A dual-robust multi-modal information bottleneck remote sensing classification method for satellite-to-ground tasks

Abstract

The application discloses a kind of double robust multimodal information bottleneck remote sensing classification methods for star-ground task, its characteristics are that the method includes: source domain disturbance injection and modal specificity double-layer probability bottleneck coding, cross-modal fusion and variation semantic compression, differentiable additive white Gaussian noise channel simulation and receiving end task decoding and source domain sensitivity regular and double robust variation optimization based on Fisher information such as steps.Compared with prior art, the application has stability improvement under the premise of not increasing any on-board computing power, storage and delay, is especially suitable for satellite platform with severely limited inference resources, especially in the star-ground remote sensing scene where downlink bandwidth is limited, transmission is unreliable, and source domain disturbance and channel disturbance coexist, can significantly improve the accuracy and robustness of ground object classification, and all robustness mechanisms only take effect in the offline training phase, the computational graph, parameter amount, memory occupancy and latency in the inference phase are consistent with the baseline, without increasing any on-board inference overhead.

Description

Technical Field

[0001] This invention relates to the fields of multimodal remote sensing and mission-oriented semantic communication technology, and in particular to a dual-robust multimodal information bottleneck remote sensing classification method for satellite-ground missions. Background Technology

[0002] Earth observation increasingly relies on multimodal fusion of hyperspectral imagery and lidar to provide spectral and geometrically complementary clues for ground feature classification. Deploying such fusion in a space-ground collaborative system faces a fundamental contradiction: the onboard platform is simultaneously constrained by multiple factors including computing power, storage, energy consumption, bandwidth, and visible time window latency, while the downlink channel is time-varying and affected by fading and additive noise. The core challenge lies in maintaining reliable mission decisions when source domain perturbations (sensor noise, atmospheric interference, local contamination) and channel-side distortions (fading, additive Gaussian noise) coexist and interact in the extraction-fusion-transmission pipeline. Under this complex perturbation mechanism, source domain noise may be amplified during cross-modal fusion and further degraded by channel distortion before reaching the classifier, thus constituting the central problem addressed by this invention.

[0003] To address the aforementioned issues, existing research primarily follows four relatively independent paths. The first category comprises multimodal remote sensing fusion methods, ranging from handcrafted feature methods based on extended attribute profiles, morphological profiles, sparse and low-rank component analysis, and superpixel-level decision fusion, to deep end-to-end architectures such as spatial-spectral cross-modal enhancement networks, cross-scale attention transformers, and information bottleneck-guided multi-source fusion. While these methods continuously improve classification accuracy, they are mostly optimized for offline inference under clean input conditions. Their training objectives typically exclude communication constraints, semantic rate budgets, and transmission-side contamination, leading to a gap between the reported algorithm accuracy and the actual deployability of satellite-to-ground systems. The second category focuses on task- and semantic-oriented communication methods, including deep joint source-channel coding and edge inference communication based on information bottlenecks, prioritizing transmission for task utility rather than bit-level fidelity. However, these methods are mostly centered on channel adaptation and rarely explicitly model source domain perturbations jointly within a single optimization objective. This deficiency is particularly critical in multimodal remote sensing where complex perturbations propagate across extraction, fusion, compression, and transmission stages. The third category is information bottleneck methods for task representation learning. The information bottleneck principle describes representation learning as a tradeoff in mutual information: maximizing task relevance sufficiency. Minimize input redundancy at the same time The variational information bottleneck framework makes the objective trainable in deep networks through variational approximation and KL divergence regularization, and has been extended to multimodal settings. However, many existing implementations rely on idealized source domain assumptions or emphasize only a single perturbation side in optimization, and their robustness improvement is often incomplete when source domain uncertainty and channel degradation occur simultaneously. The fourth category is robust learning methods under dual perturbations, including adversarial training, degenerate Bruker optimization, and input gradient regularization. These methods can improve local stability, but most are geared towards single-stage input-to-output mapping and do not address multi-stage pipelines. Satellite-ground multimodal semantic communication constitutes a coupled multi-stage link: sensing and extraction, cross-modal fusion, semantic compression, channel transmission, and receiver inference; under this link, source domain perturbations and channel perturbations interact through nonlinear error propagation. Training based solely on channel awareness cannot suppress source domain-induced instability, and source domain regularization alone may not maintain transmission robustness.

[0004] In summary, existing methods have structural gaps at three levels: at the modeling level, no optimization objective simultaneously couples task discriminability, semantic compression, channel robustness, and source domain sensitivity within a single problem; at the robustness level, source domain and channel perturbations are treated independently, while they are actually coupled through nonlinear cross-modal error propagation; at the evaluation level, reporting accuracy only at a single noisy operating point fails to characterize the joint trade-off between accuracy, robustness, and communication overhead that governs actual deployability. Figure 3 As shown in Figure a, the traditional "transmit first, compute later" pipeline faces high overhead and performance collapse due to training-deployment mismatch under bandwidth constraints and dual perturbations of source domain and channel. In contrast, the task-oriented semantic feature downlink transmission paradigm requires a robust information bottleneck to compress task-irrelevant redundancy while preserving task-related information. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a robust, multimodal remote sensing classification method for satellite-to-ground missions that addresses information bottlenecks. This method employs a two-stage information bottleneck approach, prioritizing single-modality integration followed by fusion. It utilizes an optimization framework that couples source-domain robustness and channel robustness, unifying semantic compression, channel-robust task inference, and source-domain sensitivity suppression under a single variational optimization objective. This effectively prevents single-modality input noise from contaminating shared representations, achieving a better balance between denoising and information preservation. This method achieves improved stability without increasing onboard computing power, storage, or latency, making it particularly suitable for satellite platforms with severely limited inference resources. Especially in satellite-to-ground remote sensing scenarios characterized by limited downlink bandwidth, unreliable transmission, and the coexistence of source-domain and channel disturbances, it significantly improves the accuracy and robustness of ground feature classification. Furthermore, all robustness mechanisms are effective only during the offline training phase; the computation graph, parameter count, memory usage, and latency during the inference phase remain consistent with the baseline, without increasing onboard inference overhead.

[0006] The specific technical solution to achieve the purpose of this invention is: a robust multimodal remote sensing classification method for satellite-to-ground missions, characterized by the following steps:

[0007] S1, Source Domain Perturbation Injection and Mode-Specific Two-Layer Probabilistic Bottleneck Coding

[0008] For including hyperspectral imaging and lidar One co-registered modal input, based on coverage Additive Gaussian noise is injected into a randomly selected subset of inputs within each mini-batch to simulate source domain perturbation, while unselected positions remain unchanged. Each perturbated mode is then input into its corresponding mode-specific probabilistic encoder. Each encoder outputs the mean and logarithmic variance of a diagonal Gaussian distribution, and the single-mode bottleneck characterization of that mode is obtained through reparameterized sampling.

