Degradation-aware time step guided single-step diffusion remote sensing image super-resolution reconstruction method

Abstract

The present application belongs to the technical field of image processing and computer vision, aiming at the shortcomings of existing reconstruction methods, a single-step diffusion remote sensing image super-resolution reconstruction method guided by degradation perception time step is proposed, a degradation estimation network is constructed, a number of scalar scores representing the degradation degree of the image are estimated and output, and the starting time step injected in the single-step diffusion reconstruction is converted, the adaptive mapping from the degradation degree to the diffusion time step is carried out, the adaptive time step is used as the starting point of the single-step diffusion reconstruction, the remote sensing image reconstruction from low resolution to high resolution is completed, the optimal balance of fidelity and generation effect is realized, and the single-step reasoning guarantees the fidelity and generation quality, especially suitable for real remote sensing data with complex imaging degradation causes and large differences in degradation degree between different images, and can adaptively perform super-resolution reconstruction for the degradation of each image.

Description

Technical Field

[0001] This invention belongs to the field of image processing and computer vision technology, and particularly relates to super-resolution reconstruction of remote sensing images. Background Technology

[0002] Super-resolution reconstruction (SR) of remote sensing images is a technique that recovers high-resolution (HR) images from low-resolution (LR) images without altering the imaging hardware. Its core function is to enhance the spatial resolution and texture details of images, thereby improving the identification and interpretation of ground features. It is widely used in fields such as land resource monitoring and environmental assessment.

[0003] Currently, diffusion model-based methods, such as Stable Diffusion and FLUX.1-dev, leverage their powerful generative priors obtained through pre-training on large-scale data to achieve high-quality reconstruction using low-quality images as a condition, through an iterative denoising process. This significantly outperforms traditional interpolation and pure Generative Adversarial Network (GAN) methods in terms of perceptual quality. Even when adversarial loss is introduced during fine-tuning, it participates with only a small weight, enhancing the realism of high-frequency textures while effectively mitigating the training instability and artifact problems common in pure GAN methods.

[0004] However, several bottlenecks still need to be overcome in practical remote sensing engineering applications. To reduce the computational cost of multi-step sampling, recent studies tend to distill pre-trained diffusion models into single-step models. Typical paradigms include OSEDiff introducing variational fractional distillation, TSD-SR improving detail recovery, and OMGSR observing that the noise latent distribution at intermediate time steps is closer to the latent distribution of low-quality images, injecting low-quality latent representations at fixed intermediate time steps.

[0005] The above methods share the following common problems: First, multi-step diffusion methods require dozens of iterative sampling steps, resulting in slow inference speed and high computational overhead, making it difficult to meet the engineering and efficient processing requirements of large-format remote sensing images. Secondly, while single-step diffusion methods improve efficiency, they use the same fixed starting time step for all inputs. However, the degradation levels of images vary significantly between different imaging systems and even within the same system under different regions and conditions. Diffusion models exhibit varying degrees of generative priors at different time steps. This "one-size-fits-all" strategy cannot adapt to the inherent degradation differences of each image—it easily introduces over-generation and false textures into mildly degraded images, while it fails to adequately repair severely degraded images, making it difficult to achieve a balance between fidelity and generative capability. Existing work, TADSR, has recognized the limitations of fixed time steps and introduced a time-aware VAE encoder. However, its time step variation comes from random sampling during the training phase and is unrelated to the actual degradation level of the input images. Therefore, it cannot determine a time step that matches the degradation level of each image during the inference phase, nor is it specifically designed for scenarios with large degradation differences in remote sensing images.

