CSMoE: Cross-Sensor Mixture-of-Experts

• Cross-Sensor Mixture-of-Experts (CSMoE) is a remote sensing model that integrates soft MoE into CSMAE, enabling shared cross-sensor representation with modality-specific specialization.
• It employs adaptive routing mechanisms that dynamically allocate specialized experts to handle sensor-induced heterogeneity in tasks such as land cover classification, 3D segmentation, and geo-localization.
• The architecture balances soft and sparse routing strategies to maximize computational efficiency and robust domain generalization compared to traditional models.

Cross-Sensor Mixture-of-Experts (CSMoE) denotes both a specific remote sensing foundation model and a broader architectural paradigm in which heterogeneous sensor, platform, or domain inputs are processed by specialized experts under learned routing. In the remote sensing formulation, CSMoE is obtained by integrating the Soft mixture-of-experts mechanism into the Cross-Sensor Masked Autoencoder (CSMAE), so that modality-specific expert specialization is combined with shared cross-sensor representation learning (Hackel et al., 17 Sep 2025). Closely related systems use sparse, routed, or hierarchical MoE designs to address spectral shift in multispectral land cover classification, cross-domain generalization in 3D point cloud segmentation, zero-shot fault diagnosis across equipment and sensor types, cross-platform geo-localization, and cross-subject fMRI decoding (Chen et al., 20 May 2025, Chen et al., 29 May 2025, Wang et al., 6 May 2026).

1. Problem setting and scope

Across the literature, the motivating condition is sensor-induced heterogeneity. In multispectral land cover classification, the dominant issue is “spectral shift,” arising from cross-sensor disparities and cross-geospatial disparities (Chen et al., 20 May 2025). In 3D perception, point clouds are captured by diverse sensors such as depth cameras and LiDAR, producing different scanning patterns, sampling densities, and semantic biases; these differences make unified large-scale training difficult, especially when domain labels are inaccessible at inference time (Chen et al., 29 May 2025). In mechanical equipment diagnosis, domain shift across equipment and operating conditions limits generalization, motivating architectures intended for zero-shot deployment on unseen domains and sensor types (Wang et al., 6 May 2026). In cross-modal geo-localization, severe heterogeneity arises across satellite, drone, and ground imagery, with an additional domain gap between generic training descriptions and platform-specific test queries (Li et al., 23 Oct 2025).

The resulting design goal is not merely multi-domain robustness, but adaptive computation. Instead of fitting a single dense model to all sensor regimes, these systems divide processing across experts and learn routing rules that activate, weight, or fuse experts according to the input. In the remote sensing foundation model named CSMoE, this objective is stated as enhancing efficiency while maintaining or improving representational performance; in Point-MoE and YOTOnet, it is framed as automatic specialization without requiring domain labels or external meta-data at test time (Hackel et al., 17 Sep 2025, Chen et al., 29 May 2025, Wang et al., 6 May 2026).

System	Heterogeneity addressed	Expert mechanism
CSMoE	cross-sensor remote sensing modalities	Soft MoE in modality-specific and cross-sensor encoders
Land-MoE	cross-sensor and cross-geospatial MLCC	Mixture of Low-Rank Token Experts with Frequency-Aware Filters
Point-MoE	RGB-D, synthetic data, and LiDAR	top-k routed MoE layers inside PTv3
YOTOnet	different equipment and sensor types	Domain-Conditioned Sparse Experts

2. Routing regimes and expert specialization

The defining operation in CSMoE-style systems is routing. In YOTOnet, the Domain-Conditioned Sparse Experts (DC-MoE) module takes a domain-agnostic feature vector $z$ from the Invariant Feature Distiller, computes expert probabilities with

$\mathbf{p} = \mathrm{softmax}(W_g z),$

selects only the top-$k$ experts via a Top-$k$ operation, and forms the output as

$y = \sum_i m_i\, p_i\, h_i(z),$

where $h_i(\cdot)$ is the $i$-th expert network (Wang et al., 6 May 2026). The gating is conditioned on the input signal itself, not on external domain metadata, so routing is signal-driven rather than label-driven.

