FusionFormer-X: Hierarchical Self-Attentive Multimodal Transformer for HSI-LiDAR Remote Sensing Scene Understanding

1. Introduction

Remote sensing (RS) technologies have emerged as pivotal tools for Earth observation (EO), with applications spanning land use classification [1,2,3], mineralogical surveys [4], environmental monitoring [7], urban planning [8], ecological conservation, and disaster response coordination. The availability of diverse RS data sources has catalyzed a paradigm shift from traditional manual analysis to data-driven intelligent processing pipelines, wherein both classical machine learning and contemporary deep learning (DL) methods are heavily deployed.

Despite significant progress, many prior works have been confined to unimodal sensing, particularly focusing on Hyperspectral Imaging (HSI), which captures rich spectral information but often lacks spatial granularity [9]. This intrinsic limitation hampers its ability to distinguish between semantically different landcover categories that may share similar spectral signatures (e.g., concrete roads versus rooftops). On the contrary, LiDAR systems, employing active sensing mechanisms, provide elevation and 3D shape cues, which are inherently complementary to HSI’s spectral sensitivity [10]. Thus, the fusion of HSI and LiDAR is both intuitive and advantageous for holistic scene interpretation in complex landscapes.

Over the past decade, efforts have been made to explore such fusion strategies. Classical methods such as EP-based spatial feature extraction or subspace-based learning (e.g., CoSpace) attempt to map multimodal data into a shared feature domain. While promising, these approaches often lack robustness in non-linear complex scenes. The emergence of deep neural architectures has opened new avenues, with convolutional neural networks (CNNs) [11,12] providing significant performance gains. Recently, Transformer-based architectures have garnered attention due to their self-attention capability, which excels in modeling long-range dependencies and global contextual interactions [13].

SpectralFormer [14] is one such model leveraging the attention mechanism to model inter-band relationships in HSI. However, it lacks spatial modeling and thus underperforms in joint spectral-spatial tasks. To tackle this, MFT [10] introduced a multimodal ViT framework to incorporate both HSI and secondary data. Yet, this model fails to resolve the resolution disparity challenge between modalities like HSI and LiDAR, leading to suboptimal feature alignment and inconsistent performance in cluttered scenes [15].

To this end, we propose FusionFormer-X, a novel Transformer-based fusion framework that incorporates (1) convolutional tokenization to introduce local inductive biases while preserving spatial semantics, and (2) a Hierarchical Multi-Scale Multi-Head Self-Attention (H-MSMHSA) module that performs coarse-to-fine multimodal feature fusion, effectively addressing the alignment gap between HSI and LiDAR modalities.

In summary, the contributions of this paper are fourfold:

1

We propose a new architecture, FusionFormer-X, designed explicitly for fusing high-dimensional spectral and geometric cues from HSI and LiDAR data.

2

We develop a novel Hierarchical Multi-Scale Multi-Head Self-Attention module that enables progressive cross-modal feature integration with spatial and spectral consistency.

3

We integrate convolutional inductive biases into the tokenization stage, enhancing local feature modeling and preserving fine-grained spatial structures.

4

We conduct extensive experiments on Trento and MUUFL benchmarks, showing that FusionFormer-X significantly outperforms existing state-of-the-art methods across various evaluation metrics.

By systematically unifying spectral richness with geometric structure, FusionFormer-X contributes to the advancement of multimodal learning in remote sensing and lays the groundwork for future developments in generalizable EO models.

2. Related Work and Preliminary Studies

The task of fusing multimodal remote sensing (RS) data—particularly the integration of hyperspectral imagery (HSI) with complementary modalities such as Light Detection and Ranging (LiDAR)—has received sustained attention from both the remote sensing and machine learning communities. Early research efforts focused on traditional handcrafted methods that aimed to capture spatial and spectral cues using engineered filters and statistical classifiers. These classic techniques laid the groundwork for multimodal fusion by attempting to extract meaningful patterns from structurally distinct data sources.

Among these earlier approaches, a range of morphological and profile-based methods were introduced, including morphological profiles (MPs) [16], attribute profiles (APs) [17], and extinction profiles (EPs) [18]. These techniques primarily aimed at enhancing spatial-spectral representation by generating descriptors based on structural transformations and attribute filtering. Meanwhile, statistical classifiers such as Random Forests (RF) [20] gained popularity due to their robustness on high-dimensional but limited-sample datasets, a common scenario in RS applications. Ham et al. [20] notably demonstrated the capability of RF-based hierarchical classifiers in generalizing over limited hyperspectral training samples.