[0006] Third, the degradation estimation network upon which the adaptive time step relies lacks accuracy and stability. Existing degradation estimation networks (such as DEResNet in MM-RealSR) employ independent estimation by degradation type branches, resulting in parameter redundancy. Their sigmoid output layer is prone to convergence efficiency issues during training with ranking loss due to the gentle gradient in the saturation region. Furthermore, deeper spatial downsampling weakens the mid-frequency texture information used to distinguish the severity of degradation and lacks global context modeling capabilities. These shortcomings lead to insufficient accuracy and robustness in estimating degradation levels in scenarios with complex and highly varied degradation causes in remote sensing images, making it difficult to provide reliable support for accurate mapping of the adaptive time step. Summary of the Invention

[0007] To address the issues of reconstruction efficiency and engineering applicability in existing technologies, the fidelity-generation imbalance caused by fixed time steps, and the insufficient accuracy of degradation estimation networks, this invention proposes a degradation-aware time step-guided single-step diffusion remote sensing image super-resolution reconstruction method. This method systematically improves the performance and practicality of remote sensing image super-resolution reconstruction, and includes the following steps: S1. Construct a degradation estimation network, which consists of a first-layer convolution, a shared feature extraction backbone, a channel projection neck, a self-attention bottleneck, a degradation estimation head, and a corresponding regression output layer. It takes the low-resolution remote sensing image to be reconstructed as input, estimates and outputs several scalar scores that characterize the degree of degradation of the image, and uses them to estimate the degree of degradation of the input remote sensing image on an image-by-image basis. S2. The scalar score output by the degradation estimation network in S1 is converted into the starting time step injected by the single-step diffusion reconstruction. An adaptive mapping from the degree of degradation to the diffusion time step is performed to determine the single-step diffusion starting time step that matches the degree of degradation for each image. S3. Using the adaptive time step determined by S2 as the starting point for single-step diffusion reconstruction, the remote sensing image reconstruction from low resolution to high resolution is completed.

[0008] The shared feature extraction backbone consists of several cascaded downsampling stages, which are shared across all degenerate dimensions. Each downsampling stage contains several residual blocks for local feature extraction, as well as a convolution to change the spatial resolution or number of channels, and outputs a feature map. The self-attention bottleneck applies one or more self-attention modules to the preserved feature map, introducing global context modeling for the entire image while preserving mid-frequency texture information. The degradation estimation head consists of a residual block, followed by global average pooling to compress the feature map into a feature vector, and then outputs the scalar score of the degradation dimension through a regression output layer consisting of a fully connected layer, ReLU, a fully connected layer, and bounded activation.

[0009] The adaptive mapping described in S2 is specifically as follows: First, the scalar scores output in S1 are checked for validity and truncated to between 0 and 1. Then, they are weighted and summed according to preset weights to obtain the comprehensive degradation metric d. Linear rescaling is applied to the overall degradation metric d, and (d-d_min) / (d_max-d_min) is calculated and truncated to between 0 and 1 to obtain the calibrated degradation metric. A nonlinear transformation with a power exponent less than 1 is applied to the degradation metric; The transformed degradation metric is linearly mapped to the preset time step index interval [t_min, t_max] to obtain the time step index t_index; this index is rounded and truncated to the range [t_min, t_max], and then the corresponding sigma (the initial noise level of single-step diffusion reconstruction) is obtained by looking up the index according to the length of the time step sequence.

[0010] The single-step diffusion reconstruction in S3 is specifically as follows: Using a pre-trained diffusion model as a base, low-quality images are encoded into low-quality latent representations. The sample-by-sample sigma obtained by S2 is used as the initial noise level injected into the trainable diffusion transformer. Single-step prediction is performed to obtain flow field / velocity predictions. Based on the initial noise level of each sample, the low-quality latent representation is predicted by -sigma × , and the image is restored to the zero time step one sample at a time step to obtain the reconstructed latent representation. The image is then decoded by a frozen VAE decoder to obtain the reconstructed image.

[0011] Technical effects: To address the imbalance between fidelity and generation at a fixed time step, this invention establishes an adaptive mapping mechanism between the degree of image degradation and the starting time step. Multiple degradation scores are weighted and fused, and after rescaling and nonlinear transformation, they are mapped to a preset time step range to determine the corresponding starting time step. This achieves an adaptive adjustment where the starting time step is larger for more severe degradation and smaller for milder degradation, thus achieving the optimal balance between fidelity and generation effect.

