Dual-Domain Illumination Prior for Low-Light Remote Sensing Image Enhancement

Highlights

What are the main findings?

• We identify a dual-domain prior design strategy that jointly calibrates per-color-channel Fourier amplitude and spatial-domain per-channel statistics and fuses their outputs through a selective gated mechanism (SCFF). The resulting illumination prior is integrated with the host backbone to support subsequent detail restoration.
• DDIP is jointly optimized with the tested enhancement backbones as an auxiliary illumination-prior module for low-light image restoration. Across the evaluated datasets, it yields PSNR and SSIM gains over the corresponding backbone baselines.

What are the implications of the main findings?

• Lightweight integration and prior design. DDIP is jointly optimized with the tested CNN-, Transformer-, and Mamba-based enhancement networks as an auxiliary illumination-prior branch, while their main restoration blocks remain unchanged. Within the evaluated settings, the results support combining frequency-domain and spatial-domain illumination statistics in a unified prior branch alongside end-to-end low-light enhancement networks.
• Evidence within the evaluated scope. The reported comparisons cover iSAID-dark, a limited downstream object-detection evaluation on its high-resolution scenes, and five general low-light benchmarks. They do not establish performance on untested sensors, modalities, or real nighttime remote sensing data.

Abstract

Low-light conditions degrade remote sensing imagery by reducing contrast, distorting color, and obscuring fine terrain structures and small objects critical for Earth observation. Accurate illumination adjustment under spatially varying scene content remains challenging for existing enhancement methods, and many prior-guided approaches operate exclusively in either the spatial domain or the frequency domain. In this work, we propose a Dual-Domain Illumination Prior (DDIP), a trainable dual-domain illumination-prior module that is jointly optimized with each host backbone and exploits frequency-domain and spatial-domain illumination statistics. DDIP comprises three components: a Frequency-Domain Illumination Distribution Prior (FIDP) that performs per-color-channel amplitude calibration in Fourier space to improve global brightness; a Spatial-Domain Illumination Distribution Prior (SIDP), adapted from IDP-Net, that performs multi-scale sub-region statistical correction for local illumination adjustment; and a Selective Core Feature Fusion (SCFF) module that adaptively combines the frequency-domain output, the spatial-domain output, and the original input through an attention-based gating mechanism with dual pooling. DDIP is integrated with each host backbone while leaving its main restoration blocks unchanged. In the controlled reconstruction comparisons on iSAID-dark and the evaluated general low-light benchmarks, equipping the tested backbone networks with DDIP improves PSNR and SSIM over their corresponding baselines. Complementary LPIPS and CIELAB lightness measurements characterize perceptual similarity and lightness behavior, while a fixed-detector object-detection evaluation on the tested high-resolution iSAID-dark scenes examines the effect of the enhancement pipelines under the reported synthetic low-light conditions. The ablation studies further examine the contribution of the module components within the reported experimental settings.

1. Introduction

Images captured under low-light conditions suffer from low contrast, color distortion, and amplified sensor noise. In remote sensing, these degradations can impair visual quality and downstream tasks such as aerial object detection [1], land-cover classification [2], and disaster monitoring [3]. Remote sensing imagery poses challenges beyond those of generic low-light photography: high spatial resolution with dense small objects; heterogeneous land cover types (vegetation, water, infrastructure) with distinct reflectance characteristics; and long shadows cast by terrain or buildings that produce spatially complex illumination distributions. Dedicated low-light remote sensing enhancement methods have been developed to address these challenges [4]. Low-light remote sensing image enhancement aims to improve global illumination and recover fine-grained scene details obscured in dark regions.

Low-light enhancement has been approached through three broad methodological directions. Physics-based methods, such as Retinex theory [5,6,7] and variational formulations [8,9], offer interpretable priors but rely on assumptions that can be difficult to satisfy under the heterogeneous land covers and strong directional shadows typical of remote sensing scenes. Deep learning methods, spanning CNNs [10,11], Transformers [12,13], and Mamba-based architectures [14], have expanded representational capacity and are typically designed as complete end-to-end networks. Among prior-guided approaches, histogram equalization (HE) [15] and gamma correction (GC) [16] use global pixel-level statistics for enhancement; SNR-Net [17] introduces spatially varying enhancement through SNR-guided feature fusion and attention; CSNorm [18] selectively normalizes lightness-relevant feature channels; and IDP-Net [19] establishes the relationship between per-channel statistics and illumination as an explicit prior. These methods illustrate the effectiveness of spatial-domain priors for enhancement guidance, but they operate on spatial-domain statistics and do not integrate frequency-domain illumination cues. Meanwhile, frequency-domain illumination statistics—in particular, the correlation between Fourier amplitude and global brightness [20]—have been exploited in dedicated deep architectures [20,21,22,23], where frequency-domain operations are embedded as integral network stages rather than as separate illumination-prior branches integrated with a host backbone while leaving its main restoration blocks unchanged.

Existing dual-domain low-light enhancement methods, including remote-sensing-oriented approaches [4], combine spatial and frequency information within complete architectures. Accordingly, this work formulates spatial-frequency dual-domain illumination calibration as a jointly optimized illumination-prior module for low-light remote sensing image enhancement.

Building on the above observations, we propose the Dual-Domain Illumination Prior (DDIP), a dual-domain illumination-prior module jointly optimized with each host backbone, that integrates frequency-domain amplitude calibration, spatial-domain statistical correction, and selective dual-domain fusion into a unified illumination prior. DDIP is integrated with each tested enhancement backbone while leaving its main restoration blocks unchanged.

The main contributions of this work are as follows:

• Dual-domain illumination prior with frequency-domain calibration. We formulate a dual-domain illumination-prior module (DDIP), jointly optimized with each host backbone, that integrates per-color-channel Fourier amplitude calibration with spatial-domain statistical correction into a unified prior branch. The frequency-domain component (FIDP) independently calibrates the Fourier amplitude of each RGB channel to correct global brightness while preserving phase-related structural information. The spatial calibration branch is adapted from IDP-Net [19] with a lightweight multi-scale refinement strategy.
• Selective core feature fusion (SCFF). We design an attention-based fusion mechanism with dual pooling (global average and max) that allocates per-channel branch-wise weights to adaptively combine the frequency-domain output, the spatial-domain output, and the original input. This lightweight gating module integrates the three sources while maintaining a direct contribution from the original input before the host backbone performs texture restoration.
• Evaluation across backbones and datasets. We evaluate DDIP with five CNN-, Transformer-, and Mamba-based backbones on the iSAID-dark remote sensing dataset [4] and on five general low-light benchmarks. Complementary perceptual and illumination-oriented metrics and a fixed-detector evaluation on high-resolution iSAID-dark scenes further assess the tested enhancement outputs. These experiments assess performance within the reported backbone–dataset settings.

