Dynamic Frequency Feature Fusion Network (DFFNet)

• The paper introduces DFFNet, which dynamically adapts frequency filtering and modality fusion to overcome limitations in remote sensing classification.
• It integrates a Dynamic Filter Block (DFB) with a Spectral-Spatial Adaptive Fusion Block (SSAFB) to jointly process hyperspectral and SAR/LiDAR data.
• Empirical results on Berlin and Houston2013 confirm significant improvements in OA, AA, and Kappa compared to baseline methods.

Searching arXiv for the specified DFFNet paper and closely related adaptive feature-fusion work.

arXiv results gathered. I’ll write the article centered on the direct DFFNet paper and situate it against closely related but differently named architectures.

Dynamic Frequency Feature Fusion Network (DFFNet) is a multi-source remote sensing classification architecture introduced for the joint classification of hyperspectral images (HSI) with Synthetic Aperture Radar (SAR) or Light Detection and Ranging (LiDAR) data. It is defined by the combination of a Dynamic Filter Block (DFB), which performs input-adaptive filtering in the frequency domain, and a Spectral-Spatial Adaptive Fusion Block (SSAFB), which performs cross-modal fusion between HSI and SAR/LiDAR features. In the reported formulation, DFFNet is designed to address two stated limitations of prior methods: limited adaptability to diverse land-cover frequency characteristics and suboptimal cross-modal fusion (Zhao et al., 6 Jul 2025). Closely related literature includes dynamic feature fusion in semantic edge detection, where fusion weights are conditioned on input images and locations (Hu et al., 2019), and several spatial-frequency fusion architectures that are relevant by analogy but are not the same model, such as SFAFNet (Gao et al., 20 Feb 2025), SFFNet (Yang et al., 2024), and frequency-aware fusion frameworks for pansharpening and image fusion (Xing et al., 2022, Hu et al., 2024).

1. Definition and scope

DFFNet stands for Dynamic Frequency Feature Fusion Network. The model is proposed for multi-source remote sensing land-cover classification, specifically the joint classification of hyperspectral images with SAR or LiDAR data (Zhao et al., 6 Jul 2025). In the reported problem setting, HSI provides rich spectral signatures for material discrimination, while SAR and LiDAR provide complementary structural or elevation information. The network is therefore organized around two linked objectives: adaptive modeling of frequency-domain characteristics and adaptive fusion of heterogeneous modalities.

The motivation is explicitly tied to the observation that existing methods “struggle to adapt to the frequency-domain characteristics of different land cover types,” and that prior fusion strategies often fail to fully exploit the spectral richness of HSI together with the spatial or structural details of SAR/LiDAR (Zhao et al., 6 Jul 2025). This positions DFFNet within a broader class of adaptive fusion methods, but with a specifically frequency-domain formulation rather than purely spatial feature aggregation.

A recurring source of confusion is nomenclature. “DFF” has also been used for Dynamic Feature Fusion in semantic edge detection, where the core mechanism predicts image- and location-specific fusion weights over multi-level CNN responses (Hu et al., 2019). That model is not frequency-domain. By contrast, DFFNet explicitly applies FFT, input-conditioned frequency filtering, and inverse FFT (Zhao et al., 6 Jul 2025). Other similarly named systems, including DFFN for low-light remote sensing enhancement (Yao et al., 2024) and DWFF-Net for farmland habitat segmentation (Zheng et al., 11 Nov 2025), are architecturally related through adaptive fusion principles but are not the same network.

2. Architectural organization

DFFNet is divided into a primary network and a refinement module (Zhao et al., 6 Jul 2025). The primary network contains two Dynamic Filter Fusion Modules (DFFMs), while the refinement module contains two fully connected layers for final land-cover classification. Before feature extraction, the HSI input is reduced by PCA, producing $\mathbf{I}^H$, while the SAR/LiDAR input is denoted $\mathbf{I}^X$ (Zhao et al., 6 Jul 2025).

