporating elevation data and ancillary information, such as LiDAR or radar imagery, has further improved classification outcomes by providing additional contextual cues (Liu et al., 2023b). Moreover, object-based image analysis (OBIA) has emerged as an effective strategy, segmenting imagery into meaningful objects rather than individual pixels. OBIA combined with deep learning facilitates the integration of spatial, spectral, and contextual information, resulting in more coherent and accurate classifications. The use of remote sensing data from various platforms, including Sentinel-2, Landsat, and UAVs, offers diverse spatial and spectral resolutions, which can be exploited through data fusion techniques. These techniques merge information from multiple sources, enhancing the richness of input data and improving classification accuracy (Liu et al., 2023a). CoastVisionNet contributes to this evolving landscape by introducing a transformer-based model designed for coastal land cover classification. Its architecture incorporates spatial-channel attention mechanisms, tailored to the unique challenges of coastal environments (Wang et al., 2024; Zhao et al., 2022; Deng et al., 2024). This model addresses the limitations of prior approaches by capturing long-range dependencies, emphasizing relevant spatial regions, and adapting to the heterogeneous nature of coastal land types. The design of CoastVisionNet reflects a synthesis of advances in deep learning, attention mechanisms, and remote sensing, positioning it as a state-of-the-art solution for coastal land cover analysis.

Remote sensing has long served as a pivotal modality for a wide spectrum of scientific and industrial applications, ranging from environmental monitoring to urban planning, from agricultural forecasting to military reconnaissance. The remarkable advancements in sensor technology and the advent of high-resolution, multi-spectral, and temporally-rich satellite imagery have posed both immense opportunities and significant challenges for automated analysis. At the heart of these challenges lies the fundamental issue of developing scalable, generalizable, and semantically interpretable models that can effectively reason over spatial and spectral heterogeneity. This paper introduces a novel methodology designed to address these foundational requirements.

Our methodological framework is built around a unified architecture for remote sensing scene understanding, designed to integrate symbolic formalization, architectural novelty, and strategic data alignment. To this end, the following three components constitute the core contributions of this work: an abstract formulation of remote sensing interpretation the umbrella of structured representation learning, a newly introduced model architecture that harnesses hierarchical representations for multi-scale spectral reasoning, and a domain-specific inference strategy that enables dynamic modulation of semantic priors and spectral dependencies. Together, these three pillars allow the framework to address major limitations in existing remote sensing pipelines, particularly in their scalability to unseen distributions, their rigidity in encoding multisource semantics, and their lack of adaptive reasoning mechanisms.

The first core component of the proposed approach lies in the formalization of the remote sensing problem space. Remote sensing imagery is intrinsically high-dimensional, temporally sparse, and spatially redundant. Furthermore, the semantic classes present in satellite images are entangled across scales and often exhibit intra-class variability and inter-class ambiguity. In Section 3.2, we provide a comprehensive mathematical formalism of the remote sensing domain. This involves constructing the input-output space through rigorous notations of spectral vectors, semantic categories, spatial neighborhoods, and inter-band covariance. More critically, we establish a set of transformation invariances and stochastic process assumptions which guide the formulation of the underlying inference problem. These include band-shift invariance, translation equivariance, and latent topological priors—all of which underpin the design of downstream modules. Informed by the abstract formulation, Section 3.3 introduces a new model, which we term Spectral-Topographic Encoding Network (STEN). Unlike conventional convolutional or attention-based approaches that operate in a fixed-resolution space, our model dynamically adapts to both spectral density and spatial topology. The model leverages a dual-path encoding scheme: one path captures local spectral gradient fields using multi-scale depth-wise convolutions, while the other path models topographic contours and edge distributions using a variational vector field decomposition. These two modalities are then fused through a geometry-aware self-attention mechanism that learns spectral co-occurrence patterns conditioned on topographic continuity. By decoupling the representation of spectral semantics and spatial morphology, STEN not only achieves better class separability but also significantly enhances generalization to out-of-distribution samples. Furthermore, the model includes a recursive encoding layer that iteratively refines feature maps based on residual inter-band entropy, a technique inspired by information bottleneck theory. The final component is presented in Section 3.4, wherein we propose an inference strategy referred to as Spectrum-Guided Semantic Modulation (SGSM). The motivation behind SGSM stems from the observation that remote sensing categories are often not mutually exclusive but rather spectrum-dependent. For instance, urban infrastructure and barren land may share similar spectral signatures under certain illumination and seasonal conditions. SGSM introduces a context-sensitive inference pipeline that modulates semantic predictions via a learned spectrum-attention gate. This gate dynamically adjusts the decision boundaries based on inter-band correlation coefficients and ambient reflectance priors. The strategy also incorporates a curriculum-inspired mechanism for spectral augmentation, wherein the training regime selectively emphasizes hard-to-distinguish spectra during early epochs and gradually incorporates easier spectra as training stabilizes. SGSM integrates an uncertainty-aware regularization term in the optimization objective, which penalizes semantically inconsistent predictions across neighboring spectral bands. Through the combination of structured problem formulation, tailored model architecture, and domain-aware strategy, our method achieves state-of-the-art performance on multiple remote sensing benchmarks. These include land cover classification, scene parsing, and object segmentation tasks across a diverse set of satellite platforms and resolutions. More importantly, the unified framework facilitates transferability across different regions and imaging conditions, a crucial property for real-world deployment.