2. Related Work

The literature most relevant to DDIP spans spatial-domain illumination priors, frequency-domain enhancement, and dual-domain designs developed for remote sensing and general low-light imaging.

2.1. Spatial-Domain Illumination Priors

Illumination-prior methods embed statistical knowledge about illumination distributions to constrain enhancement. They range from coarse global corrections—histogram equalization (HE) [15] and gamma correction (GC) [16], which do not explicitly adapt to spatially varying illumination—to spatially adaptive learned modules. SNR-Net [17] uses an SNR prior to guide spatially varying fusion between convolutional and Transformer features; CSNorm [18] selectively normalizes lightness-relevant feature channels to improve adaptation; IDP-Net [19] formalizes the per-channel mean–standard-deviation–illumination relationship; and LightQANet [24] learns illumination-related factors and incorporates them through light-aware prompts. Together, these methods show how spatial-domain statistics can guide illumination adjustment.

2.2. Frequency-Domain Enhancement

In Fourier-based low-light enhancement research, the core motivation is well established: the amplitude component of the Fourier transform correlates with global illumination, while the phase component encodes spatial structures and, when preserved, retains the original textural information of the image [20]. FSR-Net [25] applies a deep Fourier network to adjust amplitude for shadow removal, and UHDFour [26] extends this idea to ultra-high-definition low-light enhancement by processing amplitude and phase under a low-resolution regime before restoring high-resolution details. FourLLIE [20] introduces a two-stage pipeline in which an AmplitudeNet estimates a Fourier-domain amplitude transformation map, and a subsequent spatial-frequency fusion network (SFNet) interleaves Fourier and spatial blocks across multiple depths. Each of these methods demonstrates the effectiveness of Fourier-domain processing for illumination recovery.

2.3. Dual-Domain Enhancement: From Remote Sensing to General Low-Light Applications

In remote sensing, dual-domain designs address nighttime aerial imagery challenges. DFFN [4] proposes a spatial-frequency dual-domain feature fusion network for low-light remote sensing enhancement, and DEPFusion [27] introduces dual-domain enhancement with Fourier detail recovery for UAV multispectral object detection. FSDENet [28] applies frequency- and spatial-domain processing to remote sensing semantic segmentation, providing an example in a separate interpretation task. DFFN couples spatial–frequency feature extraction and fusion within a purpose-built end-to-end enhancement architecture. DEPFusion combines dual-domain enhancement with a multispectral UAV detection objective. These settings differ from the RGB enhancement setting studied here. DDIP is not presented as a replacement for either architecture, and the current RGB experiments do not establish performance on multispectral inputs.

Beyond remote sensing, dual-domain methods for general low-light enhancement follow similar principles. FECNet [21] employs spatial–frequency interaction where amplitude and phase branches exchange information throughout encoding and decoding. DMFourLLIE [22] adopts a dual-stage multi-branch Fourier architecture, and LTCF-Net [29] integrates Fourier transforms within a Transformer-based luminance-chrominance framework. SPJFNet [23] combines wavelet and Fourier processing in a dual-frequency design with self-mining guidance, while Joint-WF-Priors [30] constructs joint wavelet–Fourier priors for zero-shot diffusion enhancement.

These studies integrate spatial and frequency information within complete enhancement architectures. Their frequency-domain operations are embedded as integral network stages rather than formulated as separate illumination-prior branches jointly optimized with a host backbone while leaving its main restoration blocks unchanged. This distinction motivates the formulation adopted in the present work.

2.4. Summary and Motivation

Spatial-domain priors capture local illumination variations, while frequency-domain priors characterize global brightness; in remote sensing imagery, where both large-area underexposure and heterogeneous land covers coexist, the two priors complement each other. Prior work has demonstrated that dual-domain processing is effective within complete network architectures; extending this capability to a dual-domain illumination-prior module jointly optimized with each tested host backbone while leaving its main restoration blocks unchanged is the central motivation of DDIP.

3. Method

We propose a jointly optimized illumination-prior module, the Dual-Domain Illumination Prior (DDIP), which comprises three components, as shown in Figure 1. The Frequency-Domain Illumination Distribution Prior (FIDP) performs per-color-channel Fourier amplitude calibration to correct global brightness degradation in low-light images. The Spatial-Domain Illumination Distribution Prior (SIDP), adapted from IDP-Net [19], conducts multi-scale sub-region statistical correction to address local illumination non-uniformity caused by heterogeneous land covers in remote sensing scenes. The Selective Core Feature Fusion (SCFF) module adaptively integrates the calibrated outputs from both domains with the original input, balancing global brightness restoration and local detail preservation. Although the abbreviation DIP is commonly used for Deep Image Prior [31], DDIP denotes the distinct Dual-Domain Illumination Prior proposed here; it is trained jointly with the host enhancement backbone rather than used as an instance-specific optimization framework.

3.1. Frequency-Domain Illumination Distribution Prior

For an input image x, we use the orthonormal two-dimensional discrete Fourier transform, defined as

$$\mathcal{F}\left(x\right)\left(u,v\right)={\displaystyle \frac{1}{\sqrt{HW}}}{\displaystyle \sum _{h=0}^{H-1}}{\displaystyle \sum _{w=0}^{W-1}}x\left(h,w\right)e^{-j2\pi \left({\displaystyle \frac{h}{H}}u+{\displaystyle \frac{w}{W}}v\right)},$$

(1)

where j is the imaginary unit, h and w are the spatial coordinates, H and W are the height and width of the image, and u and v are the coordinates in Fourier space. The inverse transform follows the corresponding orthonormal convention, consistent with the implementation.

The inverse Fourier transform is denoted by ${\mathcal{F}}^{-1}$. The amplitude component $\mathcal{A}\left(x\right)$ and the phase component $\mathcal{P}\left(x\right)$ are given by

$$\mathcal{A}\left(x\right)\left(u,v\right)=\sqrt{R^2\left(x\right)\left(u,v\right)+I^2\left(x\right)\left(u,v\right)},$$

(2)

$$\mathcal{P}\left(x\right)\left(u,v\right)=\mathrm{atan2}\left(I\left(x\right)\left(u,v\right),R\left(x\right)\left(u,v\right)\right),$$

(3)

where $R\left(x\right)$ and $I\left(x\right)$ are the real and imaginary parts of $\mathcal{F}\left(x\right)$, respectively.