Point-MoE adopts a similar sparse principle within Point Transformer V3: each MoE layer contains $N$ independent expert MLPs, and a lightweight router selects a sparse subset of experts using a top-$k$ routing strategy, with the MoE output defined as

$$\mathrm{MoE}(\mathbf{x}) = \sum_{i \in \mathcal{S}_\mathbf{x}} G_i(\mathbf{x}) f_i(\mathbf{x}),$$

where $\mathbf{p} = \mathrm{softmax}(W_g z),$0 indexes the top-$\mathbf{p} = \mathrm{softmax}(W_g z),$1 experts for feature $\mathbf{p} = \mathrm{softmax}(W_g z),$2 (Chen et al., 29 May 2025). The paper reports that standard point cloud backbones degrade significantly on mixed-domain data, whereas this routing strategy can automatically specialize experts even without access to domain labels.

By contrast, the remote sensing model CSMoE employs Soft MoE rather than hard sparse routing. Its two-stage mechanism first forms slots from weighted combinations of tokens,

$\mathbf{p} = \mathrm{softmax}(W_g z),$3

then routes each slot to a specific expert, and finally redistributes expert outputs back to tokens (Hackel et al., 17 Sep 2025). This soft formulation is presented as allowing efficient parallelization and better utilization while preserving high capacity.

Other variants change the conditioning variable rather than the basic expert principle. In the geo-localization PE-MoE system, a 2-layer MLP takes the shared text feature as input and outputs platform weights over satellite, drone, and ground experts, with final retrieval similarity computed as a dynamically weighted sum of expert scores (Li et al., 23 Oct 2025). In MoRE-Brain, routing is hierarchical and dual-stage: voxel-to-expert assignment is learned in the encoder, then a Time Router selects hierarchical granularity across diffusion timesteps and a Space Router integrates expert embeddings conditioned on the current noisy latent (Wei et al., 21 May 2025). This suggests that, within the CSMoE paradigm, routing has become the principal mechanism for absorbing heterogeneity at token, instance, query, voxel, or timestep level.

3. The remote sensing foundation model named CSMoE

The paper “CSMoE: An Efficient Remote Sensing Foundation Model with Soft Mixture-of-Experts” defines the term most explicitly at the foundation-model level (Hackel et al., 17 Sep 2025). Its starting point is CSMAE, a dual-encoder architecture for remote sensing images in which each input is partitioned into patches, a subset is masked for self-supervised masked image modeling, modality-specific encoders capture sensor-specific content, and a cross-sensor encoder maps both modalities into a shared latent space. CSMoE replaces standard feed-forward networks in Transformer layers with Soft MoE blocks.

Architecturally, each input first passes through modality-specific encoder layers with Soft MoE blocks, enabling expert specialization within each modality. The outputs are then processed by a shared cross-sensor encoder, again composed of transformer-MoE layers, in order to promote cross-modal pattern discovery. Reconstruction is handled by modality-specific decoders without MoE. The model also includes slot repulsion loss, $\mathbf{p} = \mathrm{softmax}(W_g z),$4, to reduce redundancy among slots, and entropy loss, $\mathbf{p} = \mathrm{softmax}(W_g z),$5, to promote uniformly distributed expert usage (Hackel et al., 17 Sep 2025).

A second distinctive component is the thematic-climatic descriptor-driven sampling strategy used to construct the pretraining set. The method assigns each remote sensing image a tuple of climatic and thematic descriptors, then performs entropy-maximizing stratified sampling within each descriptor pair, using a genetic algorithm to maximize the entropy of pairwise geodesic distances. The stated purpose is to obtain a training set that is thematically, climatically, and spatially diverse while minimizing redundancy (Hackel et al., 17 Sep 2025).