[0009] S2, cross-modal fusion and variational semantic compression

[0010] The single-mode bottleneck representations are aggregated into a joint representation by a cross-modal fusion operator. The joint representation is then mapped to diagonal Gaussian distribution parameters by a fusion encoder and obtained by reparameterized sampling to obtain the semantic code to be transmitted via the satellite-to-ground downlink.

[0011] S3. Differentiable Additive White Gaussian Noise Channel Simulation and Receiver-End Decoding

[0012] Channel noise is injected into the semantic code through a differentiable additive white Gaussian noise channel simulation layer whose noise variance is controlled by peak signal-to-noise ratio, resulting in a disturbed semantic code at the receiver. This code is then input into the receiver's main decoder to output ground feature classification prediction. During the training phase, an auxiliary decoder that acts on the bottleneck representation of each single mode also outputs predictions to provide intra-modal sufficiency supervision. The channel simulation layer contains no trainable parameters and is differentiable, allowing gradients to be backpropagated from the receiver loss to each encoder at the satellite end.

[0013] S4. Source Domain Sensitivity Regularization and Double Robust Variational Optimization Based on Fisher Information

[0014] With the goals of maximizing label-related information, suppressing representation redundancy, and suppressing source domain sensitivity, the aforementioned end-to-end pipeline is modeled as a single information-theoretic optimization objective. The task-related term is evaluated on the perturbed representation at the receiver to internalize channel robustness. Unsolvable mutual information terms are replaced with cross-entropy surrogate terms and KL divergence closed-form terms. The source domain sensitivity regularization term is instantiated as the expectation of the encoder's log-likelihood with respect to the squared norm of its input, applied to all single-mode encoders and fused encoders. Under the channel robustness setting (… Training is performed using only the receiver's perturbation representation as the basis for task supervision, under a dual robust configuration ( The source domain sensitivity regularization term is further activated; finally, the trained encoder-transmitter-decoder network is output for land cover classification of the multimodal remote sensing data to be classified.

[0015] The source domain perturbation injection and single-mode probabilistic bottleneck coding in step S1 specifically include:

[0016] S11, Source Domain Disturbance Injection

[0017] The source domain degradation is modeled as additive Gaussian noise applied to a randomly selected subset of the input. For the ... The disturbed input is generated for each input position using the following formula:

[0018] .

[0019] in, For the original modal input, For the disturbed input, By coverage within each mini-batch Uniformly sampled set of disturbed locations, indicator function Indicates only The location inside receives noise. It is additive Gaussian noise. Indicates a Gaussian distribution. The identity matrix is ​​used; the sample-by-sample noise variance is normalized by the input power and is expressed as follows:

[0020] .

[0021] in, For sample-by-sample input power, For input dimensions, This is the sample-by-sample perturbation intensity factor. for Norm; this modeling will affect the perturbation coverage. With sample-by-sample intensity Decoupling allows robustness variations observed during scan coverage to be attributed to inconsistencies in the perturbation range rather than the noise scale; the source domain perturbation is used only during the training phase to construct the learning objective, and the input during the inference phase is clean data without any perturbation.

[0022] S12, Single-modal probabilistic bottleneck coding

[0023] Each modality is encoded by an independent probabilistic backbone network, which consists of a convolutional feature extractor and two linear heads, each outputting a mean. With log variance The variance head employs softplus activation with a small negative bias to prevent variance collapse and ensure numerical stability of the KL calculation. After reparameterization sampling, the single-modal bottleneck characterization is obtained as shown in the following equation. :

[0024] .

[0025] in, For the first Input of a modality, and The first The mean function and standard deviation function of the encoder output, For element-wise multiplication, To obtain from the standard normal distribution Random noise vector sampled in the middle, It is an identity matrix; by first suppressing redundancy and noise variation in each mode through the bottleneck of the single-mode layer, the source domain disturbance is prevented from being fully amplified before entering the fusion stage.

[0026] The cross-modal fusion and variational semantic compression in step S2 specifically include: using a fusion operator The bottleneck characteristics of each single mode are aggregated into a joint characteristic. The fusion operator This is achieved by linear projection after channel splicing, and then the joint representation is obtained by a fusion encoder. The parameters are mapped to a diagonal Gaussian distribution and then reparameterized and sampled to obtain the semantic code shown in the following equation:

[0027] .

[0028] in, For the fused semantic code to be transmitted, For cross-modal joint characterization, and These are the mean and standard deviation of the output of the fusion encoder, respectively. For element-wise multiplication, The bottleneck is a random noise vector sampled from a standard normal distribution. The bottleneck adopts a two-level structure with single-mode first and fusion later: the single-mode bottleneck first suppresses intra-mode interference variation, and the fusion bottleneck then removes the cross-mode redundancy that remains after complementary semantic aggregation, thereby preventing the input noise of any single mode from polluting the shared representation before downlink transmission under the dual perturbations of the source domain and the channel.

[0029] The channel simulation and receiver decoding in step S3 specifically include: modeling the satellite-to-ground downlink as an additive white Gaussian noise channel, and injecting channel noise into the semantic code to obtain the receiver-disturbed semantic code as shown in the following formula:

[0030] .

[0031] in, For the received end's scrambled semantic code, For the transmitting end semantic code, The channel is additive white Gaussian noise. For channel noise variance, The identity matrix is ​​used; channel quality is parameterized by the peak signal-to-noise ratio (PSNR) and expressed by the following formula:

[0032] .

[0033] in, Peak signal-to-noise ratio (in dB). The peak power of semantic code features calculated in mini-batches. This indicates taking the maximum value of each dimension of the semantic code. For semantic code number Each dimension component Let V be the channel noise variance; for a given target PSNR, the channel noise variance is... The receiver's main decoder takes the scrambled semantic code as input and outputs the class log probability. , To predict labels, the training phase also includes single-mode auxiliary decoders. Output prediction enhances intramodal sufficiency; during the inference phase, only the master decoder is evaluated, and each encoder outputs a mean deterministic output. The channel is transparent and no noise is injected.