[0012] To address the technical challenges of high computational overhead and slow speed in multi-step iterative inference using iterative sampling, which hinders the batch processing of wide-format remote sensing images, this invention achieves single-step inference reconstruction while maintaining high-quality reconstruction. Using the aforementioned adaptive time step as the starting point for single-step diffusion reconstruction, it completes low-resolution to high-resolution reconstruction under single-step inference, thereby avoiding false textures caused by over-generation in mildly degraded images and achieving sufficient texture restoration in severely degraded images. Compared to multi-step diffusion, this significantly reduces computational overhead and time consumption. Combined with latent space overlapping blocks and block-based VAEs, it processes large-format remote sensing images exceeding the base resolution with limited video memory, ensuring both efficiency and fidelity in the generated images.

[0013] To address the insufficient accuracy of degradation estimation networks, this invention constructs a SharedDEResNet degradation estimation network. This network employs a shared backbone combined with two parallel degradation estimation heads to extract shared features for degradation types (blurred, noisy). Furthermore, by introducing downsampled residual connections, non-saturating bounded activations, and terminal self-attention modules, the number of parameters is reduced while ensuring training stability and preserving mid-frequency information that distinguishes the severity of degradation, thus outputting a more accurate measure of degradation degree. This provides a solid basis for dynamic time step selection.

[0014] This invention can be applied to resolution enhancement and texture restoration of remote sensing images acquired by satellite, aerial, and other Earth observation platforms. It provides high-quality image data support for fields with high image clarity requirements, such as land resource surveys, urban planning, environmental and disaster monitoring, agricultural and forestry surveys, and land cover classification and target identification. It is particularly suitable for real remote sensing data with complex causes of image degradation and significant differences in degradation levels between different images, enabling adaptive super-resolution reconstruction tailored to the specific degradation characteristics of each image. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the overall process of this invention.

[0016] Figure 2 This is a block diagram of the internal structure of the degradation estimation network in this embodiment.

[0017] Figure 3 Comparison of single-step super-resolution reconstruction results of low-resolution remote sensing images with different degradation levels at different time steps.

[0018] Figure 4 This is a scatter plot comparing the estimation results of this embodiment with existing degradation estimation networks under different degradation levels. Detailed Implementation

[0019] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This embodiment provides a degradation-sensing time-step guided method for super-resolution reconstruction of single-step diffusion remote sensing images, such as... Figure 1 As shown, the low-resolution remote sensing image is first read. The input low-resolution remote sensing image is then subjected to a pixel-unshuffle operation, which rearranges and folds spatially adjacent pixels to the channel dimension. For example, the H×W×3 input is rearranged into (H / 2)×(W / 2)×12 with a factor of 2. This halves the spatial size while retaining all high-frequency information without loss, avoiding the loss of high-frequency information caused by conventional stride convolution or pooling downsampling. Correspondingly, the number of input channels of the degradation estimation network is set to be consistent with the number of channels after the unshuffle.

[0021] For the high-resolution remote sensing images used for training, a high-order degradation model is used to synthesize the corresponding low-resolution images, forming low-quality (LQ) - high-quality (HQ) image pairs, with pixel values ​​of both normalized to [-1, 1]. In a preferred embodiment, the Real-ESRGAN degradation pipeline is used. Under single-precision (fp32) conditions, before any image enlargement or padding, the low-quality image is first restored from [-1, 1] to [0, 1] and resampled to a preset size, such as 512×512.

[0022] For large images exceeding the preset block size (i.e., 512×512), they are cropped into several image blocks before pixel inverse shuffling. For images not exceeding the preset block size, they are first filled to the block size before a single estimation is performed.