Remote sensing data encapsulate a highly structured and hierarchical form of information, composed of spectral, spatial, and temporal components. To formulate our methodology rigorously, we first provide a formal mathematical description of the remote sensing problem space. Our objective is to establish a unified symbolic foundation that guides the construction of learning objectives, model architectures, and semantic strategies. This section introduces a notational framework for remote sensing image representation, defines key invariances and structural assumptions, and builds a structured inference formulation for downstream semantic tasks such as classification, segmentation, and scene interpretation.

Let I : Ω → R B denote a remote sensing image defined on a spatial domain Ω ⊂ Z 2 , where each pixel x ∈ Ω is associated with a B -dimensional spectral vector s x ∈ R B . Each element s x ( b ) of s x corresponds to the reflectance value in the b -th spectral band.

We define a global image tensor as Equation 1: T = s x ∣ x ∈ Ω ∈ R H × W × B , (1) where H and W denote the image height and width, respectively.

To encode the local spatial context, we consider a square neighborhood N r ( x ) of radius r centered at pixel x Equation 2: N r x = x ′ ∈ Ω ∣ ‖ x ′ − x ‖ ∞ ≤ r . (2)

The concatenated spatial-spectral neighborhood feature of pixel x is defined as Equation 3: z x = vec s x ′ x ′ ∈ N r x ∈ R 2 r + 1 2 ⋅ B . (3)

Let Y = { 1 , … , C } be the set of semantic categories such as vegetation, water, urban, and barren land. Each pixel x is associated with a (possibly latent) label y x ∈ Y .

Define the decision function as Equation 4: f θ : R 2 r + 1 2 ⋅ B → Δ C , (4) where θ denotes the set of learnable parameters, and Δ C is the C -dimensional probability simplex Equation 5: Δ C = p ∈ 0,1 C ∑ c = 1 C p c = 1 . (5)

Spectral vectors of homogeneous land cover regions lie on low-dimensional manifolds embedded in R B . Formally, for a semantic class c ∈ Y , there exists a manifold M c ⊂ R B such that Equation 6: ∀ x ∈ Ω , y x = c ⇒ s x ∈ M c + ϵ x , (6) where ϵ x ∼ N ( 0 , Σ c ) represents Gaussian perturbation due to noise and atmospheric distortion.

Let M = ⋃ c = 1 C M c be the global spectral manifold. Then, the embedding function ϕ maps s x to a latent representation h x such that Equation 7: ϕ : R B → H , H ⊂ R d , d ≪ B , (7) and ideally preserves geodesic distances Equation 8: e x t d i s t H ϕ s x , ϕ s x ′ ≈ dist M s x , s x ′ . (8)

Let C ∈ R B × B denote the spectral covariance matrix over the entire image Equation 9: C = E x ∼ Ω s x − μ s x − μ ⊤ , (9) where μ = E x ∼ Ω [ s x ] is the mean spectrum.

We define a redundancy penalty operator as Equation 10: R C = ∑ i ≠ j ρ i j , ρ i j = C i j C i i ⋅ C j j , (10) to capture redundant information among spectral bands.

A key property of remote sensing is spatial translation equivariance Equation 11: f θ z x + δ = f θ z x ∀ δ ∈ Z 2 ,  when  T  is homogeneous. (11)

This justifies the application of convolutional or locally-shared architectures. However, edge effects and topographic variance often violate this assumption locally, motivating the use of adaptive filters.