[0034] The unified variational optimization objective and its trainability in step S4 specifically include:

[0035] S51, Unified Information Theory Goal

[0036] remember To obtain mutual information, solve for encoder parameters that simultaneously make the representation compact, task-discriminable, and robust to both channel and source domain perturbations. With decoder parameters The optimization objective can be expressed as follows:

[0037] .

[0038] in, For the random variable of land cover category labels, For the random variable of the disturbed semantic code at the receiving end, and These are random variables representing the fused semantic code and the fused representation, respectively. and The first Bottleneck representation of each modality and input random variables, and These are the trainable parameters of the encoder and decoder, respectively; the first set of task-related terms is represented by the perturbation at the receiver. The above evaluation internalizes channel robustness into the optimization; the compression term is composed of coefficients. Controlled compression—a trade-off of discriminativity; source domain sensitivity regularization From coefficients control: Degraded to channel robust setting, Activate dual robust training.

[0039] S52, Cross-Entropy Proxy and KL Closed-End Compression Item

[0040] A variational decoder is introduced to provide a lower bound for task-related terms, and the upper bound for negative mutual information is the cross-entropy proxy (constant entropy is omitted). ,in (Information entropy of label distribution)

[0041]

[0042] in, Represents the mathematical expectation. For parameters Variational decoders with trainable weights for class Conditional probability prediction, For the disturbed semantic code sample; introduce a fixed prior. (Standard normal distribution) With an upper bound on the compression term, the KL divergence between the diagonal Gaussian encoder and the standard normal prior has the following closed-form equation:

[0043] .

[0044] in, For parameters The conditional distribution of the output of the probability encoder. For the standard normal prior, Traversing the components of each dimension of the bottleneck representation ( , (as a bottleneck dimension) and The encoder pairs the first The mean and variance of the output in each dimension; this closed-form expression only involves element-wise operations on the encoder output, without sampling or matrix operations, and has a complexity of O(n log n). ( (For the bottleneck dimension); reparameterization is used throughout. ( This enables differentiable sampling, allowing the gradient to be back-propagated from the decoder loss to the encoder parameters without the need for high-variance scoring function estimation.

[0045] S53. Reconstructability of Fisher's Information Source Domain Regularization Terms

[0046] The expectation of traces Instantiated using the Fisher information matrix in the input space, the probability encoder is... The Fisher information matrix is ​​expressed by the following formula:

[0047] .

[0048] in, Given the Fisher information matrix in the input space, Indicates about input gradient operator, For the log-likelihood of the encoder, Represents the transpose of a vector or matrix; Definition Expectations for its traces ,in (For matrix trace operations); using identities Given the linearity of the trace and expectation, it can be simplified to the expectation of the log-likelihood with respect to the squared norm of the gradient of the input, as shown in the following equation:

[0049]

[0050] Therefore, there is no need to construct, store, or invert any... Fisher matrix, This can be obtained with a single automatic differentiation call; the regularization term is instantiated in the fusion encoder and each single-mode encoder respectively:

[0051] ;

[0052] .

[0053] in, For the Fisher regularization term of the fusion encoder, For fusion characterization gradient, Conditional distribution for the fusion encoder; For the first Fisher regularization terms for a single-mode encoder, For the reason about the Modal input gradient, For the first Conditional distribution of a single-mode encoder, The total number of modes; it is crucial to apply the regularization term uniformly to all encoders: if only the fused encoder is regularized, disturbances within each individual mode may still be amplified by the individual mode backbone before reaching the fusion stage; uniform coefficient Apply consistent sensitivity suppression to the entire coding chain.

[0054] S54, Final Trainable Loss

[0055] Combining the cross-entropy-KL proxy with the Fisher regularization term, we obtain the complete trainable loss as shown in the following equation:

[0056] ;

[0057] ;

[0058] .

[0059] in, The main branch loss consists of the receiver cross-entropy surrogate term (first term) and the fusion encoder KL compression term (second term). Master decoder, For semantic codes affected by channel perturbation, To compress weights; For the first A single-modal auxiliary branch loss, For the first One auxiliary decoder; For complete dual-robust multimodal information bottleneck loss, and They are respectively the fusion encoder and the first Fisher source domain regularization terms for a single-mode encoder For source domain sensitivity regularization weights, Traverse all One mode; where the channel robust setting ( ) and dual robust setting ( The only difference is whether the last two Fisher regular expression terms are activated; the rest of the architecture, optimization configuration, data partitioning, and evaluation protocol are completely identical, so any performance difference can be attributed to the Fisher regular expression terms.

[0060] In step S4, the tradeoff coefficients are... and A two-stage sequential tuning strategy is adopted: Stage 1 in Below, select via grid search. A stable compression-precision operating point is established under the target training PSNR, and then fixed. Phase Two in Fixed Under the conditions On the logarithmic scale The algorithm performs an upscan and uses a linear warm-up from 0 to the target value in the first few rounds to prevent the Fisher term from dominating the optimization before the latent variable distribution stabilizes; where If the value is too small, compression will be insufficient and downlink bandwidth will be wasted. If the size is too large, it will excessively smooth out bottlenecks and discard discriminative information; If the value is too small, the encoder will be sensitive to input disturbances, and source domain noise will be amplified along the pipeline. If the value is too large, the encoder tends to become a constant mapping, and the class separability collapses; the two-stage tuning utilizes and The two compromises have approximately orthogonality, without needing to... A joint grid search can achieve stable convergence.

[0061] The source domain perturbation injection is a pure training phase mechanism. Under the dual robust setting, the Fisher regularization term... Calculated through the following steps for each encoder branch: The input tensor is set to the required gradient, the distribution parameters are calculated and the bottleneck code is obtained through reparameterized sampling, and the bottleneck code is evaluated based on the diagonal Gaussian log-likelihood. Its gradient with respect to the input is obtained through automatic differentiation. and take During training, the global gradient is thresholded. Pruning is performed to suppress gradient explosion; during the inference phase, each encoder performs a deterministic forward pass with the mean as the output, without sampling, injecting channel noise, or evaluating any Fisher terms, so that... and The trained model has the same computational graph, number of parameters, memory usage, and latency during inference.

[0062] Compared with the prior art, the present invention has the following beneficial technical effects and significant technical progress:

[0063] 1) This invention unifies semantic compression, channel robust task inference, and source domain sensitivity suppression into a single variational optimization objective. It optimizes task relevance, semantic code rate control, channel robust inference, and source domain sensitivity suppression as coupling dimensions. Compared with existing methods that process the above dimensions separately, the objective of this invention explicitly applies a joint trade-off. Its structure exposes the inherent coupling between classification quality and bilateral robustness, which is more in line with the actual constraints of limited bandwidth, transmission reliability, and disturbance tolerance in satellite-to-ground systems.