In parallel, subspace learning methods emerged as another effective strategy for multimodal fusion. Techniques such as Canonical Correlation Analysis (CCA) and its nonlinear variants were widely used to project disparate modalities into a common latent space, facilitating joint feature extraction and classification [19]. These methods, while effective in reducing feature redundancy and aligning cross-modal representations, often relied heavily on linear assumptions, limiting their adaptability in highly nonlinear scenes.

The limitations of traditional approaches—especially in scalability, flexibility, and semantic expressiveness—have fueled the widespread adoption of deep learning (DL) methods in recent years. Convolutional Neural Networks (CNNs), with their powerful local receptive fields and hierarchical feature extraction capability, have shown great promise in modeling spectral-spatial relationships inherent in HSI data. For instance, Makantasis et al. [11] proposed a dual-branch CNN model that separately encoded the spatial and spectral features of pixel neighborhoods, demonstrating considerable improvements over prior handcrafted methods.

Building upon the success of CNNs, attention has gradually shifted to Transformer-based architectures, which overcome the spatial locality limitations of CNNs by modeling long-range dependencies across the entire input. Originally introduced in natural language processing [13], Transformers have been successfully adapted for image understanding tasks and, more recently, for multimodal RS data fusion. Unlike CNNs, Transformers offer global context modeling through self-attention, allowing for dynamic interaction across input tokens without fixed receptive fields.

A notable milestone in this direction is SpectralFormer [14], which employed a cross-layer encoder built upon the Vision Transformer (ViT) backbone to model inter-band spectral relationships. By leveraging self-attention across adjacent spectral channels, SpectralFormer effectively captured spectral continuity, though it largely ignored the spatial dimension, thereby limiting its applicability in full-scene classification tasks.

To address this, Swalpa et al. [10] proposed the Multimodal Fusion Transformer (MFT), which extends ViT to incorporate both HSI and LiDAR modalities. The MFT model demonstrated the potential of transformer-based architectures in multimodal fusion; however, it suffered from several critical drawbacks. Most notably, MFT did not explicitly address the inherent resolution gap between modalities—particularly the high spectral resolution of HSI versus the low spatial resolution and sparse nature of LiDAR. This mismatch often led to feature misalignment and compromised fusion quality in complex environments.

In contrast to these previous models, our proposed FusionFormer-X builds upon the foundational Transformer design but incorporates three key innovations to better handle multimodal fusion in RS contexts. First, we introduce Convolutional Tokenization Blocks prior to Transformer encoding, which utilize stacked convolutional layers to embed local spatial patterns while preserving positional integrity. This design incorporates inductive biases that are known to be beneficial for remote sensing imagery with structured spatial layouts.

Figure 1. Overview of the overall framework.

Figure 1. Overview of the overall framework.

Second, we formulate a Hierarchical Multi-Scale Multi-Head Self-Attention (H-MSMHSA) mechanism that extends standard self-attention to perform progressive interaction across different spatial scales and spectral domains. Unlike vanilla multi-head attention, our H-MSMHSA module groups transformer heads by resolution bands and dynamically adapts attention weighting across modalities. Lastly, we employ a cross-modal fusion bottleneck layer that aligns the final stage embeddings from both modalities via learnable transformation matrices, enabling better joint prediction at the classification head.

In summary, the body of related work reveals a clear evolution from handcrafted feature engineering to deep and transformer-based architectures. However, existing models often overlook key challenges such as modality disparity, spatial-spectral inconsistency, and fusion adaptability. By introducing architectural enhancements and principled multi-scale attention mechanisms, FusionFormer-X seeks to address these gaps, offering a more robust and generalizable solution for multimodal RS image classification.

3. Methodology

3.1. Hierarchical FusionFormer-X Framework

The overall design of our proposed FusionFormer-X follows a hybrid fusion architecture tailored to integrate hyperspectral (HSI) and LiDAR modalities, aimed at enhancing landcover classification in remote sensing. Unlike conventional Vision Transformer (ViT) models, which often neglect modality discrepancies and spatial fidelity, FusionFormer-X leverages a shallow yet expressive transformer backbone (depth=2) combined with convolutional inductive priors and multiscale attention to achieve enhanced representation.