[0023] The preprocessed image patch is fed into a degradation estimation network to obtain a degradation score for the image patch. Further, the degradation estimation network (SharedDEResNet) takes the low-resolution remote sensing image to be reconstructed as input, estimates and outputs several scalar scores characterizing the degree of degradation of the image. These scalar scores range from 0 to 1, corresponding to the severity of degradation in different degradation dimensions; a larger value indicates a more severe degradation in that dimension. In a preferred embodiment, the degradation dimension includes two dimensions: blur and noise. The degradation estimation network outputs the blur and noise scores of the corresponding image blocks, and averages the scores of all image blocks to obtain the blur and noise scores of the entire image.

[0024] Furthermore, the degradation estimation network consists of a first convolutional layer, a shared feature extraction backbone, a channel projection neck, a self-attention bottleneck, a degradation estimation head, and a corresponding regression output layer, with the following data flow: The first convolutional layer maps the preprocessed low-resolution image feature map to the basic channel dimension of the backbone to obtain initial features. Preferably, the basic channel dimension is 64 channels.

[0025] The shared feature extraction backbone consists of several cascaded downsample stages, sharing features across all degenerate dimensions, meaning all degenerate dimensions reuse features extracted from the same backbone. Each downsample stage contains several residual blocks for local feature extraction, as well as a convolution to change the spatial resolution or number of channels; furthermore, each downsample stage has a residual shortcut connected in parallel to its main path.

[0026] In a preferred embodiment, the shared feature extraction backbone is configured with four stages. The number of channels in each stage is 64, 64, 64, and 128, the number of residual blocks is 1, 1, 2, and 2, and the downsampling factor is 2, 2, 2, and 1, respectively. That is, the shallow stage is thinner (single residual block), and the deep stage is thicker (two residual blocks). The total spatial downsampling factor of the shared feature extraction backbone is 8 times. Taking a 256×256 preprocessed input as an example, a 32×32 feature map is obtained after passing through the shared feature extraction backbone.

[0027] The channel projection neck involves projecting the feature map output from the shared feature extraction backbone through a 1×1 convolution to the channel dimension required by the self-attention bottleneck, thereby controlling the computational cost of the self-attention module. In a preferred embodiment, the projection is onto 64 channels.

[0028] The self-attention bottleneck applies one or more self-attention modules to the preserved feature map. In a preferred embodiment, this is a 32×32 module, which flattens out to correspond to 1024 tokens. Each self-attention module consists of a residual block, a multi-head self-attention layer, and a feedforward network: the residual block further extracts local features; the multi-head self-attention layer (4 attention heads in a preferred embodiment) flattens the feature map into a sequence, calculates the query, key, and value after layer normalization, and performs scaled dot product attention, injecting it into the global context in a residual manner; the feedforward network then performs nonlinear refinement of each channel in a residual manner. Thus, the self-attention bottleneck introduces global context modeling for the entire image while preserving mid-frequency texture information.

[0029] The features of the degradation estimation head and regression output, and the self-attention bottleneck output, are processed separately by lightweight estimation heads corresponding to each degradation dimension. In a preferred embodiment, each degradation estimation head consists of a residual block, followed by global average pooling to compress the feature map into a feature vector, and then outputs the degradation score of that degradation dimension through a regression output layer consisting of a fully connected layer—ReLU—fully connected layer—bounded activation; the bounded activation in the preferred embodiment uses an unsaturated bounded activation function to constrain the score between 0 and 1.

[0030] In a preferred embodiment, the first convolutional layer, the shared feature extraction backbone, and each degradation estimation head are initialized with a unit scale, while the regression output layer is initialized with a smaller scale (e.g., 0.1 times), so that the degradation scores are around 0.5 and the activations are in the linear region in the initial training stage, thereby ensuring that the ranking loss based on metric learning can effectively backpropagate gradients in the early training stage.

[0031] The degradation estimation network can be pre-trained independently and used as a predictor of degradation degree in the method of the present invention by freezing parameters; its training adopts existing unsupervised metric learning methods (such as metric learning based on ranking loss), and the training method itself does not constitute a limitation of the present invention.