[0064] 2) This invention employs a two-level information bottleneck approach: single-modality bottleneck first, followed by fusion. The single-modality bottleneck suppresses intramodal redundancy and noise variation, preventing source domain disturbances from being amplified in the cross-modal fusion process. The fusion-level bottleneck then removes the cross-modal redundancy remaining after complementary semantic aggregation. Compared to a scheme that applies a single-level bottleneck only after fusion, this two-level design can more effectively prevent single-modal input noise from contaminating the shared representation under dual perturbations, achieving a better balance between denoising and information preservation.

[0065] 3) Starting from the expectation of the trace of the Fisher information matrix in the input space, this invention uses the trace-norm identity to reconstruct the source domain sensitivity regularization term into the expectation of the squared gradient norm of the encoder's log-likelihood with respect to the input. It does not require the construction, storage, or inversion of any high-dimensional Fisher matrix throughout the entire process; it can be obtained by a single automatic differentiation. It is compatible with the standard automatic differentiation framework. This regularization term applies uniformly to all single-mode encoders and fusion encoders, suppressing disturbance amplification at each encoding stage.

[0066] 4) This invention calculates the task loss at the receiver after channel disturbance. Non-clean coding representation Above all, receiver alignment supervision directly improves decision reliability under channel degradation, achieves a smoother accuracy degradation curve and stronger stability in the low peak signal-to-noise ratio range, and achieves higher task-related semantic density per unit bit under a fixed transmission budget.

[0067] 5) All robustness mechanisms of this invention (source domain perturbation injection, latent variable random sampling, and Fisher gradient calculation) are effective only during the offline training phase. During the inference phase, a deterministic encoding → transmission → decoding forward pass with the mean as the output is executed. Its computation graph, number of parameters, memory usage, and latency are exactly the same as the channel robustness baseline. Thus, bilateral robustness is implemented as a purely offline investment, and the deployed model achieves improved stability without increasing any onboard computing power, storage, or latency, making it particularly suitable for satellite platforms with severely limited inference resources. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the end-to-end framework of the dual robust information bottleneck of the present invention;

[0069] Figure 2 This is a schematic diagram of the dual robust training phase of the present invention;

[0070] Figure 3 This is a comparison diagram between the present invention and the semantic feature downlink transmission paradigm of transmitting first and then computing;

[0071] Figure 4 This is a comparison diagram of the architecture of the present invention and traditional pipelines;

[0072] Figure 5 This is a diagram of the four-objective joint optimization framework of the present invention;

[0073] Figure 6 This is a schematic diagram of the trainable path and reparameterization mechanism for the variational information bottleneck of this invention. Detailed Implementation

[0074] This invention employs a two-stage information bottleneck approach, prioritizing single-modality integration followed by fusion. Utilizing an optimization framework that couples source domain robustness and channel robustness, it unifies semantic compression, channel robust task inference, and source domain sensitivity suppression under a single variational optimization objective. Specifically, it includes the following steps:

[0075] Step S1: Source domain perturbation injection and mode-specific two-layer probabilistic bottleneck coding

[0076] For including hyperspectral imaging and lidar Each co-registered modal input is based on coverage. Additive Gaussian noise is injected into a randomly selected subset of the input to simulate source domain perturbations. After the diagonal Gaussian distribution parameters are output by a mode-specific probabilistic encoder, a single-mode bottleneck characterization is obtained through reparameterization sampling. Specifically, this includes:

[0077] S1-1: Source domain degradation is modeled as additive Gaussian noise applied to a random subset. ( ),in By coverage Each batch is sampled uniformly, and the noise variance for each sample is normalized according to the input power. ( ), to increase coverage With sample-by-sample intensity Decoupling: the perturbation is used only for training, and the input during inference is clean data.

[0078] S1-2: Each modality consists of a probabilistic backbone encoding formed by a convolutional feature extractor followed by two linear headers, which output the mean values ​​respectively. With log variance The variance head employs softplus activation with a small negative bias to prevent variance collapse, and single-modal bottleneck characterization is obtained through reparameterized sampling. ( ).

[0079] Step S2: Cross-modal fusion and variational semantic compression

[0080] The two-stage bottleneck approach, with single-mode input first and fusion second, first suppresses intra-mode interference and then removes cross-mode redundancy to prevent input noise from any mode from contaminating the shared characterization before downlink transmission. Specifically, this includes:

[0081] S2-1: Characterizing the bottleneck of each single mode By fusion operator After being spliced ​​together by channels, the composite representation is obtained by linear projection. ;

[0082] S2-2: Joint characterization The semantic code to be transmitted is obtained by mapping the parameters to a diagonal Gaussian distribution using a fusion encoder and then reparameterizing the samples. .

[0083] Step S3: Differentiable additive white Gaussian noise channel simulation and receiver task decoding

[0084] S3-1: Model the downlink as an additive white Gaussian noise channel ( ),

[0085] Channel quality is determined by peak signal-to-noise ratio. Parameterization, where For batch calculation of special

[0086] Peak power Let V be the channel noise variance. For a given target's PSNR, the channel noise variance is: The additive white Gaussian noise channel It contains no trainable parameters and is differentiable, allowing the gradient to be backpropagated from the receiver loss to the satellite encoder;

[0087] S3-2: The receiver's master decoder outputs the scrambled semantic code for classification and prediction. During the training phase, each single-mode auxiliary decoder... Provides in-modal sufficiency supervision.

[0088] Step S4: Source Domain Sensitivity Regularization and Double Robust Variational Optimization Based on Fisher Information

[0089] S4-1: Using a unified variational objective to evaluate the bottleneck loss of the encoder and decoder for dual robust multimodal information. Perform the joint optimization shown in the following formula:

[0090] .

[0091] Among them, the main branch loss The task items are evaluated on the receiver's disturbance characterization to internalize channel robustness, and the loss of each single-mode auxiliary branch is considered. KL divergence diagonal Gaussian prior Having a closed form complexity .

[0092] S4-2: The source domain sensitivity regularization term is the expectation of the trace of the Fisher information matrix in the input space. Export, use Reconstructed as the expectation of the squared norm of the input gradient Therefore, no construction or inversion is required. Fisher matrix;

[0093] S4-3: Instantiate the Fisher regularization terms in the fusion encoder respectively. With each single-mode encoder To unify source domain sensitivity regularization weights Apply consistent sensitivity suppression throughout the entire coding chain;

[0094] S4-4: Set channel robustness to: The dual robustness setting is as follows: The only difference is whether the Fisher regular expression is activated; otherwise, they are completely identical. and A two-stage sequential tuning method is adopted, that is, first in Selected Fix it Post-scan and to Apply linear preheating and utilize the approximate orthogonality of the compromise between the two to achieve stable convergence.