Initial modality-specific processing is asymmetric. The model applies 3D convolution to $\mathbf{I}^H$ and 2D convolution to $\mathbf{I}^X$, reflecting the different structural properties of the modalities (Zhao et al., 6 Jul 2025). The resulting features are then passed into the repeated DFFM blocks. Within each DFFM, the feature stream is split into two complementary processing paths: one routed through the Dynamic Filter Block (DFB) for frequency-domain enhancement, and one routed through the Spectral-Spatial Adaptive Fusion Block (SSAFB) for cross-modal feature fusion (Zhao et al., 6 Jul 2025).

The network’s organization can be summarized as a sequence of modality-specific preprocessing, two repeated fusion modules, and a refinement head. The paper states that the best setting uses 2 DFFMs and 30 principal components after PCA (Zhao et al., 6 Jul 2025).

Component	Function	Reported role
PCA + 3D/2D convolutions	Input preprocessing	Reduce HSI redundancy; initialize HSI and SAR/LiDAR features
DFB	Frequency-domain adaptive filtering	Model class-dependent frequency characteristics
SSAFB	Spectral-spatial cross-modal fusion	Combine HSI and SAR/LiDAR features
Two FC layers	Refinement/classification	Produce final land-cover prediction

This design suggests a division of labor: DFB enhances modality features through adaptive frequency filtering, while SSAFB performs modality-aware fusion after the features have been enriched by frequency-domain processing (Zhao et al., 6 Jul 2025).

3. Dynamic Filter Block

The DFB is the most distinctive component of DFFNet. Given an input feature map $\mathbf{X}_{in}$, it first computes a frequency-domain representation using FFT:

$$f(u,v)=\sum^{H-1}_{y=0}\sum^{W-1}_{x=0}\mathbf{X}_{in}(x,y)e^{-j2\pi(\frac{ux}{W}+\frac{vy}{H})}.$$

The transformed feature is then filtered by an input-conditioned kernel:

$$\hat{f}(u,v)=\mathcal{K}(\mathbf{X}_{in}) \odot f(u,v),$$

where $\odot$ denotes element-wise multiplication (Zhao et al., 6 Jul 2025). The filtered response is mapped back to the spatial domain by inverse FFT:

$$\mathbf{X}_{out}=\frac{1}{HW}\sum^{H-1}_{v=0}\sum^{W-1}_{u=0}\hat{f}(u,v)e^{j2\pi(\frac{ux}{W}+\frac{vy}{H})}.$$

The dynamic kernel itself is not fixed. The paper defines a set of learnable basis filters

$\mathbf{F}=\{F_1, F_2, \ldots, F_N\},$

and constructs the effective filter by global average pooling, an MLP, Softmax normalization, and tensor multiplication:

$\mathbf{I}^X$0

This means the filter is an input-conditioned weighted combination of basis filters rather than a static frequency mask (Zhao et al., 6 Jul 2025). The paper describes this as injecting “frequency contextual knowledge” into the filter kernels.

After inverse FFT, the feature is passed through a feed-forward network that contains an identity branch, a spatial convolutional branch, and a frequency convolutional branch, with both convolutional branches using depth-wise convolution (Zhao et al., 6 Jul 2025). The explicit equations for this feed-forward network are not provided in the supplied text, but its stated purpose is to exploit complementary information between spatial and frequency representations.

This adaptive synthesis of filters is the aspect that most directly justifies the term dynamic in DFFNet. In contrast to fixed fusion strategies in earlier feature-fusion work (Hu et al., 2019), DFB makes the frequency response sample-dependent (Zhao et al., 6 Jul 2025).

4. Spectral-Spatial Adaptive Fusion Block

The SSAFB is the cross-modal fusion block. It operates on an HSI feature $\mathbf{I}^X$1 and a SAR/LiDAR feature $\mathbf{I}^X$2, and applies different attention mechanisms to each modality. For HSI, the paper uses a channel- or spectral-attention form:

$\mathbf{I}^X$3

where $\mathbf{I}^X$4 is ReLU and $\mathbf{I}^X$5 denotes convolution (Zhao et al., 6 Jul 2025). For SAR/LiDAR, the paper uses a spatial-attention form:

$\mathbf{I}^X$6

The two attended features are then concatenated, channel-shuffled, and projected by a $\mathbf{I}^X$7 convolution:

$\mathbf{I}^X$8

The channel shuffle operation is used to “effectively exchange information between modalities” and “enhance interactions between spectral and spatial attention weights” (Zhao et al., 6 Jul 2025). This formulation reflects a strong modality prior: HSI is treated as spectrally rich, while SAR/LiDAR is treated as spatially or structurally informative.