We present the Spectral-Topographic Encoding Network (STEN), a hybrid architecture that integrates spectral analysis and topographic pattern learning to enhance semantic representation of multi-band imaging data. STEN is built upon three core innovations: a residual spectral encoder for capturing cross-band dependencies, a differential topographic encoder to extract spatial-geometric cues, and a transformer-based fusion mechanism that aligns the heterogeneous modalities for robust feature learning.

The architecture follows a four-stage hierarchical structure with progressively reduced spatial resolution and increased channel dimensionality. Each stage includes patch embedding, convolutional blocks, and residual connections. For better clarity and visual alignment, The overall architectural pipeline, including the stage-wise spatial and channel dimensions, is shown in Figure 1. The detailed internal structure of the STEN module is illustrated in Figure 2.

Remote sensing imagery presents profound spectral diversity and spatial ambiguity, often exacerbated by domain shifts across sensors, seasons, and geographies. To address this, we propose a Spectrum-Guided Semantic Modulation (SGSM) strategy that dynamically adjusts STEN’s internal behavior during training and inference. SGSM achieves adaptivity through three core mechanisms: spectral prior encoding, uncertainty-aware semantic gating, and a spectrum-driven curriculum scheduler (As shown in Figure 4).

It is important to note that both the Spectral-Topographic Encoding Network (STEN) and the Spectrum-Guided Semantic Modulation (SGSM) modules are employed during both training and inference. STEN operates as the core feature extraction backbone in all phases, while SGSM is integrated into the prediction head to refine outputs via spectral priors and confidence-based gating. These modules are fully differentiable and impact gradient flow during training, and during inference, they retain their functionality to enhance robustness and adaptivity under unseen or noisy spectral distributions.

The BigEarthNet dataset Sumbul et al. (2021) is a large-scale benchmark consisting of over 590,000 Sentinel-2 image patches across 10 European countries, each annotated with one or more land-cover class labels based on the CORINE Land Cover (CLC) nomenclature. The dataset spans 43 semantic categories, including artificial surfaces, agricultural zones, forests, wetlands, and water bodies. Each image patch covers a 120 × 120 m area and contains 12 spectral bands, facilitating multi-label scene classification, land use monitoring, and deep representation learning under diverse seasonal and geographic conditions. The OSCD Dataset (Onera Satellite Change Detection Dataset) Fu et al. (2021) includes bi-temporal multispectral image pairs captured by SPOT-6 and SPOT-7 satellites across 24 urban and rural regions worldwide. Annotated with binary change masks, the dataset supports supervised change detection tasks and includes 13 spectral bands. OSCD enables robust evaluation under spatial misalignment, atmospheric variation, and domain shift, and is widely used in research involving urban dynamics, environmental monitoring, and post-disaster analysis. The LandCoverNet dataset Alemohammad and Booth (2020) is a globally distributed Sentinel-2-based land cover dataset curated by the Radiant Earth Foundation, comprising more than 200,000 scene-labeled image chips across five continents. It adheres to the Dynamic World schema with semantic classes including built-up areas, trees, crops, water, wetlands, and bare ground. Each chip is georeferenced and temporally aligned with expert-validated labels, supporting tasks such as global-scale semantic segmentation, domain generalization, and label robustness studies. The EuroSAT dataset Bhatt and Bhatt (2024) is a medium-scale classification benchmark derived from Sentinel-2 imagery, containing 27,000 labeled image patches across 10 land use and land cover types including residential, industrial, forest, river, pasture, and highway. Each patch is 64 × 64 pixels and includes all 13 spectral bands, allowing both RGB-based and multispectral training. Its balanced class distribution and wide accessibility make EuroSAT a popular dataset for remote sensing classification, deep model prototyping, and educational use in satellite image analysis.