[0095] To better understand this technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.

[0096] Example 1

[0097] This embodiment uses co-registered hyperspectral images and lidar data for urban land cover classification. The method provided by this invention is illustrated on three publicly available multimodal remote sensing benchmark datasets: the Houston2013 dataset (144 hyperspectral bands, single-band digital surface model, 15 classes, 15029 labeled pixels), the Augsburg dataset (180 hyperspectral bands, single-band digital surface model, 7 classes), and the Berlin dataset (244 hyperspectral bands, single-band digital surface model, 8 classes). This embodiment uses modal number... Fusion operator The implementation is a linear projection after channel splicing; all methods use consistent preprocessing (channel normalization, co-registration), training / testing partitioning and evaluation protocols.

[0098] See Figure 1The overall framework consists of a satellite-end coding pipeline and a ground-end decoding pipeline connected by a differentiable additive white Gaussian noise channel simulation layer. This framework shows the complete data flow of dual-modal probabilistic bottleneck coding, reparameterized sampling, cross-modal fusion and variational semantic compression, additive white Gaussian noise channel simulation and receiver decoding, and annotates the classification loss, KL regularization and Fisher source domain regularization terms for each branch.

[0099] See Figure 3 a. The traditional transmission paradigm of "transmit first, then compute": The satellite compresses and encodes the raw (pixel-level) multimodal remote sensing data and transmits it to the ground via the downlink. The data is then reconstructed and ground feature classification is performed at the ground. Under the conditions of limited downlink bandwidth and coexistence of source domain disturbances and channel noise, this paradigm faces performance collapse caused by high downlink overhead and training-deployment mismatch.

[0100] See Figure 3 b. The task-oriented semantic feature downlink transmission paradigm adopted in this invention: the satellite end extracts and compresses only the semantic features related to the ground object classification task after passing through the information bottleneck, and then transmits them downlink. While retaining task-related information, task-irrelevant redundancy is compressed. The ground decoder directly outputs the classification result based on the received semantic code. Thus, reliable task decision-making is maintained with lower downlink overhead under limited bandwidth and source domain-channel dual perturbation.

[0101] See Figure 4 a. Conventional fusion → transmission → decoding pipeline affected by source domain disturbances and channel noise.

[0102] See Figure 4 b. The present invention introduces Fisher source domain regularization and channel robust information bottleneck to form a dual robust framework. Compared with the conventional fusion → transmission → decoding pipeline which is affected by source domain disturbance and channel noise, the present invention introduces Fisher source domain regularization on the coding chain and constructs a dual robust target with channel robust information bottleneck.

[0103] S1, Source Domain Perturbation Injection and Mode-Specific Two-Layer Probabilistic Bottleneck Coding

[0104] See Figures 1-2 The satellite operates three modules sequentially, with the specific steps as follows:

[0105] 1) Source domain perturbation injection is performed on both the hyperspectral and lidar modes separately: within each training batch, perturbation is performed according to coverage. Uniformly extract the set of disturbed locations Additive Gaussian noise is applied to one of the locations. ( Sample-by-sample noise variance normalized by input power In this embodiment, the intensity is fixed for each sample. Coverage exist Up scan; this perturbation models situations where only some sensor readings are affected in a real-world scenario (such as local cloud cover or local sensor degradation), and is applied only during the training phase, with clean data as input during the inference phase.

[0106] 2) Each mode is encoded by a mode-specific probabilistic encoder. The backbone network consists of a convolutional feature extractor connected to two linear heads, which output the mean values ​​respectively. With log variance The variance header employs softplus activation with a small negative bias to constrain the variance from zero and ensure the stability of the KL calculation values; the single-modal bottleneck code is reparameterized and sampled. ( This is obtained. The single-mode bottleneck in the two-level bottleneck first suppresses intramodal redundancy and noise variation to prevent source domain disturbances from entering the fusion stage with full amplitude.

[0107] S2, Cross-modal fusion and variational semantic compression

[0108] Cross-modal fusion module The two single-mode bottleneck codes are concatenated by channel and then linearly projected to form a joint characterization. Fusion encoder The joint representation is compressed into diagonal Gaussian distributed parameters. The semantic code to be transmitted is obtained through reparameterized sampling. ( The fusion-level bottleneck removes residual cross-modal redundancy after complementary semantic aggregation. In this embodiment, the bottleneck dimension... The value should be no less than 128, and the encoder logarithmic variance should be constrained by softplus.

[0109] S3. Simulation of Differentiable Additive White Gaussian Noise Channel and Decoding at the Receiver

[0110] The channel is simulated by an additive white Gaussian noise layer with peak signal-to-noise ratio controlled variance, which injects noise into the semantic code. ( ),in To calculate the characteristic peak power in mini-batch, take the following for a given target PSNR: This layer contains no trainable parameters and is differentiable, allowing gradients to be backpropagated from the receiver loss to each encoder at the satellite end. Ground-end master decoder. Disturbed semantic codes Mapped to class log probabilities, auxiliary decoder During the training phase, the code is applied to a single-mode code to provide intra-mode supervision; the decoder is a lightweight two-layer multilayer perceptron. During the inference phase, only the master decoder is evaluated, and each encoder outputs a mean deterministic value. The channel is transparent.

[0111] S4. Source Domain Sensitivity Regularization and Dual Robust Variational Optimization Based on Fisher Information

[0112] See Figure 2 The forward calculation and loss structure are shown in three segments: satellite end, additive white Gaussian noise channel, and ground end. A comparison table is provided to give the channel robustness settings. ) and dual robust setting ( The differences in sampling, channel, Fisher regularization, and auxiliary decoder are optimized according to a unified variational objective during training.

[0113] See Figure 5 The invention presents four types of objective terms—task discrimination, transmission rate, channel noise, and source domain sensitivity—and their joint optimization objective function, categorized into an input layer, a semantic representation layer, and a task output layer. This invention models a semantic system for multimodal tasks as a joint optimization of four objective terms: task discrimination, transmission rate, channel noise, and source domain sensitivity. These correspond to the source domain sensitivity in the input layer, the compressed code rate in the semantic representation layer, the link noise risk in the channel layer, and the discrimination loss in the task output layer, respectively.