A plausible implication is that SSAFB functions as a typed fusion operator rather than a generic concatenation block. The attention mechanism differs by modality before the actual fusion step, which distinguishes DFFNet from architectures that use a shared attention rule for all inputs. The paper further notes that $\mathbf{I}^X$9 is fused again with HSI and SAR/LiDAR features via element-wise summation, although explicit equations for that later summation are not given (Zhao et al., 6 Jul 2025).

This cross-modal design places DFFNet near other adaptive fusion frameworks, but its combination of frequency filtering and modality-specific attention is more specialized than the fixed grouped-convolution fusion used in earlier DFF-style edge detectors (Hu et al., 2019).

5. Empirical performance and ablation evidence

DFFNet is evaluated on two benchmark datasets: Berlin for HSI+SAR and Houston2013 for HSI+LiDAR (Zhao et al., 6 Jul 2025). On Berlin, the reported results are:

• OA: 75.42
• AA: 64.85
• Kappa: 63.22

These exceed the listed baselines TBCNN, S$\mathbf{I}^H$0ENet, FusAtNet, DFINet, and MICF-Net (Zhao et al., 6 Jul 2025). On Houston2013, the reported performance is:

• OA: 92.35
• AA: 93.48
• Kappa: 91.70

The paper states that DFFNet surpasses the others by at least 2.79% OA on Houston2013 (Zhao et al., 6 Jul 2025).

Classwise, the Berlin results highlight Industrial area = 54.69 and Low plants = 81.88 for DFFNet (Zhao et al., 6 Jul 2025). On Houston2013, selected class accuracies include Stressed grass = 99.72, Commercial = 93.46, Highway = 84.27, Parking lot2 = 92.98, Tennis court = 100.0, and Running track = 100.0 (Zhao et al., 6 Jul 2025).

The ablation study isolates the contributions of DFB and SSAFB. On Houston2013:

• w/o SSAFB and DFB: 87.51 OA
• w/o SSAFB: 90.81 OA
• w/o DFB: 89.95 OA
• Proposed DFFNet: 92.35 OA

On Berlin:

• w/o SSAFB and DFB: 72.29 OA
• w/o SSAFB: 74.53 OA
• w/o DFB: 73.32 OA
• Proposed DFFNet: 75.42 OA (Zhao et al., 6 Jul 2025)

These numbers support the claim that both modules are necessary, and they suggest that DFB is especially important on Berlin. The architecture search over DFFM count shows that 2 DFFMs is optimal, while the PCA study shows that 30 principal components yields the best performance (Zhao et al., 6 Jul 2025).

The reported computational profile is also lightweight relative to some baselines:

Method	Params (M)	FLOPs (G)	Inference time (s)
TBCNN	0.3875	0.0063	0.2236
S$\mathbf{I}^H$1ENet	1.4701	0.1778	0.2489
FusAtNet	37.7177	3.5619	0.2469
DFINet	1.3155	0.1191	0.3182
MICF-Net	2.1375	0.0536	0.2965
DFFNet	1.2829	0.0303	0.2387

This indicates that DFFNet achieves a moderate parameter count and low FLOPs while improving accuracy (Zhao et al., 6 Jul 2025).

6. Position within the literature

DFFNet belongs to a broader family of adaptive feature-fusion models, but it differs from them in both domain and fusion target. The earlier Dynamic Feature Fusion model for semantic edge detection learns input- and location-adaptive weights over multi-level CNN responses and shows that location-adaptive fusion outperforms fixed and location-invariant fusion (Hu et al., 2019). However, that work operates entirely on spatial feature maps and does not perform frequency-domain analysis.