Before tokenization, both HSI and LiDAR data are preprocessed to a common spatial scale via upsampling. Specifically, raw input cubes of size $11\times 11$ are padded to $16\times 16$ using zero-padding, ensuring consistent patch-wise operations across modalities.

Convolutional Feature Encoding. Instead of partitioning frames into isolated patches and applying naive linear projections, we incorporate 3D and 2D convolutional blocks to jointly encode spectral and spatial features. For HSI, a combination of Conv3D [12] and HetConv2D [21] is used to project the original spectral channels to a reduced embedding space of 64 channels:

$${\mathbf{X}}_{\mathrm{HSI}}^{\left(64\right)}=\mathrm{HetConv}2\mathrm{D}\left(\mathrm{Conv}3\mathrm{D}\left({\mathbf{X}}_{\mathrm{HSI}}^{\left(C\right)}\right)\right),$$

(1)

where C is the number of original spectral bands.

For LiDAR, a Conv2D layer is applied to upsample its single-band or low-channel feature into a 64-channel representation:

$${\mathbf{X}}_{\mathrm{LiDAR}}^{\left(64\right)}=\mathrm{Conv}2\mathrm{D}\left({\mathbf{X}}_{\mathrm{LiDAR}}^{\left(1\right)}\right).$$

(2)

We then concatenate both encoded features along the channel axis:

$${\mathbf{X}}_{\mathrm{fusion}}\in {\mathbb{R}}^{128\times H\times W}=\mathrm{Concat}({\mathbf{X}}_{\mathrm{HSI}}^{\left(64\right)},{\mathbf{X}}_{\mathrm{LiDAR}}^{\left(64\right)}).$$

(3)

Convolutional Tokenization and Projection. The concatenated multimodal feature ${\mathbf{X}}_{\mathrm{fusion}}$ is processed by a shared convolutional embedding to produce the token embeddings for transformer attention:

$$\mathbf{Q},\mathbf{K},\mathbf{V}=\mathrm{Conv}2\mathrm{D}({\mathbf{X}}_{\mathrm{fusion}},k=(1,1)).$$

(4)

For stability and better gradient flow, we replace traditional LayerNorm and linear projection with:

$$\begin{array}{cc}\hfill \mathbf{Y}& =\mathrm{LeakyReLU}\left(\mathrm{Conv}2\mathrm{D}\right(\mathbf{X},k=(3,3),p=(1,1)),\alpha =0.2),\hfill \\ \hfill \mathbf{Z}& =\mathrm{BN}\left(\mathbf{Y}\right),\hfill \end{array}$$

(5)

where $\mathbf{Z}$ is the normalized feature map input to MSMHSA.

3.2. Multi-Scale Self-Attention via Spatial Pyramids

The key component of FusionFormer-X is the Multi-scale Multi-head Self-Attention (MSMHSA), which fuses features at varying spatial resolutions using a pyramid-based design. We denote the input token maps $\mathbf{Q}/\mathbf{K}/\mathbf{V}\in {\mathbb{R}}^{C\times H\times W}$. The total number of heads is fixed to 3, and each head operates at a different resolution.

Head-wise Partitioning.

Each head $i\in \{1,2,3\}$ takes a portion of channels $C_i=C/3$, and feature maps are divided as:

$$\begin{array}{c}\hfill {\mathbf{Q}}_i,{\mathbf{K}}_i,{\mathbf{V}}_i\in {\mathbb{R}}^{C_i\times H_i\times W_i},\end{array}$$

(6)

where $H_i,W_i\in \{H,H/2,H/4\}$ respectively, via spatial splitting.

Self-Attention within Each Scale.

For head i, the scaled dot-product attention is computed:

$${\mathbf{A}}_i=\mathrm{Softmax}\left({\displaystyle \frac{{\mathbf{Q}}_i·{\mathbf{K}}_i^T}{\sqrt{d_i}}}\right),{\mathbf{H}}_i={\mathbf{A}}_i·{\mathbf{V}}_i,$$

(7)

where $d_i$ is the key dimension.

Hierarchical Aggregation.

The final attention output is:

$${\mathbf{H}}_{\mathrm{MSMHSA}}=\mathrm{Concat}(\mathrm{Upsample}\left({\mathbf{H}}_1\right),\mathrm{Upsample}\left({\mathbf{H}}_2\right),\mathrm{Upsample}\left({\mathbf{H}}_3\right)).$$

(8)

Positional Bias Injection.