[0114] See Figure 6 Predicted and compressed channels, reparameterized sampling, diagonal Gaussian KL closed-form calculation and parameters The backpropagation update involves reparameterizing the encoder to output the task prediction in the prediction channel and calculating the constrained bit rate in the compression channel using a diagonal Gaussian KL closed-form algorithm. The gradients of both are propagated back to update the encoder and decoder parameters. The main branch loss is Its task item is evaluated on the disturbance characterization at the receiver to internalize channel robustness; the loss of each single-mode auxiliary branch is The losses of each branch are jointly optimized in each iteration, rather than being sequentially pre-trained and fine-tuned. A standard normal distribution is taken as the prior. The KL of the diagonal Gaussian encoder has the following closed form:

[0115] .

[0116] in, Traverse all dimensions of the bottleneck representation. and The encoder pairs the first The mean and variance of the output in each dimension.

[0117] In a dual robust configuration ( Further activating the Fisher source domain regularization term, the log-likelihood of the diagonal Gaussian encoder is expanded as follows:

[0118] .

[0119] in, The bottleneck code Each dimension component and The encoder pairs the first The mean and variance of the output in each dimension. As a bottleneck dimension, Given pi, summate. Traverse all One dimension. Input gradient. Automatic Differential Pair Differentiation shows that the gradient propagates simultaneously through the mean network. With variance network This simultaneously penalizes the input sensitivity of the encoding distribution in terms of both location and uncertainty. For each encoder branch... The process involves sequentially setting the input tensor to the desired gradient, calculating and sampling the distribution parameters, evaluating the log-likelihood, and obtaining the gradient through automatic differentiation. ,Pick The complete loss is expressed by the following formula:

[0120] .

[0121] The motivation for choosing the Fisher trace is that the spectral norm of the input Jacobian only characterizes the first-order sensitivity of the deterministic code while ignoring the uncertainty quantization of the encoder; the loss with respect to the Hessian of the input requires the second derivative of the decoder, which is costly and noisy; the Fisher trace is only a property of the encoder itself, characterizing the sensitivity of the distribution space rather than the code space, and allows for the above-mentioned efficient gradient norm reconstruction, thus possessing both statistical validity and computational feasibility for high-dimensional multimodal inputs.

[0122] The training and tuning strategy in this embodiment is as follows:

[0123] 1) Two-stage overparameter tuning: Stage 1 in Selected via grid search A stable compression-precision operating point is established under the target training PSNR, and then fixed. Phase Two in Fixed General On the logarithmic scale The algorithm performs an upscan and uses a linear warm-up from 0 to the target value for the first 10 to 20 rounds to prevent the Fisher term from dominating the optimization before the latent variable distribution stabilizes. This sequential tuning utilizes... and The two compromises have approximately orthogonality, without needing to... Joint grid search.

[0124] 2) Stability Practice: Apply a threshold to the global gradient Pruning is used to suppress gradient explosion caused by Fisher gradient computation; the encoder variance branch uses a softplus activation with negative bias to constrain the variance to be far from zero and ensure KL boundedness; in different Values ​​are synchronized with random seeds for isolation. The effect; monitoring the amplitude of component-wise loss to detect imbalances between cross-entropy, KL and Fisher terms.

[0125] 3) Implementation details: All experiments are based on PyTorch, using the Adam optimizer and initial learning rate. With cosine annealing scheduling, batch size 64, maximum 200 rounds, and early stopping based on overall accuracy verification; the only method-specific hyperparameter is... The remaining configurations are shared across all settings and baselines. In the channel robustness assessment, the training PSNR is set to 15 dB; in the joint disturbance assessment, the training PSNR is set to 10 dB, and the source domain disturbance coverage is set to default. ; exist Search on the internet, among which For the channel robust special case, additive white Gaussian noise is injected during both the training and testing phases to maintain training-test consistency.

[0126] 4) Complexity: Let the channel robustness setting be the cost of each iteration. Double robust setting The input gradient calculation is added to each single-modal branch and each fusion branch. ;exist With a single NVIDIA RTX 4090, a batch size of 64, and 200 rounds, the measured training overhead for the wall clock is approximately 15% to 25%. The memory overhead stems from reserving intermediate activations for gradient calculation, which can be mitigated by mixing precision and gradient accumulation when memory is limited. During the inference phase, Fisher regularization and random sampling are not involved; the pipeline degenerates into a deterministic forward pass encoded by mean. Its computational graph, number of parameters, memory usage, and latency significantly impact the overall performance. and Completely identical, bilateral robustness requires purely offline investment with zero deployment overhead.

[0127] To verify the effectiveness of the method provided in this invention, experiments were conducted on the three publicly available datasets mentioned above, and comparisons were made with representative multimodal fusion methods. The comparison methods included: the Spatial-Spectral Cross-Modal Enhancement Network (S2ENet, employing squeezed excitation fusion), the Encoder-Decoder Compact Fusion Network (EndNet, with assisted reconstruction task), the Spectral Spatial Multimodal Fusion Method based on Extended Morphological Attribute Profiles and Multiresolution Analysis (denoted as MRS-Hong), and the Channel Robustness Setting of this invention (…). ) and dual robust setting ( Both inventions share the same network backbone, ensuring that performance differences stem from training objectives rather than architectural capacity. Evaluation metrics include overall accuracy (OA), average accuracy (AA), and Cohen's... The coefficient of robustness in testing PSNR from 0 to 20 dB with source domain perturbation coverage. The next evaluation is performed, with each result averaged under three different random seeds and the mean value used as the averaging factor. Standard deviation report.

[0128] Experimental results show that, at a representative operating point with a peak signal-to-noise ratio of 15 dB, the channel robustness setting of this invention (…) It consistently outperforms structure-guided baselines across all three datasets, achieving overall accuracy improvements relative to the strongest baselines on each dataset. (Houston2013) (Augsburg) and (Berlin) percentage points, with an average improvement of approximately [percentage missing] across datasets. Percentage points; among them, the Berlin dataset has the most bands (244) and the highest input dimension, and bottleneck compression brings the greatest gain by suppressing redundant spectral information. In the test scan of PSNR from 0 to 20dB, the present invention exhibits a smoother degradation curve and stronger stability in the low PSNR range. The relative gain is most significant in the low to medium PSNR range of 0 to 10dB and converges in the high PSNR range (greater than 15dB), confirming that receiver alignment supervision is the main source of channel robustness. Under a fixed transmission budget, the present invention achieves higher unit bit task accuracy, and improves relative MRS-Hong by a budget of 16MB. Improvements compared to S2ENet percentage points.