The principal quantitative framing is the capacity-to-compute ratio,

$\mathbf{p} = \mathrm{softmax}(W_g z),$6

used to compare efficiency across remote sensing foundation models. In the reported comparison, CSMoE has 271M parameters, 10.11B FLOPs, and $\mathbf{p} = \mathrm{softmax}(W_g z),$7, whereas CSMAE has 87M parameters, 5.64B FLOPs, and $\mathbf{p} = \mathrm{softmax}(W_g z),$8 (Hackel et al., 17 Sep 2025). The paper states that, on average, CSMoE achieves more than twice the computational efficiency of existing remote sensing foundation models while maintaining competitive performance across scene classification, semantic segmentation, and content-based image retrieval. Reported downstream results include 66.5% mAP on m-bigearthnet, 86.2% AA on m-eurosat, 48.0% AA on m-so2sat, 94.3% AA on m-brick-kiln for CSMoE with $\mathbf{p} = \mathrm{softmax}(W_g z),$9, and up to 59.4% IoU on m-cashew-plant for $k$0 (Hackel et al., 17 Sep 2025).

4. Multispectral land cover classification and 3D scene understanding

Land-MoE extends the CSMoE logic to multispectral land cover classification by coupling a Frequency-aware Mixture of Low-rank Token Experts (MoLTE) with Frequency-Aware Filters (FAF) (Chen et al., 20 May 2025). The adapters are inserted after every Vision Foundation Model layer, while the backbone remains frozen. MoLTE uses top-$k$1 noisy routing at the token level, diverse low-rank token experts of varying ranks, and a shared MLP to synthesize token adjustments. FAF then converts the refined features to the frequency domain with RFFT, applies a learnable elementwise frequency mask, and returns to the spatial domain with RIFFT. The paper states that this design enhances semantic frequency components and suppresses frequency noise irrelevant to the task.

The reported performance is explicitly cross-sensor and cross-geospatial. On the Five-Billion-Pixels cross-sensor benchmark, Land-MoE achieves 64.4 mIoU, compared with 53.4 for frozen DINOv2, 55.7 for SET, 59.1 for Rein, and 56.6 for FADA. On the cross-geospatial setting, Land-MoE reaches 57.2 mIoU (Chen et al., 20 May 2025). It is also described as parameter-efficient, requiring only 2–4M additional parameters for MoLTE and FAF, and as robust across CLIP, SAM, EVA02, and DINOv2 backbones.

Point-MoE translates the same specialization principle to 3D semantic segmentation (Chen et al., 29 May 2025). It replaces the encoder and decoder feed-forward layers of Point Transformer V3 with MoE layers and uses top-$k$2 routing, typically with $k$3. The system is evaluated on indoor datasets including ScanNet, S3DIS, Structured3D, and Matterport3D, and on outdoor datasets including nuScenes, SemanticKITTI, and Waymo. These cover RGB-D sensors, synthetic data, and LiDAR. The paper states that Point-MoE outperforms strong multi-domain baselines and generalizes better to unseen domains; on Matterport3D, which is treated as unseen, Point-MoE achieves the best mIoU. It further reports that domain-agnostic MoE matches or exceeds domain-aware MoE, and that domain randomization improves robustness to new sensors. The analysis of expert routing via Jensen-Shannon Divergence shows that decoder-layer routing is strongly domain-specific, encoding implicit domain recognition. This provides an explicit instance of the “CSMoE paradigm” for scalable, unified 3D scene understanding (Chen et al., 29 May 2025).

5. Industrial, cross-platform, and neuro-inspired adaptations

In industrial fault diagnosis, YOTOnet combines an Invariant Feature Distiller, Domain-Conditioned Sparse Experts, and a dual-head classification system with auxiliary supervision (Wang et al., 6 May 2026). The Invariant Feature Distiller uses multi-scale dilated convolutions and FFT-based time-frequency fusion to produce “physical tokens,” after which DC-MoE routes samples to specialized lightweight feed-forward experts. The model is evaluated on five public bearing datasets—CWRU, MFPT, XJTU, OTTAWA, and HUST—through 30 cross-dataset protocols. Reported results include an F1 of 0.9862 on MFPT, 0.9065 on OTTAWA, 0.7485 on XJTU, and 0.5821 on HUST in the selected comparison table. Average test F1 improves from 0.5339 with 1 training dataset to 0.705 with 4 datasets, and ablations show that removing DC-MoE gating, replacing it with random routing, or averaging expert outputs performs worse than the full sparse, domain-conditioned gating design (Wang et al., 6 May 2026).