All experiments were conducted using PyTorch framework on a workstation equipped with NVIDIA A100 GPUs. We adopted a mini-batch size of 64 for all datasets, and trained the models using the Adam optimizer with a weight decay of 1 e − 4 . The initial learning rate was set to 0.001 and decayed using a cosine annealing schedule over the course of 100 epochs. For fair comparison, all models were trained under the same computational budget and data augmentation strategies. For image classification tasks, we applied random resized cropping to 224 × 224 pixels, horizontal flipping with a probability of 0.5, and normalization using the BigEarthNet Dataset mean and standard deviation. No additional data augmentation tricks like mixup or CutMix were employed unless explicitly stated. For the backbone network, we utilized a standard ResNet-50 architecture pretrained on BigEarthNet Dataset, followed by a lightweight transformer-based feature aggregator tailored to improve discriminative representation learning. During training, the final fully connected layer was modified to match the number of classes for each respective dataset. For datasets with fine-grained categories like LandCoverNet Dataset, we added a channel attention module to the feature extractor to enhance the learning of subtle local patterns. For the EuroSAT Dataset, we employed group normalization over batch normalization to maintain stability across small batch sizes, given the variability of texture patterns. Each experiment was repeated three times with different random seeds, and the average results are reported to ensure robustness. For hyperparameter tuning, we performed grid search on the validation set using 10% of the training data. Cross-validation was used only for datasets with fewer samples, such as OSCD Dataset, where stratified k-fold (k = 5) was employed to mitigate class imbalance. To ensure reproducibility, we fixed random seeds across numpy, PyTorch, and CUDA environments, and logged all experimental configurations using Weights & Biases. During testing, only center cropping was applied, and top-1 accuracy was used as the primary evaluation metric. For detailed analysis, confusion matrices and per-class accuracy were also computed. Our implementation also supports gradient checkpointing to save memory during training, which was particularly useful for high-resolution texture images in EuroSAT Dataset. To accelerate convergence, label smoothing with a factor of 0.1 was used, especially for datasets prone to overfitting due to small sample size. The experiments were benchmarked under consistent environmental conditions, and no hyperparameter tuning was performed on the test set. All scripts and configuration files will be made publicly available to facilitate reproducibility and further research.

To comprehensively evaluate the effectiveness of our proposed method, we conduct extensive comparisons against several state-of-the-art (SOTA) models, including ResNet50 Theckedath and Sedamkar (2020), ViT Touvron et al. (2022), EfficientNet Koonce (2021), DenseNet Dalvi et al. (2023), ConvNeXt Feng et al. (2022), and DeiT Touvron et al. (2022). As shown in Tables 2, 3, our method consistently achieves superior performance across all four benchmark datasets. On the large-scale BigEarthNet Dataset, our method achieves an accuracy of 84.91%, surpassing the best baseline ConvNeXt by 2.88%. For OSCD Dataset, which features higher intra-class variance and fewer training samples per class, our model outperforms all SOTA methods by a notable margin of 1.75% in Accuracy and 2.18% in AUC. This performance boost is largely attributed to our architecture’s ability to incorporate both global semantic context and local feature discrimination via the transformer-augmented aggregation module, which allows for dynamic feature reweighting that adapts to diverse visual patterns across varying datasets.

On fine-grained classification tasks, such as LandCoverNet Dataset, the superiority of our method becomes even more pronounced. Our method reaches a top-1 Accuracy of 91.35%, which is 2.34% higher than ConvNeXt, and outperforms ViT and EfficientNet by larger margins. In fine-grained tasks, where categories differ by subtle texture, shape, and color variations, the strength of our architecture lies in its ability to preserve fine-level details while suppressing irrelevant background noise. This is further supported by the high F1 Score (90.08%) and AUC (93.62%) achieved, indicating stable generalization under intra-class ambiguity. The channel attention module and the localized token enhancement approach in our framework are particularly effective in detecting discriminative floral features. On the EuroSAT dataset, our model again achieves the best performance with 79.29% Accuracy and 81.74% AUC, outperforming all competitors. Unlike object classification datasets, texture datasets require models to reason about style and abstract visual attributes. The effectiveness of our method on EuroSAT Dataset is a strong testament to the flexibility of our representation learning mechanism, which integrates hierarchical texture semantics through feature pyramids and context-aware refinement. The use of group normalization instead of batch normalization on EuroSAT Dataset effectively stabilizes training under smaller batch regimes, which is crucial for capturing nuanced texture patterns.

In Figures 6, 7, these consistent improvements can be largely attributed to several design components in our model, as described in the method section. The hybrid feature extractor ensures both hierarchical abstraction and spatial precision. The design of multi-resolution fusion in the transformer encoder contributes to the ability to model long-range dependencies, enhancing recognition in complex visual scenes. Our approach also benefits from a lightweight architecture that maintains computational efficiency while achieving top-tier performance. Unlike models such as ViT and ConvNeXt, which are computationally intensive, our model achieves higher accuracy without sacrificing training and inference speed. These results not only validate the effectiveness of our method across a diverse set of datasets but also highlight its generalizability and robustness, demonstrating strong potential for real-world deployment in both generic and fine-grained classification tasks.