As shown in Figure 2, FIDP operates on the red, green, and blue channels of the low-light image X independently. This separation allows the prior to accommodate channel-specific amplitude distributions without coupling the color channels during Fourier calibration. For channel $c\in \{R,G,B\}$, let $A_c$ and $P_c$ denote the Fourier amplitude and phase, respectively. An independent pointwise mapper $g_c$ produces an amplitude residual ${\Delta }_c$; the three mappers have the structure $\mathrm{Conv}(1\to 8)\to \mathrm{LeakyReLU}\left(0.2\right)\to \mathrm{Conv}(8\to 1)$ and do not share weights. The residual is converted into a bounded multiplicative modulation, after which the original phase is combined with the calibrated amplitude for inverse transformation. A channel-wise affine adjustment and clipping complete the branch. The formulation is as follows:

$$A_c=\left|\mathcal{F}\left(X_c\right)\right|,P_c=\angle \mathcal{F}\left(X_c\right),c\in \{R,G,B\},$$

(4)

$${\Delta }_c=g_c\left(A_c\right),A_c^{\prime }=A_c\left[1+\eta \tanh \left({\Delta }_c\right)\right],$$

(5)

$${\tilde{X}}_c=\mathrm{Re}\left\{{\mathcal{F}}^{-1}\left[A_c^{\prime }\exp \left(j{P}_c\right)\right]\right\},$$

(6)

$$\begin{array}{cc}\hfill X_c^{\mathrm{FIDP}}& =\mathrm{clip}\left({\gamma }_c{\tilde{X}}_c+{\beta }_c,0,1\right),\hfill \\ \hfill {\gamma }_c& ={\gamma }_{\min }+\left({\gamma }_{\max }-{\gamma }_{\min }\right)\mathrm{sigmoid}\left(a_c\right),{\beta }_c={\beta }_{\max }\tanh \left(b_c\right),\hfill \end{array}$$

(7)

$$X^{\mathrm{FIDP}}={\mathrm{Concat}}_{\mathrm{ch}}\left(X_R^{\mathrm{FIDP}},X_G^{\mathrm{FIDP}},X_B^{\mathrm{FIDP}}\right).$$

(8)

Here, $\eta$ controls the amplitude-residual range; $a_c$ and $b_c$ are unconstrained learned channel-wise affine parameters; $\mathrm{sigmoid}(·)$ denotes the logistic sigmoid function; and ${\gamma }_{\min }$, ${\gamma }_{\max }$, and ${\beta }_{\max }$ determine the admissible gain and shift ranges. The operator ${\mathrm{Concat}}_{\mathrm{ch}}(·)$ denotes concatenation along the RGB channel dimension. In the reported implementation, $\eta =0.5$, $({\gamma }_{\min },{\gamma }_{\max })=(0.8,1.2)$, and ${\beta }_{\max }=0.1$. Hence, the amplitude multiplier remains in $(0.5,1.5)$, the affine gain remains positive, and the channel-wise shift is bounded. The operator $\mathrm{clip}(·,0,1)$ restricts the reconstructed intensity to the normalized image range.

Frequency-domain processing offers a useful decomposition for low-light enhancement: the Fourier amplitude reflects global intensity statistics, while the phase component encodes the spatial arrangement of image structures. FIDP calibrates the amplitude and reuses the original phase during inverse reconstruction. Figure 3 visualizes the low-light input, the FIDP prior, the final enhanced output, and the log-amplitude change induced by FIDP; Figure 4 shows the corresponding phase behavior. These figures illustrate the operation of FIDP but are not used as quantitative evidence of exact spectral matching. In optical remote sensing imagery, the three RGB channels may exhibit different intensity distributions; FIDP therefore processes the channels independently.

For a real-valued input channel, the Fourier spectrum is conjugate symmetric. Since $A_c$ is symmetric and the same pointwise $1\times 1$ amplitude mapper $g_c$ is applied at every frequency location, the calibrated amplitude $A_c^{\prime }$ retains this symmetry. Combining $A_c^{\prime }$ with the unchanged phase $P_c$ therefore preserves conjugate symmetry up to floating-point precision. We verified this numerically on the 66 paired iSAID-dark validation images used for the fixed-checkpoint diagnostic: before taking the real component after inverse FFT, the aggregate mean and maximum magnitudes of the imaginary residual were $3.40\times {10}^{-8}$ and $8.85\times {10}^{-7}$, respectively, and the ratio between mean imaginary and mean real magnitudes was $1.40\times {10}^{-7}$. Thus, discarding the residual imaginary component removes only numerical round-off rather than image information.

Scope beyond RGB. The present implementation and all reported experiments operate on three-channel RGB optical imagery. In principle, the formulation could be extended to a B-band remote sensing image by applying FIDP amplitude calibration and SIDP local-statistics correction on a per-band basis. However, multispectral data can exhibit band-dependent radiometric properties and cross-band correlations; whether the corresponding mappers should be independently learned, partially shared, or explicitly coupled requires dedicated empirical study. This is therefore a theoretical extension only and does not establish performance on multispectral imagery.

3.2. Spatial-Domain Illumination Distribution Prior (Based on IDP-Net [19])

Building on the statistical relationship among per-channel mean, standard deviation, and illumination established by IDP-Net [19], we adapted the Spatial-Domain Illumination Distribution Prior (SIDP) with a stagewise multi-scale refinement strategy. As shown in Figure 5, the input is partitioned into a medium-scale $4\times 4$ grid, a fine-scale $16\times 16$ grid, and a single global region. At each level, per-channel mean and standard deviation are computed within every region, and learnable affine mappings adjust these statistics toward their normal-light counterparts. The calibrated outputs are applied sequentially from medium to fine to global scale, producing the final output $X_{\mathrm{SIDP}}$.

The three-scale refinement strategy is adapted from IDP-Net [19]. The representative configuration uses the ordered grid sizes $[4,16,1]$, corresponding to medium-scale, fine-scale, and global calibration. This compact design introduces 36 learnable scalar parameters and 0.16 GFLOPs at $512\times 512$ resolution; the underlying statistical formulation follows IDP-Net [19].

3.3. Selective Core Feature Fusion

Simple feature aggregation methods such as summation or concatenation may not adequately balance the distinct characteristics of frequency-domain and spatial-domain branch outputs. Summation imposes uniform weighting, while concatenation stacks the outputs without modeling inter-domain interactions. Inspired by prior feature fusion strategies [10,32], we designed the Selective Core Feature Fusion (SCFF) to adaptively combine the FIDP branch, the SIDP branch, and the original input image. SCFF employs dual pooling (global average pooling, GAP, and global max pooling, GMP) to generate channel descriptors: GAP captures the global distribution of each input channel, while GMP retains the most salient activation, together providing complementary characterizations of channel-wise importance.

As shown in Figure 6, SCFF balances the three branch contributions through weight estimation and information fusion. The former generates per-channel attention weights from dual-pooling descriptors and a branch-wise softmax, whereas the latter combines the sources by channel-wise weighted summation.