To retain positional semantics during multiscale fusion, we inject a sinusoidal position embedding $\mathbf{P}$ at each scale:

$${\mathbf{Q}}_i\leftarrow {\mathbf{Q}}_i+{\mathbf{P}}_i,{\mathbf{K}}_i\leftarrow {\mathbf{K}}_i+{\mathbf{P}}_i.$$

(9)

3.3. Cross-Modality Gated Fusion Layer

To ensure balanced learning from HSI and LiDAR inputs, we design a Gated Fusion Unit:

$$\begin{array}{cc}\hfill \mathbf{G}& =\sigma \left(\mathrm{Conv}2\mathrm{D}\left([{\mathbf{X}}_{\mathrm{HSI}},{\mathbf{X}}_{\mathrm{LiDAR}}]\right)\right)\hfill \\ \hfill {\mathbf{F}}_{\mathrm{fused}}& =\mathbf{G}\odot {\mathbf{X}}_{\mathrm{HSI}}+(1-\mathbf{G})\odot {\mathbf{X}}_{\mathrm{LiDAR}},\hfill \end{array}$$

(10)

where ⊙ denotes element-wise multiplication and $\sigma$ is the sigmoid gate.

3.4. Feedforward Projection with Dual-Scale Normalization

Following MSMHSA, a modified FFN layer is applied:

$$\begin{array}{cc}\hfill \mathbf{Y}& =\mathrm{Conv}2\mathrm{D}\left(\mathrm{BN}\left(\mathrm{Conv}2\mathrm{D}\left({\mathbf{H}}_{\mathrm{MSMHSA}}\right)\right)\right)\hfill \\ \hfill \mathbf{Z}& =\mathrm{LN}\left(\mathbf{Y}\right)+{\mathbf{H}}_{\mathrm{MSMHSA}},\hfill \end{array}$$

(11)

where BN and LN refer to BatchNorm and LayerNorm respectively, ensuring local-global feature calibration.

3.5. Objective Function and Regularization

The model is optimized via a composite objective:

$${\mathcal{L}}_{\mathrm{total}}={\mathcal{L}}_{\mathrm{cls}}+{\lambda }_1{\mathcal{L}}_{\mathrm{fusion}}+{\lambda }_2{\mathcal{L}}_{\mathrm{entropy}},$$

(12)

where ${\mathcal{L}}_{\mathrm{cls}}$ is the cross-entropy classification loss, ${\mathcal{L}}_{\mathrm{fusion}}={∥{\mathbf{Z}}_{\mathrm{HSI}}-{\mathbf{Z}}_{\mathrm{LiDAR}}∥}_2^2$ ensures modality alignment, and ${\mathcal{L}}_{\mathrm{entropy}}=-\sum plogp$ encourages confident predictions.

3.6. MLP Classifier Head

The final fused tokens are passed through a two-layer MLP head:

$${\mathbf{Z}}_{\mathrm{class}}=\mathrm{MLP}\left(\mathrm{Flatten}\left(\mathbf{Z}\right)\right),$$

(13)

with softmax output over landcover classes.

This design enables FusionFormer-X to perform joint spectral-spatial feature modeling with cross-modal consistency, and delivers state-of-the-art performance on multimodal remote sensing benchmarks.

4. Experiments

4.1. Benchmark Datasets and Evaluation Protocols

To comprehensively evaluate the performance and generalization of our proposed FusionFormer-X, we conduct extensive experiments on two well-established multimodal remote sensing datasets that provide co-registered hyperspectral and LiDAR data: the Trento and MUUFL benchmarks. These datasets represent both rural and urban environments with diverse spatial and material characteristics.