To investigate the contribution of each key component in our proposed architecture, we conducted a comprehensive ablation study by systematically removing individual modules and evaluating the performance on all four benchmark datasets. The components under study include Residual Spectral Encoding, Differential Topographic Encoding, and Spectral Prior Encoding. The results are summarized in Tables 4, 5. We denote the full model as Ours and use Residual Spectral Encoding, Differential Topographic Encoding, and Spectral Prior Encoding to represent variants with the respective component removed.

On the BigEarthNet Dataset and OSCD Dataset, the removal of Residual Spectral Encoding causes a noticeable performance drop, reducing accuracy by 2.56% and 1.68% respectively. This indicates that the transformer-based feature aggregator plays a critical role in capturing long-range dependencies and contextual relationships between image regions. Without this module, the model tends to rely heavily on local features, resulting in inferior generalization on diverse and large-scale datasets. The absence of Differential Topographic Encoding also leads to a performance decline, though to a lesser extent. Accuracy drops by approximately 1.64% on OSCD Dataset, emphasizing the importance of channel-wise recalibration in enhancing the discriminative power of learned features. Interestingly, removing Spectral Prior Encoding impacts BigEarthNet Dataset more significantly than OSCD Dataset, suggesting that the multi-resolution fusion is particularly beneficial in scenarios with high visual complexity. These results reinforce our design intuition that fusing multi-scale features is essential for building hierarchical representations adaptable to variable object scales and contexts.

On the LandCoverNet Dataset and EuroSAT Dataset, we observe a similar trend. The elimination of Residual Spectral Encoding results in accuracy drops of 1.95% and 2.45% respectively, confirming its relevance even in fine-grained or texture-heavy classification tasks. The EuroSAT Dataset, in particular, benefits from global reasoning enabled by the transformer module, since textures often span irregular patterns beyond local receptive fields. The effect of removing Differential Topographic Encoding is again significant, reducing F1 Score by over 1.5% across both datasets. In Figures 6, 7, this aligns with our hypothesis that channel-level reweighting is crucial for recognizing subtle attribute differences in fine-grained categories. Spectral Prior Encoding, although having slightly less influence on Oxford Flowers, still contributes meaningfully on the EuroSAT Dataset, where multi-scale features help capture both micro and macro texture patterns. The full model consistently outperforms all ablated versions across all datasets, validating the complementary nature of each proposed component. These findings highlight the necessity of an integrated design, where attention, fusion, and global context work in unison to boost both classification accuracy and generalization robustness.

In addition to widely-used CNN and ViT-based baselines, we further evaluate CoastVisionNet against several domain-specific transformer models tailored for remote sensing segmentation tasks. These include SpectralFormer, TransUNet, Swin-Unet, and U-Former—each of which leverages multi-scale attention mechanisms and spectral modeling strategies suitable for RS data. The experiments were conducted on both the LandCoverNet and EuroSAT datasets under consistent training settings. As shown in Table 6 and Figure 8, our method outperforms all the compared RS-specific transformer models in terms of classification accuracy, F1 score, and AUC. CoastVisionNet achieves 91.35% accuracy and 90.08% F1 score on LandCoverNet, along with 79.29% accuracy on EuroSAT. These results demonstrate the superiority and scalability of our proposed framework in handling complex coastal and terrestrial environments.

To provide empirical support for the efficiency claims, we evaluated the number of parameters and average inference time of each compared model. All measurements were conducted using a single NVIDIA A100 GPU on 224 × 224 patches. As presented in Table 7 and Figure 9, CoastVisionNet contains only 47.2 million trainable parameters and achieves an average inference latency of 32.5 ms per image, making it the most efficient model in the set. This confirms its potential for scalable deployment in operational remote sensing systems where resource constraints are critical.

To further assess the model’s ability to delineate complex coastal land cover types, we present qualitative comparisons in Figure 10. The examples are taken from the LandCoverNet dataset and cover a wide range of coastal conditions including water-land boundaries, urban-sand transitions, and vegetated zones. Each row shows: (1) the original Sentinel-2 RGB image, (2) the corresponding ground truth label map, (3) the prediction from Swin-Unet as a strong baseline, and (4) the prediction from CoastVisionNet. It can be observed that our model yields cleaner segmentation boundaries, reduced fragmentation in small land patches, and more accurate identification of mixed land classes in coastal regions. This visual evidence complements the quantitative performance metrics and highlights the interpretability and precision of CoastVisionNet.