In cross-modal geo-localization, a parameter-efficient MoE framework addresses inter-platform heterogeneity across satellite, drone, and ground imagery (Li et al., 23 Oct 2025). The system uses frozen pre-trained encoders—BGE-M3 for text and EVA-CLIP for image—together with three platform experts trained with a progressive two-stage, hard-negative mining strategy. At inference, a dynamic gating network receives the shared text feature and outputs platform weights; the final similarity score is

$k$4

The accompanying preprocessing pipeline includes platform-wise partitioning, satellite augmentation, removal of orientation words, and an LLM-based caption refinement pipeline to align textual semantics with platform-specific visual characteristics. The paper reports that the system tops the official leaderboard in RoboSense 2025 Track 4 (Li et al., 23 Oct 2025).

MoRE-Brain provides a neuro-inspired formulation in which experts are organized hierarchically and shared across subjects, while routers are subject-specific (Wei et al., 21 May 2025). The encoder assigns voxels to experts at each level via trainable routers, and the experts map fMRI into frozen CLIP space. A finetuned SDXL model then synthesizes images under dual-stage Time and Space routing. For cross-subject transfer, the experts are frozen and only the routers are fine-tuned on a new subject; the paper states that robust cross-subject decoding is obtained with minimal data, as little as 2.5% of a subject’s data. The paper explicitly notes that the design is applicable to cross-sensor MoE scenarios, stating a general principle in which experts are agnostic to input channel while routers are sensor-specific (Wei et al., 21 May 2025).

6. Design tensions, misconceptions, and evolving directions

A recurring misconception is that MoE-based cross-sensor systems require explicit domain or sensor labels at inference time. Several of the reported systems reject that assumption. YOTOnet conditions gating on the signal itself rather than external domain metadata, Point-MoE is motivated by the realistic constraint that domain labels are typically inaccessible at inference time, and its domain-agnostic routing matches or exceeds domain-aware routing, while the geo-localization PE-MoE uses a query-adaptive gating network rather than a fixed platform assignment (Wang et al., 6 May 2026, Chen et al., 29 May 2025, Li et al., 23 Oct 2025).

A second misconception is that expert balancing is uniformly beneficial. The literature is not uniform on this point. YOTOnet includes a load-balancing regularization loss to avoid expert collapse, Land-MoE trains with an expert balancing loss that penalizes high variance in expert routing mass, and MoRE-Brain uses load balancing at each hierarchy level to encourage even expert utilization (Wang et al., 6 May 2026, Chen et al., 20 May 2025, Wei et al., 21 May 2025). Point-MoE, however, reports that removing the standard MoE load-balancing loss improves performance, likely due to domain and data imbalance in 3D scenarios (Chen et al., 29 May 2025). This divergence indicates that balancing interacts strongly with the structure of heterogeneity rather than constituting a universal recipe.

A third tension concerns sparse versus soft routing. Sparse top-$k$5 routing appears in YOTOnet, Land-MoE, and Point-MoE, where specialization and compute efficiency are tied to activating only a subset of experts (Wang et al., 6 May 2026, Chen et al., 20 May 2025, Chen et al., 29 May 2025). Soft routing defines the remote sensing CSMoE model, where every token contributes to slot formation and expert outputs are redistributed to all tokens, and the formulation is presented as offering a superior trade-off among representational capacity, accuracy, and computational efficiency (Hackel et al., 17 Sep 2025). A plausible implication is that “CSMoE” now refers less to a single routing rule than to a broader strategy: preserve shared representation learning across sensors while permitting expert specialization wherever heterogeneity is most acute.

An earlier antecedent in this trajectory is CameNN, which introduced softmax gating networks over transformer experts for image-text alignment and conversion-rate prediction in a multi-task recommender system (Li et al., 2020). Although not framed as cross-sensor learning, it already combined input-dependent expert selection, transformer-based experts, and cross-modal alignment tasks. The subsequent literature extends that logic to sensor-defined, platform-defined, and subject-defined heterogeneity, suggesting an ongoing convergence between efficient conditional computation and robust generalization across disparate data sources.