• Trento Dataset. This dataset captures a rural zone located south of Trento, Italy. It comprises a hyperspectral image with 63 spectral bands and a corresponding single-band LiDAR-derived digital surface model (DSM). The spatial resolution is 1 meter, and the image spans $166\times 600$ pixels. There are six labeled landcover categories, including buildings, trees, and terrain classes such as grass and agricultural land. The simplicity in background but complexity in class overlap makes Trento ideal for analyzing cross-modal synergy.
• MUUFL Dataset. The MUUFL Gulfport dataset was acquired over the University of Southern Mississippi campus. After noise band removal, 64 effective hyperspectral bands remain, and the LiDAR modality contains 2 elevation-related channels. With a spatial size of $325\times 220$ pixels, this dataset features 11 fine-grained landcover types including road markings, curbs, trees, and man-made structures. MUUFL is more challenging due to narrow classes, urban clutter, and spectral ambiguities.
• Training Configuration. All models are implemented in PyTorch 1.12.1 and trained on a CentOS 7.9 workstation equipped with a single NVIDIA RTX 3090 GPU (24 GB). The batch size is fixed to 64 for all models to ensure comparability. Optimization is performed using the Adam optimizer with an initial learning rate of $5\times {10}^{-4}$, decayed by $\gamma =0.9$ every 50 epochs using a step-based scheduler. A weight decay of $5\times {10}^{-3}$ is used for regularization. Each model is trained for 500 epochs and evaluated across 3 independent seeds, reporting the mean and standard deviation.
• Evaluation Metrics. We adopt three standard metrics for classification: Overall Accuracy (OA), Average Accuracy (AA), and Cohen’s Kappa coefficient ($\kappa$). OA reflects pixel-wise global accuracy. AA measures the average per-class accuracy, accounting for class imbalance. Kappa provides a chance-corrected agreement:

  $$\kappa ={\displaystyle \frac{p_o-p_e}{1-p_e}},$$

  where $p_o$ is the observed agreement and $p_e$ is the expected agreement. These metrics collectively ensure both absolute and balanced performance evaluation.

4.2. Performance Comparison on Trento and MUUFL

We benchmark FusionFormer-X against a spectrum of baselines including classical machine learning and recent deep learning methods: Random Forest (RF) [20], CNN2D [11], Vision Transformer (ViT) [22], SpectralFormer [14], and MFT [10]. These models represent the evolution of remote sensing classification from traditional to transformer-based architectures.

On the Trento dataset, FusionFormer-X achieves state-of-the-art results with $99.18\%$ OA, $97.91\%$ AA, and $\kappa =98.90\%$, consistently outperforming all baselines. The gains over MFT (previous best) are clear, particularly in challenging terrain classes. On MUUFL, where class boundaries are subtle and multiple narrow objects coexist, FusionFormer-X yields $94.73\%$ OA, $84.57\%$ AA, and $\kappa =93.02\%$, showing strong generalization.

These results confirm that FusionFormer-X not only performs well in ideal rural scenes but also maintains robustness in dense urban domains. Notably, class-level improvements are significant. For example, ViT struggles with class 10 “Yellow Curb” in MUUFL, achieving only $31.99\%$, while FusionFormer-X improves this to $36.97\%$—a substantial $+5\%$ gain. Similar trends hold for hard classes in Trento like low-height vegetation and shaded objects.

4.3. Ablation Studies on Trento

To investigate the architectural effectiveness of FusionFormer-X, we conduct targeted ablation studies on two components: (1) modality fusion strategy, and (2) multi-scale attention design.

• Multimodal Fusion Effectiveness. In Table 1, we compare the performance of FusionFormer-X trained with only HSI, only LiDAR, and both modalities. While HSI alone provides rich spectral cues, and LiDAR contributes spatial geometry, their fusion achieves the best results in every metric. Specifically, multimodal fusion improves OA by +2.62% over HSI-only and +8.89% over LiDAR-only, with similar improvements in AA and Kappa. This demonstrates the complementary nature of elevation and spectral information.
• Multi-Scale Self-Attention Variants.Table 2 evaluates the MSMHSA module under various scale settings. We observe that using a coarse-to-fine hierarchy (e.g., $16\times 16$, $4\times 4$, $2\times 2$) yields the best results. Adding too many scales may slightly degrade performance due to over-fragmentation, while using only one scale fails to capture both local and global interactions. This validates our hypothesis that hierarchical attention enhances fine-grained segmentation boundaries and maintains scene context.

4.4. Extended Quantitative Insights

Beyond aggregate scores, FusionFormer-X demonstrates superior class-specific behavior. For instance, it effectively distinguishes classes with overlapping spectral distributions by leveraging spatial priors from LiDAR. In MUUFL, the model correctly segments “Concrete” and “Painted Metal” regions which are often confused by CNN2D and ViT. FusionFormer-X also shows resilience in minority classes with few pixels, an essential trait for real-world deployment.