Weight Estimation. SCFF considers the FIDP output $\left(L_1\right)$, the input low-light image $\left(L_2\right)$, and the SIDP output $\left(L_3\right)$ as three candidate sources and forms $L=L_1+L_2+L_3$. Global average pooling and global max pooling are applied to L, and the resulting descriptors are aggregated as s. The pooled descriptor is first compressed by a $1\times 1$ convolution and PReLU; a second $1\times 1$ projection produces the branch-wise logits $V_1$, $V_2$, and $V_3$. After stacking and reshaping to $[B,3,C,1,1]$, these logits are normalized by a branch-wise softmax with temperature $\tau$. The resulting tensor $\mathbf{s}\in {[0,1]}^{B\times 3\times C\times 1\times 1}$ supplies channel-wise reliability scores whose three branch values sum to one. In the reported implementation, $\tau =2.0$. The scores are broadcast over the spatial dimensions when applied to the corresponding branch outputs, allowing SCFF to emphasize the source that is more informative for a given channel and degradation condition.

Information Fusion. The three branch outputs $L_1$, $L_2$, and $L_3$ lie in $[0,1]$, so no additional rescaling is required. Their reliability-weighted aggregation is $U=s_1\odot L_1+s_2\odot L_2+s_3\odot L_3$, where $s_1,s_2,s_3\in {[0,1]}^{B\times C\times 1\times 1}$ are the FIDP, original-input, and SIDP scores extracted from $\mathbf{s}$, and ⊙ denotes element-wise multiplication with spatial broadcasting. An explicit fidelity path then produces $U_{\mathrm{out}}=(1-\alpha )U+\alpha L_2$, where $\alpha \in [0,1]$ controls the contribution of the original input. We use $\alpha =0.3$ in the reported implementation. Consequently, the original input maintains a direct contribution to the fused representation even when selective attention favors one calibrated branch. After projection, $U_{\mathrm{out}}$ is added to the shallow features of the host backbone, as shown in Figure 7. The main restoration blocks remain unchanged, and MIRNet-v2 retains its original image-residual path.

Through dual-pooling reliability estimation, selective branch weighting, and an explicit input-fidelity path, SCFF integrates the candidate dual-domain cues without requiring equal branch contributions. Its role is to select useful illumination information while maintaining a direct path from the input before the host backbone performs subsequent texture restoration.

3.4. Loss Function

The training objective consists of two components: a reconstruction loss that drives pixel-level enhancement quality, and a frequency-domain amplitude loss that supervises global illumination calibration.

For the reconstruction loss, we adopt the ${\mathcal{L}}_1$ loss between the enhanced image $\widehat{X}$ and the ground-truth image $GT$:

$${\mathcal{L}}_r={∥GT-\widehat{X}∥}_1,$$

(9)

where ${\parallel ·\parallel }_1$ denotes the ${\mathcal{L}}_1$ norm.

For the amplitude loss, we directly supervise the FIDP output in the frequency domain, since the Fourier amplitude is closely correlated with illumination. This term encourages FIDP to function as a global illumination calibrator, while the reconstruction loss allows the downstream backbone to refine residual textures and local structures. The amplitude loss is defined as:

$${\mathcal{L}}_{amp}={∥\mathcal{A}\left(GT\right)-\mathcal{A}\left(X^{\mathrm{FIDP}}\right)∥}_1,$$

(10)

where $\mathcal{A}\left(GT\right)$ and $\mathcal{A}\left(X^{\mathrm{FIDP}}\right)$ denote the amplitude components of the ground truth and the FIDP output, respectively.

The total loss is a weighted sum of the two terms:

$${\mathcal{L}}_{total}={\mathcal{L}}_r+\lambda {\mathcal{L}}_{amp},$$

(11)

where $\lambda$ is the amplitude-loss weight. The reported main results use $\lambda =0.01$; its sensitivity within the tested range is examined in Section 4.5.

4. Results and Discussion

The experiments cover enhancement quality on iSAID-dark, a fixed-detector object-detection evaluation on high-resolution scenes, supplementary results on five general low-light benchmarks, and controlled analyses of the DDIP components and design parameters.

4.1. Datasets and Experimental Setup

Datasets. The primary evaluation was conducted on the publicly available iSAID-dark dataset released by Yao et al. [4]. This dataset was constructed from the iSAID aerial image benchmark by applying a physically motivated illumination degradation model to high-quality daytime images, producing varying exposure levels and spatially uneven lighting. The prepared split used by our enhancement experiments contained 3755 paired training images and 66 paired validation images at a spatial resolution of $512\times 512$ pixels. The scenes contain complex terrain textures and dense multi-scale objects, making illumination correction and detail preservation challenging.

General Benchmarks. Supplementary results are reported on five low-light datasets: LOL-v1 [33], LOL-v2-Real [34], LOL-v2-Syn [34], LSRW-Nikon [35], and LSRW-Huawei [35] (training/test splits: 485/15, 689/100, 900/100, 3150/20, and 2450/30, respectively).

Comparison Baselines. The controlled results were the adjacent backbone/DDIP pairs, for which the baseline and DDIP-equipped variant were trained and evaluated under matched settings within each pair. FECNet, SNR-Net, FourLLIE, LLFormer, and ZERO-IG are literature-based contextual references reported with citations to their source methods, whereas DFFN was evaluated using the authors’ publicly released iSAID checkpoint. These contextual rows were not treated as results reproduced under the matched host-backbone protocols. Accordingly, only the adjacent backbone/DDIP pairs and their explicitly reported “+” gains were interpreted as controlled comparisons in this study; no gain was computed against a contextual method row.

Training Configuration. For each controlled backbone/DDIP pair, the baseline and DDIP-equipped variant used matched data splits, preprocessing, optimization schedules, patch and batch settings, iteration budgets, and evaluation scripts. No official pre-trained weights were used for either member of a controlled pair. DDIP was integrated as an illumination-prior branch before the main restoration blocks. This description excludes the separately reported DFFN reference, for which the released checkpoint was used as stated below. Unless otherwise stated, the principal reported setup used PyTorch 1.13 [36] and an NVIDIA RTX 4090 GPU. The learning rate was initialized at $3.0\times {10}^{-4}$ and decreased to $1.0\times {10}^{-6}$ using cosine annealing [37]. During training, input images were randomly cropped to $256\times 256$ patches. We employed the Adam optimizer [38] with ${\beta }_1=0.9$ and ${\beta }_2=0.999$. The principal iteration budget was $3.0\times {10}^5$ with a batch size of four. The Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) [39] were used as standard evaluation metrics. In all tables, the upward arrow (↑) indicates that higher values are preferable, and the downward arrow (↓) indicates that lower values are preferable.

For the complementary perceptual and illumination-oriented analysis, we additionally report the Learned Perceptual Image Patch Similarity (LPIPS) [40], the mean absolute error of the CIELAB lightness channel (Lab $L^{*}$ MAE), and the signed mean $L^{*}$ bias on the 66 paired images in the iSAID-dark validation set under the Standard setting. LPIPS was computed using the official AlexNet linear model. LPIPS and Lab $L^{*}$ MAE compare the enhanced result with the paired normal-light reference, while the signed mean $L^{*}$ bias indicates whether the output is globally darker or brighter than that reference; a bias magnitude closer to zero indicates closer global lightness. In addition to the controlled backbone/DDIP pairs, DFFN was evaluated using the authors’ publicly released iSAID checkpoint as an external architecture-level reference. These metrics were computed after inference and were not used as training losses.