[0129] Robust weights for source domain The ablation experiments showed an inverted U-shaped trend: with From 0 to Performance improvement, The price fell back, consistent with the trade-off between deviation and variance. Too small a source domain sensitivity to underregulation Excessive smoothing of class boundaries and suppression of discriminative features; the optimal operating point is This embodiment uses this as the default dual robust configuration. On Houston 2013, the overall accuracy is... time Upgraded to time This trend is evident in OA, AA, and Consistent with the above.

[0130] Source domain disturbance coverage As the value gradually increases, the dual robust setting of this invention ( Consistently outperforms channel robustness settings across all coverage levels. ), and the performance gap increases with Increased and widened: Overall accuracy gap in Houston 2013 time Expand to time Similar trends were observed on Augsburg and Berlin. Since the perturbation strength was fixed and only the coverage varied, the aforementioned gains can be attributed to coverage-driven regularization rather than noise scale inconsistency, confirming the coverage-adaptive nature of Fisher-based source-domain robust targets. Class analysis further showed that classes with higher intra-class variance (e.g., commercial areas, land parcels) benefited more significantly from dual-robust training, while spectrally homogeneous classes (e.g., water bodies, tennis courts) gained less, consistent with the expectation that source-domain regularization is most valuable where the decision boundary is inherently ambiguous.

[0131] In summary, the experiments support the central hypothesis of this invention: robustness in satellite-to-ground multimodal mission-oriented communication should be optimized as a coupled source-channel problem rather than two independent modules. Under the same architecture and training protocol, the channel robustness setting of this invention improves both fixed-point accuracy and channel robustness. The dual robustness setting provides further gains under joint source-channel perturbations without increasing any inference overhead. The gains are most significant in challenging regions where single-sided robustness strategies are insufficient, such as low PSNR and high source-channel perturbation coverage.

[0132] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. This embodiment takes co-registered hyperspectral images and lidar data as an example. The framework can be extended to other mode combinations, including synthetic aperture radar, provided that the alignment and labeling are consistent. Scenarios with missing modes and severe asynchronous acquisition are not within the scope of this embodiment. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A robust, multimodal remote sensing classification method for satellite-to-ground missions addressing information bottlenecks, characterized in that... The method includes: Step S1: Source domain perturbation injection and mode-specific two-layer probabilistic bottleneck coding S1-1: For applications including hyperspectral imaging and lidar One co-registered modal input, based on coverage Additive Gaussian noise is injected into a randomly selected subset of the input within each mini-batch to simulate source domain perturbation, while unselected positions remain unchanged; S1-2: Each mode after the perturbation is injected is input into the corresponding mode-specific probabilistic encoder. Each encoder outputs the mean and logarithmic variance of a diagonal Gaussian distribution, and the single-mode bottleneck characterization of the mode is obtained through reparameterized sampling. Step S2: Cross-modal fusion and variational semantic compression S2-1: Aggregate the bottleneck representations of each single mode into a joint representation using the cross-modal fusion operator; S2-2: The joint representation is mapped to diagonal Gaussian distribution parameters by the fusion encoder, and the fusion semantic code to be transmitted via satellite-to-ground downlink is obtained through reparameterized sampling; Step S3: Differentiable additive white Gaussian noise channel simulation and receiver task decoding S3-1: A differentiable additive white Gaussian noise channel simulation layer that controls the noise variance through peak signal-to-noise ratio injects channel noise into the semantic code to obtain the perturbed semantic code at the receiver. This code is then input into the receiver's main decoder to output ground object classification prediction. The channel simulation layer allows the gradient to be backpropagated from the receiver loss to each encoder at the satellite end. S3-2: During the training phase, an auxiliary decoder characterized by bottlenecks in each single mode outputs predictions to provide intramodal sufficiency supervision; Step S4: Source Domain Sensitivity Regularization and Double Robust Variational Optimization Based on Fisher Information Matrix S4-1: Maximizing label-related information, suppressing representation redundancy, and suppressing source domain sensitivity are modeled as a single information-theoretic optimization objective, where the task-related terms are evaluated on the receiver's perturbed representation, and channel robustness is incorporated into the optimization objective; S4-2: Replace the unsolvable mutual information term with a cross-entropy surrogate term and a KL divergence closed term; S4-3: Instantiate the source domain sensitivity regularization term as the expectation of the encoder's log-likelihood with respect to the square of the gradient norm of its input, and apply it to all single-modal encoders and fusion encoders. S4-4: Adjust the source domain sensitivity regularization weights for channel robustness settings. Training is performed using task supervision applied to the receiver's perturbation representation, with source domain sensitivity regularization weights set in a dual-robust manner. Activate the source domain sensitivity regularization term; S4-5: Output the trained encoding → transmission → decoding network, which is used to classify land cover from the multimodal remote sensing data to be classified; The dual robust multimodal information bottleneck adopts a two-level bottleneck structure with single-modality first and fusion later. The single-modality level bottleneck first suppresses intramodal interference variation, and the fusion level bottleneck removes cross-modal redundancy that remains after complementary semantic aggregation.

2. The dual-robust multimodal information bottleneck remote sensing classification method for space-to-ground missions according to claim 1, characterized in that, The source domain perturbation injection in step S1 models the source domain degradation as additive Gaussian noise applied to a randomly selected subset of the input, for the first... The disturbed input is generated at each input position using the following formula. : ； in, This is the original modal input; By coverage within each mini-batch A uniformly extracted set of disturbed locations; To indicate An indication function for receiving noise at the location within; It is additive Gaussian noise; It follows a Gaussian distribution; It is the identity matrix; The following formula represents the sample-by-sample noise variance normalized to the input power: ； in, Input power per sample; For input dimensions; For each sample, the per-sample perturbation intensity factor is used. for Norm; The source domain degradation modeling will affect the perturbation coverage. With per-sample perturbation intensity factor Decoupling.

3. The dual-robust multimodal information bottleneck remote sensing classification method for space-to-ground missions according to claim 1, characterized in that, The modality-specific two-layer probabilistic bottleneck coding in step S1 specifically includes: 1) Each modality is encoded by an independent probabilistic backbone network, which consists of a convolutional feature extractor and two linear heads, each outputting a mean. With log variance The variance head is activated using softplus with a small negative bias; 2) The single-mode bottleneck characterization is obtained through reparameterized sampling, expressed by the following formula. : ； in, For the first Input of each modality; and The first The mean function and standard deviation function of the encoder output; For element-wise multiplication; To obtain from the standard normal distribution A random noise vector sampled in the middle; It is an identity matrix.