The margin of improvement in AA (class-wise average) over SpectralFormer is over +10% on Trento, suggesting better representation generality. This is crucial since SpectralFormer, while effective in 1D spectral modeling, lacks spatial fusion flexibility.

4.5. Visual Quality Assessment

While visualizations are not presented here, we report qualitative findings. FusionFormer-X generates significantly smoother classification maps, especially around object boundaries. Unlike RF and CNN2D that often produce noisy or blocky predictions, our model preserves structural continuity, owing to the multi-scale receptive fields and token mixing.

Even in shadowed or occluded regions, the model maintains high confidence predictions, an outcome likely attributable to the positional-aware encoding and spatial-spectral fusion at multiple scales.

4.6. Inference Robustness and Repeatability

We further evaluate the statistical robustness of our model across random initializations. FusionFormer-X exhibits lower standard deviations in OA, AA, and Kappa than any baseline. For example, on Trento, the OA variance is less than $\pm 0.02\%$, confirming that our architecture is stable during training. This is critical for downstream applications requiring repeatability (e.g., environmental monitoring).

4.7. Summary and Takeaways

In summary, FusionFormer-X achieves superior results on both Trento and MUUFL benchmarks, validating its design through both accuracy and stability. Key factors driving performance include: (1) convolutional tokenization for spatial context, (2) hierarchical attention for resolution-aware learning, and (3) modality-aligned fusion through cross-modal embedding.

Our model not only outperforms state-of-the-art alternatives in raw accuracy but also delivers smoother predictions, better generalization to rare classes, and consistent outcomes across training runs, highlighting its promise for real-world multimodal remote sensing applications.

5. Conclusion and Future Directions

In this work, we propose a novel multimodal transformer framework, termed FusionFormer-X, specifically designed to enhance remote sensing (RS) image classification by effectively integrating complementary modalities—Hyperspectral Imaging (HSI) and Light Detection and Ranging (LiDAR). The core objective of FusionFormer-X is to harness both the fine-grained spectral discrimination power of HSI and the structural elevation cues provided by LiDAR, facilitating a more robust and comprehensive scene understanding.

Our architecture introduces a carefully crafted Multi-scale Multi-Head Self-Attention (MSMHSA) module, which enables hierarchical fusion across different spatial resolutions. This design alleviates the resolution mismatch between HSI and LiDAR modalities and empowers the network to capture both global dependencies and local contextual details. Furthermore, we integrate convolutional layers into the tokenization and projection stages, thereby embedding local inductive biases that are essential for preserving fine spatial structures in RS imagery. This hybrid design bridges the gap between CNNs and pure Transformers, achieving an optimal trade-off between computational efficiency and classification accuracy.

Extensive experiments conducted on two widely recognized benchmarks—Trento and MUUFL—validate the effectiveness of our approach. FusionFormer-X consistently outperforms prior methods, including both convolution-based networks and recent transformer-based fusion frameworks. The superior performance across Overall Accuracy (OA), Average Accuracy (AA), and Kappa coefficient ($\kappa$) confirms the robustness and generalizability of our model in both rural and urban RS scenarios. Additionally, ablation studies confirm the critical contribution of both the multimodal fusion strategy and the hierarchical attention mechanism.

Future Work. While FusionFormer-X demonstrates strong capabilities, several directions remain open for future exploration. One natural extension is to incorporate additional modalities beyond LiDAR and HSI, such as Synthetic Aperture Radar (SAR), RGB imagery, or thermal data, which could further enrich the semantic context and enhance classification performance in more diverse or adverse conditions.

Another promising direction involves adapting our framework for dynamic or temporal RS tasks, such as change detection, multi-temporal land use monitoring, or disaster assessment. This could involve extending FusionFormer-X into a temporal multimodal transformer with recurrent or attention-based temporal modeling capabilities.

Moreover, future efforts could explore lightweight variants of FusionFormer-X for deployment in real-time or edge-based RS systems. Techniques such as model pruning, quantization, and knowledge distillation could be applied to reduce inference latency while maintaining classification accuracy.

Finally, integrating uncertainty quantification and active learning into FusionFormer-X could make it suitable for semi-supervised or low-label RS tasks, thereby reducing reliance on costly annotated data.

In conclusion, FusionFormer-X presents a principled and scalable solution for multimodal remote sensing classification. Its flexible design, strong empirical performance, and potential for cross-modal generalization pave the way for broader adoption in real-world geospatial applications.