4.2. Low-Light Remote Sensing Image Enhancement

Experiment Settings. DDIP was evaluated with CNN-, Transformer-, and Mamba-based backbones on iSAID-dark [4] under two settings. In the Standard setting, models were trained and tested on $512\times 512$ images. In the High-resolution setting, the models trained at $512\times 512$ were applied directly to full-resolution $1920\times 1080$ aerial images without fine-tuning. This cross-resolution setting tested whether illumination correction and fine-grained geospatial detail could be retained at a substantially higher input resolution.

Analysis of Results.

Table 1. Quantitative comparison on the iSAID-dark dataset. Boldface identifies the DDIP-equipped results; parenthetical values give the gain over the corresponding backbone. The DFFN row provides an architecture-level contextual reference; the controlled comparisons are the adjacent backbone/DDIP pairs.

Method	iSAID-Dark (Standard)		iSAID-Dark (High-Resolution)
	PSNR ↑	SSIM ↑	PSNR ↑	SSIM ↑
DFFN (2024) [4]	25.30	0.784	22.47	0.721
RetinexMamba (2024) [14]	25.30	0.834	22.88	0.730
RetinexMamba+DDIP	25.80 (+0.50)	0.842 (+0.008)	23.20 (+0.32)	0.738 (+0.008)
SPJFNet (2026) [23]	21.50	0.743	18.55	0.521
SPJFNet+DDIP	23.50 (+2.00)	0.771 (+0.028)	20.24 (+1.69)	0.586 (+0.065)
MIRNet-v2 (2022) [11]	25.50	0.791	22.83	0.672
MIRNet-v2+DDIP	25.94 (+0.44)	0.857 (+0.066)	23.42 (+0.59)	0.684 (+0.012)
Restormer (2022) [13]	24.10	0.763	22.40	0.640
Restormer+DDIP	25.60 (+1.50)	0.776 (+0.013)	22.90 (+0.50)	0.660 (+0.020)
Retinexformer (2023) [12]	25.65	0.778	22.96	0.670
Retinexformer+DDIP	26.04 (+0.39)	0.795 (+0.017)	23.10 (+0.14)	0.680 (+0.010)

reports the quantitative results on iSAID-dark. Each DDIP-equipped backbone improves upon its corresponding baseline in the reported PSNR and SSIM comparisons. In the Standard setting, Retinexformer+DDIP achieves the highest PSNR of 26.04 dB, while SPJFNet [23] + DDIP improves by 2.0 dB over its base network.

Complementary Illumination and Perceptual Analysis.

Table 2. Complementary perceptual and illumination-oriented evaluation on the 66 paired images in the iSAID-dark validation set under the Standard setting. All metrics use the paired normal-light reference. Lower LPIPS and Lab $L^{*}$ MAE are preferable; an absolute mean $L^{*}$ bias closer to zero is preferable.

Input/Enhancement Pipeline	LPIPS ↓	Lab ${\mathit{L}}^{*}$ MAE ↓	Mean ${\mathit{L}}^{*}$ Bias (→0)
Raw low-light input	0.8464	29.1586	$-28.9468$
DFFN (official checkpoint) [4]	0.2170	4.2533	$+0.0320$
RetinexMamba	0.2222	4.6589	$-2.1474$
RetinexMamba+DDIP	0.2261	4.2976	$-1.3362$
MIRNet-v2	0.2223	4.6245	$-1.2311$
MIRNet-v2+DDIP	0.2143	4.2304	$-1.1555$
Restormer	0.2320	5.0772	$-2.2570$
Restormer+DDIP	0.2322	4.4443	$-1.7555$
Retinexformer	0.2271	4.1006	$-0.9762$
Retinexformer+DDIP	0.2158	4.2614	$-1.2401$

provides complementary evidence beyond the primary PSNR/SSIM evaluation. With RetinexMamba, DDIP reduces Lab $L^{*}$ MAE from 4.6589 to 4.2976 and the absolute mean $L^{*}$ bias from 2.1474 to 1.3362, while LPIPS increases from 0.2222 to 0.2261. With MIRNet-v2, DDIP reduces LPIPS from 0.2223 to 0.2143, Lab $L^{*}$ MAE from 4.6245 to 4.2304, and the absolute mean $L^{*}$ bias from 1.2311 to 1.1555. With Restormer, DDIP reduces Lab $L^{*}$ MAE by 12.47% and the absolute mean $L^{*}$ bias by 22.22%, while LPIPS remains nearly unchanged. With Retinexformer, DDIP reduces LPIPS by 4.98%, but Lab $L^{*}$ MAE increases by 3.92% and the absolute mean $L^{*}$ bias changes from 0.9762 to 1.2401. Thus, changes in the complementary metrics vary across host architectures, and no single perceptual or illumination metric is used to claim uniform superiority. DFFN is included as an external architecture-level reference rather than a paired DDIP ablation. All values were computed after inference using saved outputs and did not enter the training objectives.

DFFN is a representative dual-domain architecture for low-light remote sensing image enhancement. In the Standard setting of Table 1, the PSNR values of Retinexformer+DDIP (26.04 dB) and MIRNet-v2+DDIP (25.94 dB) exceed the contextual DFFN result (25.30 dB) by 0.74 dB and 0.60 dB, respectively. Because DFFN was evaluated as a separate end-to-end architecture, these differences are not treated as paired DDIP gains.

The qualitative comparisons in Figure 8 and Figure 9 further illustrate the reported results on the iSAID-dark dataset under the Standard and High-resolution settings. The baseline SPJFNet [23] shows a performance decline when applied to the high-resolution inputs (PSNR decreasing from 21.50 dB to 18.55 dB). In this setting, SPJFNet+DDIP’s PSNR is 1.69 dB higher than that of SPJFNet. These results suggest that dual-domain illumination calibration merits further study for remote-sensing imagery at varying resolutions. The gains reported for the tested CNN-, Transformer-, and Mamba-based backbones are specific to iSAID-dark and the stated experimental configuration. DDIP is jointly optimized with each host backbone; the frequency-domain branch is intended to calibrate global brightness, while the spatial-domain branch is intended to address localized illumination variation.