4. The dual-robust multimodal information bottleneck remote sensing classification method for space-to-ground missions according to claim 1, characterized in that, Step S2 specifically includes: 1) Employing a fusion operator The bottleneck characteristics of each single mode are aggregated into a joint characteristic. , The fusion operator The aggregation is achieved by linear projection after splicing the channels; 2) Joint characterization The fusion encoder maps the parameters to a diagonal Gaussian distribution, and the fusion semantic code to be transmitted is obtained through reparameterized sampling, as shown in the following formula. : ； in, and These are the mean and standard deviation of the output of the fusion encoder, respectively. For element-wise multiplication; This is a random noise vector sampled from a standard normal distribution.

5. The dual-robust multimodal information bottleneck remote sensing classification method for space-to-ground missions according to claim 1, characterized in that, Step S3 specifically includes: 1) Model the satellite-to-ground downlink as an additive white Gaussian noise channel; 2) Inject channel noise into the fused semantic code to obtain the receiver-side disturbed semantic code as shown in the following formula. : ； in, The channel is additive white Gaussian noise; The channel noise variance; It is the identity matrix; 3) Channel quality is defined by the peak signal-to-noise ratio (PSNR) shown in the following formula: ； in, The peak power of semantic code features calculated in mini-batches. This indicates taking the maximum value of each dimension of the semantic code. For semantic code number Each dimension component; Let V be the channel noise variance. For a given target's PSNR, the channel noise variance is: ; 4) The receiver's main decoder takes the scrambled semantic code as input and outputs the class log probability. ,in For predicting labels; 5) During the training phase, each single-mode auxiliary decoder... The output predicts the output, the inference stage evaluates the master decoder, and each encoder outputs a deterministic output with the mean. .

6. The dual-robust multimodal information bottleneck remote sensing classification method for space-to-ground missions according to claim 1, characterized in that, Step S4 specifically includes: 1) Unified Information Theory Goal a. Solve for the encoder parameters based on the optimization objective expressed in the following formula. With decoder parameters : ； in, For random variables representing feature category labels; For the random variable representing the disturbance at the receiving end; and These are random variables representing the fused semantic code and the fused representation, respectively. and The first Bottleneck representation of each modality and input random variables; and These are the trainable parameters for the encoder and decoder, respectively; To compress weights; Source domain sensitivity Regular weights; For mutual information; This is a source domain sensitivity regularization term; b. Disturbance representation of the first group of task-related items at the receiving end The evaluation incorporates channel robustness into the optimization objective; the compression term is composed of compression weights. Controlled compression → trade-offs in discriminativity; source domain sensitivity regularization term Source domain sensitivity regularization weights Control, when Degenerates into channel robust setting. Activate dual robust training; 2) Cross-entropy proxy and KL closed-form compression terms a. Introduce a variational decoder to set a lower bound for task-related terms, and use cross-entropy proxy as the upper bound for negative mutual information. b. Introduce fixed priors The standard normal distribution, with an upper bound on the compression term, gives the KL divergence between the diagonal Gaussian encoder and the standard normal prior as a closed-form equation: ； in, For parameters The conditional distribution of the output of the probability encoder; The prior is the standard normal distribution. To traverse the various dimensional components of the bottleneck representation, , As a bottleneck dimension; and The encoder pairs the first The mean and variance of the output in each dimension; The closed-loop element-wise operation on the encoder output employs reparameterization throughout. ( To achieve differentiable sampling; 3) The reconstructability of Fisher's information source domain regularization terms a. The expectation of the trace Instantiated using the input space Fisher information matrix, its Fisher information matrix is ​​expressed as follows: ； in, The Fisher information matrix is ​​the input space. For input The gradient operator; For the log-likelihood of the encoder; Transpose a vector or matrix; b. Define source domain sensitivity regularization terms Expectations for its traces ,in For the trace operation of a matrix, use the identity. And the linearity of the trace and expectation, the expectation of the trace The simplified expression is the expectation of the squared gradient norm of the input, expressed as follows: ； in, To evaluate the diagonal Gaussian log-likelihood obtained by automatic differentiation; c. Instantiate the Fisher information source domain regularization term in the fusion encoder and each single-modal encoder respectively, and express it as follows: ； ； in, For the first Fisher regularization terms for a single-modal encoder; For the first Modal input The gradient; For the first Conditional distribution of a single-mode encoder; The total number of modes; 4) Trainable loss Based on the combined cross-entropy of the KL agent and the Fisher regularization term, the main branch loss is described by the following formula. , No. Single-modal auxiliary branch loss and dual robust multimodal information bottleneck loss Trainable loss: ； ； ； in, For the receiver's cross-entropy proxy item; For the KL compression term of the fusion encoder; Main decoder; Semantic codes for channel disturbances; To compress weights; For the first One auxiliary decoder; and They are respectively the fusion encoder and the first Fisher source domain regularization terms for a single-mode encoder; To traverse all One mode; The source domain sensitivity regularization weights are set to the following channel robustness settings: The dual robustness setting is as follows: .

7. The dual-robust multimodal information bottleneck remote sensing classification method for space-to-ground missions according to claim 1, characterized in that, The source domain sensitivity regularization weight in step S4 A two-stage tuning strategy is adopted: Stage 1 is in Next, the compression weights are selected through grid search. And establish a compression-precision working point under the target training PSNR, and then fix it. Phase Two in Fixed Under the conditions, On the logarithmic scale The system performs an up-scan and uses linear preheating from 0 to the target value in the first few rounds; the two-stage tuning strategy utilizes... and Two compromises in approximate orthogonality achieve stable convergence.

8. The dual-robust multimodal information bottleneck remote sensing classification method for space-to-ground missions according to claim 1 or claim 2, characterized in that, The source domain perturbation injection is performed under a dual robust setting with Fisher regularization. Calculate using the following steps: 1) For each encoder branch The input tensor is set to the required gradient, the distribution parameters are calculated and the bottleneck code is obtained by reparameterized sampling. 2) Use automatic differentiation to obtain the diagonal Gaussian log-likelihood evaluation. input gradient and take ; 3) During training, the global gradient is thresholded. Pruning is performed to suppress gradient explosion; 4) During the inference phase, each encoder executes a deterministic forward pass with the mean as the output, and... and The two trained models have identical computational graphs, parameter counts, memory usage, and latency during inference.

Images