A second neighboring line is represented by SFAFNet, which proposes a spatial-frequency domain adaptive fusion network for image deblurring. Its GSFFBlock includes a frequency-domain information dynamic generation module and gated cross-domain fusion, but the decomposition is formulated through learnable low-pass and high-pass filtering in feature space rather than through the FFT-based basis-filter synthesis used in DFFNet (Gao et al., 20 Feb 2025). Another relevant analogue is SFFNet for remote sensing segmentation, where wavelet-based low- and high-frequency features are aligned with spatial features through multiscale dual-representation attention (Yang et al., 2024). These models reinforce the broader principle that spatial and frequency representations are complementary, but they differ from DFFNet’s multi-source classification setting and its explicit dynamic filter bank in the frequency domain.

Frequency-aware fusion has also appeared in remote sensing pansharpening, where FAFNet uses DWT/IDWT and a high-frequency feature similarity loss to fuse PAN and MS features (Xing et al., 2022), and in infrared-visible fusion, where SFDFusion constructs a parallel spatial-frequency architecture with FFT-based amplitude and phase fusion (Hu et al., 2024). Those works are relevant by analogy, but DFFNet differs in that its primary task is HSI+SAR/LiDAR land-cover classification and its dynamic component is a basis-filter synthesis mechanism rather than amplitude-phase fusion or wavelet-domain reconstruction (Zhao et al., 6 Jul 2025).

This suggests that DFFNet is best interpreted as a specialized member of the adaptive frequency-feature fusion family: it combines the adaptive weighting idea seen in dynamic fusion literature (Hu et al., 2019) with explicit frequency-domain filtering and modality-aware fusion (Zhao et al., 6 Jul 2025).

7. Interpretation, significance, and limitations

The significance of DFFNet lies in the way it binds two design claims into a single classifier. First, it treats frequency-domain modeling as something that should be adaptive to land-cover content rather than fixed across samples. Second, it treats cross-modal fusion as something that should be modality-aware, using channel-oriented enhancement for HSI and spatial-oriented enhancement for SAR/LiDAR (Zhao et al., 6 Jul 2025). The ablation results support both claims.

A plausible implication is that DFFNet’s frequency branch is not merely a denoising device. Because the dynamic filter is synthesized from pooled input features and applied before inverse FFT, it serves as a sample-specific spectral selector whose behavior can vary across land-cover types (Zhao et al., 6 Jul 2025). This is distinct from architectures that simply append a fixed spectral transform to a conventional backbone.

At the same time, some implementation details remain unspecified in the provided text. The paper does not give the exact loss function, patch size, scheduler, weight decay, or detailed layer widths (Zhao et al., 6 Jul 2025). It also does not provide explicit formulas for the post-IFFT feed-forward network inside DFB or for the later element-wise summation after SSAFB. These omissions do not obscure the core architectural idea, but they do limit strict reproducibility from the textual description alone.

Another important caveat is conceptual. Not every model with “DFF” or “feature fusion” in its name is frequency-domain. DWFF-Net for farmland habitat segmentation performs dynamic weighting over transformer layers rather than frequency bands (Zheng et al., 11 Nov 2025), and DFF for semantic edge detection predicts dynamic fusion weights over side outputs in the spatial CNN domain (Hu et al., 2019). DFFNet, in the strict sense, refers specifically to the HSI+SAR/LiDAR classification architecture built around the Dynamic Filter Block and Spectral-Spatial Adaptive Fusion Block (Zhao et al., 6 Jul 2025).

In that restricted meaning, Dynamic Frequency Feature Fusion Network denotes a model in which input-conditioned frequency filters and modality-specific attention are jointly used to improve multi-source remote sensing classification. Its direct contribution is therefore not only higher reported OA, AA, and Kappa on Berlin and Houston2013, but also a concrete architectural formulation of adaptive frequency-domain modeling for heterogeneous remote sensing data (Zhao et al., 6 Jul 2025).