4.3. Downstream Object Detection Evaluation

Experiment Settings. To assess whether the tested enhancement outputs benefited a downstream remote-sensing task, we trained a YOLOv8n detector once on normal-illumination iSAID images derived from DOTA v1.0, using the corresponding official iSAID instance annotations. The detector was then fixed; no detector retraining was performed for any low-light or enhanced input. The test set comprised 24 high-resolution iSAID-dark validation scenes, partitioned into 144 detector tiles. For each scene, we evaluated the raw synthetic low-light images at three illumination levels (Low-15, Low-20, and Low-30), together with the outputs of RetinexMamba, RetinexMamba+DDIP, MIRNet-v2, MIRNet-v2+DDIP, Restormer, Restormer+DDIP, Retinexformer, Retinexformer+DDIP, and DFFN. All variants used the same image tiles and annotations. DFFN used the authors’ publicly released iSAID checkpoint and was included as an external architecture-level reference; it was not treated as a controlled DDIP pair. We report COCO-style ${\mathrm{mAP}}_{50:95}$ (denoted as mAP), ${\mathrm{mAP}}_{50}$, and small-object AP. These checkpoints are used here as application-oriented comparisons of enhancement pipelines; this experiment does not replace the controlled backbone-level ablations in Section 4.5.

Analysis of Results. The fixed-detector evaluation in

Table 3. Downstream object-detection performance (%) under three synthetic low-light levels. A single fixed YOLOv8n detector is used for all inputs. Each entry reports ${\mathrm{mAP}}_{50:95}/{\mathrm{mAP}}_{50}$; the final column is the mean ${\mathrm{AP}}_{\mathrm{small}}$ over the three low-light levels.

Input Image	Low-15	Low-20	Low-30	Mean	Mean ${\mathbf{AP}}_{\mathbf{small}}$
Raw low-light	2.71/4.12	5.91/9.26	18.07/28.06	8.90/13.82	4.65
DFFN (official checkpoint) [4]	26.18/40.99	29.50/44.80	33.58/51.01	29.75/45.60	15.76
RetinexMamba	25.66/40.49	28.79/43.52	33.59/51.80	29.35/45.27	14.83
RetinexMamba+DDIP	25.23/39.80	29.91/45.00	33.13/50.56	29.42/45.12	15.41
MIRNet-v2	26.50/41.22	30.61/46.17	33.47/50.56	30.19/45.98	17.56
MIRNet-v2+DDIP	26.41/41.65	31.39/47.09	34.20/51.58	30.67/46.77	16.64
Restormer	24.26/38.15	29.18/43.67	33.15/50.24	28.86/44.02	15.90
Restormer+DDIP	25.51/41.18	29.53/45.37	32.92/51.76	29.32/46.10	14.82
Retinexformer	25.94/41.13	30.01/45.30	34.22/52.88	30.06/46.44	16.61
Retinexformer+DDIP	26.15/40.90	30.30/46.02	33.85/51.59	30.10/46.17	16.10

shows that the tested enhancement pipelines substantially improve detection performance relative to the raw synthetic low-light inputs under the evaluated conditions. Relative to the corresponding host backbones, DDIP increases mean mAP by 0.48, 0.46, and approximately 0.08 percentage points for MIRNet-v2, Restormer, and RetinexMamba, respectively; the last value was calculated from the unrounded results. For Retinexformer, the mean mAP changes only marginally, whereas its mean ${\mathrm{mAP}}_{50}$ and mean ${\mathrm{AP}}_{\mathrm{small}}$ decrease. For RetinexMamba, the mean ${\mathrm{AP}}_{\mathrm{small}}$ increases from 14.83% to 15.41%, and the largest gain occurs under the Low-20 condition; its mean ${\mathrm{mAP}}_{50}$ decreases slightly from 45.27% to 45.12%. The improvement is therefore more evident for the MIRNet-v2 and Restormer hosts and is not uniform across all backbones, degradation levels, or metrics. In particular, the present experiment does not support the claim that DDIP consistently improves small-object AP. The official DFFN checkpoint provides an external architecture-level reference of 29.75% mean mAP and 45.60% mean ${\mathrm{mAP}}_{50}$; this row is not used to infer a DDIP gain.

4.4. Supplementary Evaluation on Standard Low-Light Benchmarks

We additionally evaluated DDIP on five standard low-light benchmarks. While the primary contribution targets remote sensing imagery (Section 4.2), these experiments assess DDIP in the reported non-remote-sensing settings and do not constitute additional remote-sensing validation.

Experiment Settings. MIRNet-v2 [11], Restormer [13], Retinexformer [12], and EvLight [41] served as base networks, with DDIP inserted before each for controlled backbone-versus-backbone+DDIP comparisons (see Figure 7). Additional baseline methods are listed in Section 4.1. A cross-dataset evaluation was performed between LOL-v2-Real [34] and LSRW-Huawei [35]: models trained on one dataset were tested on the other under the specified train–test pairing.

Analysis of Results.

Table 4. Quantitative comparison of different methods on LOL-v1, LOL-v2-Real, and LOL-v2-Syn datasets. The controlled evidence is the adjacent backbone/DDIP pairing; other method rows are contextual references and are not used to calculate the reported gains.

Method	LOL-v1		LOL-v2-Real		LOL-v2-Syn
	PSNR ↑	SSIM ↑	PSNR ↑	SSIM ↑	PSNR ↑	SSIM ↑
FECNet (2022) [21]	23.44	0.838	20.67	0.795	22.57	0.894
SNR-Net (2022) [17]	24.15	0.840	21.48	0.848	24.57	0.899
FourLLIE (2023) [20]	24.15	0.839	22.34	0.848	24.57	0.899
LLFormer (2023) [42]	23.65	0.843	22.40	0.838	24.60	0.925
ZERO-IG (2024) [43]	22.18	0.772	20.76	0.833	22.66	0.890
MIRNet-v2 (2022) [11]	24.74	0.851	21.98	0.833	22.64	0.890
MIRNet-v2+DDIP	25.38	0.856	22.48	0.835	23.98	0.907
Restormer (2022) [13]	22.43	0.823	21.98	0.837	24.11	0.926
Restormer+DDIP	23.37	0.862	22.59	0.842	25.12	0.930
Retinexformer (2023) [12]	25.16	0.845	22.80	0.840	25.67	0.930
Retinexformer+DDIP	25.44	0.847	23.25	0.845	26.08	0.939
EvLight (2024) [41]	25.54	0.848	22.65	0.850	25.36	0.932
EvLight+DDIP	25.84	0.852	23.07	0.853	25.88	0.937

and

Table 5. Quantitative comparison on LSRW-Huawei and LSRW-Nikon datasets. Boldface identifies the DDIP-equipped results; parenthetical values give the gain over the corresponding backbone. The controlled evidence is the adjacent backbone/DDIP pairing; other method rows are contextual references and are not used to calculate the reported gains.

Method	LSRW-Huawei		LSRW-Nikon
	PSNR ↑	SSIM ↑	PSNR ↑	SSIM ↑
FECNet (2022) [21]	21.09	0.612	17.06	0.500
SNR-Net (2022) [17]	20.67	0.591	17.54	0.482
FourLLIE (2023) [20]	21.30	0.622	17.82	0.504
LLFormer (2023) [42]	20.36	0.607	17.94	0.519
ZERO-IG (2024) [43]	19.84	0.594	16.22	0.471
MIRNet-v2 (2022) [11]	19.81	0.608	17.30	0.506
MIRNet-v2+DDIP	20.11 (+0.30)	0.614 (+0.006)	17.67 (+0.37)	0.508 (+0.002)
Restormer (2022) [13]	21.01	0.609	18.18	0.501
Restormer+DDIP	21.42 (+0.41)	0.615 (+0.006)	18.39 (+0.21)	0.505 (+0.004)
Retinexformer (2023) [12]	21.37	0.631	18.06	0.517
Retinexformer+DDIP	21.61 (+0.24)	0.641 (+0.010)	18.44 (+0.38)	0.523 (+0.006)
EvLight (2024) [41]	22.34	0.629	18.54	0.527
EvLight+DDIP	22.73 (+0.39)	0.644 (+0.015)	18.79 (+0.25)	0.531 (+0.004)

report the quantitative results on the LOL-v1, LOL-v2-Real, LOL-v2-Syn, LSRW-Nikon, and LSRW-Huawei datasets. Across these benchmarks, equipping the reported base networks with DDIP yields positive PSNR and SSIM differences. The largest PSNR increases on the five datasets are 0.94 dB, 0.61 dB, 1.34 dB, 0.41 dB, and 0.38 dB, respectively. The qualitative comparison in Figure 10 suggests that DDIP contributes to correcting the illumination distribution of the base networks while preserving spatial structures. These results are limited to the evaluated datasets and training settings.

Table 6. Cross-dataset evaluation results. Models were trained on one dataset and tested on the other. Boldface identifies the DDIP-equipped results; parenthetical values give the gain over the corresponding backbone. The paired host-backbone rows constitute the controlled DDIP comparisons; the FECNet, FourLLIE, and LLFormer rows are contextual references.

Method	LOL-v2-Real → Huawei		Huawei → LOL-v2-Real
	PSNR ↑	SSIM ↑	PSNR ↑	SSIM ↑
FECNet [21]	17.19	0.583	18.40	0.675
FourLLIE [20]	16.23	0.555	18.57	0.662
LLFormer [42]	15.33	0.486	17.95	0.582
MIRNet-v2 [11]	17.87	0.516	18.16	0.682
MIRNet-v2+DDIP	18.58 (+0.71)	0.538 (+0.022)	19.46 (+1.30)	0.724 (+0.042)
Restormer [13]	19.58	0.547	20.14	0.705
Restormer+DDIP	19.94 (+0.36)	0.555 (+0.008)	20.76 (+0.62)	0.728 (+0.023)
Retinexformer [12]	19.02	0.583	20.22	0.717
Retinexformer+DDIP	19.31 (+0.29)	0.594 (+0.011)	20.74 (+0.52)	0.734 (+0.017)

reports the cross-dataset results. FECNet, FourLLIE, and LLFormer are included as reference baselines; because their architectures are not used as modular host backbones here, DDIP is integrated into MIRNet-v2, Restormer, and Retinexformer for the paired comparisons. In this setting, the DDIP-equipped backbones retain positive PSNR and SSIM gains over their corresponding baselines. This finding is restricted to the two reported train–test dataset pairs.

4.5. Ablation Study

To examine the contribution of each component, we conducted controlled backbone-level ablations on the iSAID-dark [4] dataset with MIRNet-v2 under the Standard setting.

Table 7. Backbone-level ablation studies on the DDIP components. Backbone: MIRNet-v2. Dataset: iSAID-dark (Standard setting).

Method	FIDP	SIDP	SCFF	PSNR ↑	SSIM ↑
Only FIDP	Yes	No	No	25.54	0.7985
Only SIDP	No	Yes	No	25.57	0.8004
No SCFF	Yes	Yes	No	25.68	0.7994
Baseline	No	No	No	25.50	0.7913
Ours	Yes	Yes	Yes	25.94	0.8570

evaluates FIDP, SIDP, and SCFF within the full enhancement pipeline.

Table 8. Ablation studies on feature fusion strategies. The reported FLOPs and parameters denote the total complexity of the standalone fusion-analysis configuration used for this component-level comparison, rather than the parameter count of the fusion block alone. Differences among rows indicate the additional cost incurred by replacing the fusion strategy.

Method	PSNR ↑	SSIM ↑	FLOPs (G)	Params (M)
Sum	20.73	0.741	0.544	43.03
Concat	20.70	0.742	0.544	43.05
SKFF	21.37	0.751	0.546	43.08
SCFF (Ours)	21.49	0.754	0.546	43.09

provides a separate standalone fusion analysis on LOL-v1 [33]; because it uses a different dataset and evaluation configuration, it is not combined with Table 7 to estimate the total DDIP overhead. The ablation focused on controlled component and fusion comparisons. We did not extract modules from other complete architectures for module-only comparisons because doing so would not have reproduced their original end-to-end systems.

Component Ablation. The “Only FIDP” and “Only SIDP” configurations isolate the frequency-domain and spatial-domain priors, respectively, while “No SCFF” retains both prior branches without the selective fusion module. Table 7 shows that the full DDIP configuration performs better than either single-domain configuration and the version without SCFF. It achieves a PSNR of 25.94 dB and SSIM of 0.8570, compared with 25.50 dB and 0.7913 for the baseline. Within this controlled setting, the results support the joint use of the two illumination priors and selective cross-domain fusion.

SIDP Scale and Order Sensitivity. The default sequence $[4,16,1]$ combines medium-scale ($4\times 4$), fine-scale ($16\times 16$), and global calibration so that the spatial branch can model heterogeneous local illumination together with scene-wide brightness. In the retraining-controlled rows of

Table 9. SIDP scale-count and order sensitivity on the iSAID-dark Standard validation setting with MIRNet-v2+DDIP. “Retrained” denotes independently optimized scale-count configurations. “Fixed checkpoint” denotes inference using the same saved network weights while changing only the order of the three learned SIDP levels on 66 paired validation images. Results should be compared within the same protocol. The grid size n denotes the number of sub-regions per spatial dimension; $n=1$ denotes global calibration.

SIDP Grid Sizes n	Spatial Levels	Protocol	No. of Levels	Parameters	PSNR ↑	SSIM ↑
Three-level default $[4,16,1]$	Medium, fine, global	Retrained	3	36	25.9400	0.8570
Two-level variant $[4,1]$	Medium, global	Retrained	2	24	25.9368	0.8564
Default $[4,16,1]$	Medium, fine, global	Fixed checkpoint	3	36	25.9471	0.8522
Local-order-reversed $[16,4,1]$	Fine, medium, global	Fixed checkpoint	3	36	25.9481	0.8522
Global-first $[1,4,16]$	Global, medium, fine	Fixed checkpoint	3	36	25.9436	0.8522

, removing the fine-scale level changes PSNR by $-0.0032$ dB and SSIM by $-0.0006$. Thus, after retraining, the tested scale-count change produces only a small difference. The fixed-checkpoint rows examine processing order while preserving the same learned parameters. The default, local-order-reversed, and global-first sequences obtain 25.9471/0.8522, 25.9481/0.8522, and 25.9436/0.8522, respectively. The 0.0045 dB PSNR span and unchanged four-decimal SSIM show limited order sensitivity for the tested images. This fixed-checkpoint result is interpreted as an inference-time diagnostic rather than an independently trained architectural ablation, and values from the two protocols are not compared directly. The three-level formulation remains the hierarchical default rather than being presented as uniquely optimal; the preferred spatial partition may depend on image resolution and illumination degradation.

Amplitude-Loss-Weight Sensitivity.

Table 10. Sensitivity to the FIDP amplitude-loss weight $\lambda$ on the iSAID-dark Standard validation setting.

$\mathit{\lambda }$	PSNR ↑	SSIM ↑
0	25.9604	0.8545
0.001	26.0386	0.8552
0.01	25.9400	0.8570

evaluates $\lambda \in \{0,0.001,0.01\}$. The maximum PSNR variation is 0.0986 dB. The setting $\lambda =0.001$ obtains the highest observed PSNR, whereas $\lambda =0.01$ obtains the highest observed SSIM. These small differences indicate that DDIP is not highly sensitive to the tested range of amplitude-loss weights; no statistical-significance claim is made from these single runs. The reported main results and the scale comparison retain $\lambda =0.01$ to preserve the original training protocol.

SCFF Ablation. We conducted a component-level comparison on LOL-v1 [33] between SCFF and three alternative aggregation strategies: Sum, Concat, and Selective Kernel Feature Fusion (SKFF) [10]. Only the fusion strategy was replaced, while the remaining standalone dual-domain analysis configuration was kept unchanged. Therefore, the FLOPs and parameter counts in Table 8 should be interpreted as the total complexity of the standalone fusion-analysis configuration rather than the parameter size of each fusion block alone. The differences among rows reflect the relative cost introduced by each fusion strategy. Compared with SKFF, SCFF obtains a 0.12 dB higher PSNR and 0.003 higher SSIM under the reported configuration, with the same rounded FLOPs.

SCFF Parameter Sensitivity.

Table 11. Fixed-checkpoint inference sensitivity of SCFF on 66 paired iSAID-dark validation images. Changes are reported relative to the default setting $(\tau ,\alpha )=(2.0,0.3)$ using the same saved checkpoint; this post hoc diagnostic does not retrain the model or replace the controlled results in Table 7.

Setting	$\mathit{\tau }$	$\mathit{\alpha }$	ΔPSNR (dB)	ΔSSIM
Default	2.0	0.3	0.0000	0.0000
Lower temperature	1.0	0.3	0.0000	0.0000
Higher temperature	4.0	0.3	0.0000	0.0000
No fidelity blending	2.0	0.0	$-0.0899$	$+0.0041$
Stronger fidelity blending	2.0	0.6	$-0.6230$	$-0.0124$

reports a fixed-checkpoint inference diagnostic. Varying $\tau$ from 1.0 to 4.0 leaves the measured PSNR and SSIM unchanged at the reported precision. For the fidelity coefficient, the default $\alpha =0.3$ gives the highest PSNR among the tested values. Setting $\alpha =0$ disables the explicit input-fidelity path and reduces PSNR by 0.0899 dB, whereas increasing $\alpha$ to 0.6 reduces both PSNR and SSIM. The small SSIM increase at $\alpha =0$ reflects a PSNR–SSIM trade-off rather than a uniform gain. This fixed-checkpoint analysis characterizes local sensitivity around the selected configuration and is not interpreted as a retraining-controlled architectural ablation.

Inference Overhead of DDIP.

Table 12. Inference latency and throughput comparison of MIRNet-v2 with and without DDIP, measured on an NVIDIA RTX 4090 GPU at two input resolutions. Parameter counts are static implementation values; DDIP adds 502 parameters.

Method	Params (M)	512 × 512		1920 × 1080
		Latency (ms)	FPS	Latency (ms)	FPS
MIRNet-v2 (2022)	5.858560	262.0	3.82	2082.0	0.48
MIRNet-v2+DDIP	5.859062	276.6	3.62	2208.0	0.45
Overhead	$+0.000502$	$+14.6$	$-0.20$	$+126.0$	$-0.03$

reports end-to-end inference latency and throughput for MIRNet-v2 with and without DDIP at two representative resolutions. At $512\times 512$, the reported DDIP configuration adds 14.6 ms of latency (a 5.6% relative increase), with FPS changing from 3.82 to 3.62. At $1920\times 1080$, the reported absolute latency difference is 126.0 ms, with FPS changing from 0.48 to 0.45. Static inspection of the tested implementation further shows that DDIP adds 502 parameters (from 5,858,560 to 5,859,062 parameters; 0.009% of the host parameter count). The numbers of incremental parameters are 81 in FIDP, 36 in SIDP, 65 in SCFF, and 320 in the prior-projection layer. These measurements describe the tested implementation, hardware, and host backbone; they do not provide module-level profiling or establish latency behavior for untested backbones or resolutions.

5. Conclusions

This paper presented a Dual-Domain Illumination Prior (DDIP), a jointly optimized illumination-prior module for low-light remote sensing image enhancement. Component ablations examined the contributions of FIDP, SIDP, and SCFF by comparing the full dual-domain configuration with single-domain and reduced-fusion variants. The controlled reconstruction comparisons on iSAID-dark reported PSNR and SSIM gains for the tested CNN-, Transformer-, and Mamba-based backbones. The fixed-detector analysis on high-resolution iSAID-dark scenes showed that the tested enhancement pipelines improved detection performance relative to raw low-light inputs under the evaluated protocol, whereas the incremental DDIP effect was backbone- and metric-dependent and did not consistently improve mean ${\mathrm{AP}}_{\mathrm{small}}$ across the reported pairs. The supplementary evaluation on five general low-light benchmarks reported results beyond the remote-sensing setting but did not establish general performance for untested data or sensors.

For remote sensing applications, DDIP addresses large-area underexposure through frequency-domain amplitude calibration and localized illumination non-uniformity through multi-scale spatial statistical correction. The evaluation included PSNR, SSIM, LPIPS, and Lab $L^{*}$ measurements, together with a fixed-detector object-detection experiment on synthetic high-resolution iSAID-dark scenes. Nevertheless, the current evidence is limited to RGB optical imagery, synthetic degradation, 24 high-resolution scenes comprising 144 detector tiles, and one detector. It does not establish performance on real nighttime remote-sensing imagery, other sensing modalities, or untested sensors and platforms. The downstream comparison is an application-oriented evaluation rather than a retraining-controlled ablation of DDIP. Broader real-world and multi-sensor validation remains necessary.