[0032] The differences and improvements of the degradation estimation network in this embodiment compared to the existing degradation estimation network (DEResNet) are mainly reflected in the following four aspects: 1. Shared Feature Extraction Backbone and Multi-Head Estimation Structure. Existing DEResNet sets up a complete feature extraction branch with independent parameters for each degradation type, essentially replicating the entire backbone according to the number of degradation dimensions. This embodiment uses only one shared feature extraction backbone, sharing feature extraction across all degradation dimensions, and connecting a lightweight estimation head to each degradation dimension after the backbone. Since the shallow and mid-layer features relied upon by different degradation dimensions highly overlap, the shared backbone eliminates redundant branches; even with the additional introduction of a self-attention module, the overall parameter count is significantly reduced compared to the two-branch DEResNet, with a measured reduction of approximately 40%.

[0033] 2. Bounded and Unsaturated Output Activation (SoftClamp). Existing DEResNet uses a sigmoid activation at the output, whose gradient decays rapidly exponentially when the input magnitude is large, entering the saturation region. This causes the gradient to vanish and converge slowly during training for the ranking loss based on metric learning. This embodiment uses a bounded activation function with a gradually varying gradient at the output, constraining the output between 0 and 1, while the gradient decays polynomially. This avoids gradient vanishing caused by activation saturation, ensuring the ranking loss remains effective throughout the training process. In a preferred embodiment, the gradient decays inversely proportional to the square of the input magnitude.

[0034] 3. Residual shortcut at downsampling transitions. In existing DEResNet, residual connections exist only within each residual block (ResBlock). The downsampling layer responsible for reducing spatial resolution is a regular strided convolution, directly connected in the network. There is no shortcut across this convolution at the downsampling transition, and the gradient must pass through the strided convolution to decay at this point. In this embodiment, each downsampling stage is constructed as a main path of "several residual blocks and downsampling convolutions". A residual shortcut consisting of average pooling and 1×1 convolution is connected in parallel outside this main path, allowing the gradient to cross the downsampling transition and be directly backpropagated in a manner close to unity gain, thereby improving the stability of deep network training.

[0035] 4. Self-attention bottleneck replaces direct global pooling. Existing DEResNet directly applies global average pooling after the backbone output, compressing the entire feature map into a single feature vector before regressing the degradation score. This lacks modeling of the global spatial context, and global averaging also erases mid-frequency texture information used to distinguish the severity of degradation. This embodiment retains a feature map of a certain size at the end of the backbone and introduces a self-attention module consisting of residual blocks, multi-head self-attention, and a feedforward network on this feature map. It first models the global context and retains mid-frequency texture before performing pooling and regression, thereby improving the accuracy and stability of degradation estimation.

[0036] An adaptive mapping from degradation level to diffusion time step (deg_to_sigma) determines the starting time step for each sample in the batch, resulting in a sample-by-sample sigma vector. The sigma represents the initial noise level of the single-step diffusion reconstruction, characterizing the starting position of the input low-quality image on the diffusion trajectory. This sigma is obtained by looking up the time step index in the flow matching schedule and is a continuous value. Its magnitude directly determines the modification magnitude of the input image by the single-step reconstruction.

[0037] The specific steps of the adaptive mapping are as follows: 1. Weighted Fusion. First, the degradation scores output by the degradation estimation network are checked for validity and truncated to between 0 and 1. Then, they are weighted and summed according to preset weights to obtain the comprehensive degradation metric d. In a preferred embodiment, the comprehensive degradation metric d is obtained by weighting the fuzziness score and the noise score according to their respective weights (e.g., each is taken as 0.5).

[0038] 2. Interval Calibration. Considering that the comprehensive degradation metric d does not cover the entire range of values ​​in the actual remote sensing data, but is distributed in a relatively low effective sub-interval, d is linearly rescaled using a preset lower bound d_min and upper bound d_max. That is, (d - d_min) / (d_max - d_min) is calculated and truncated to between 0 and 1 to obtain the calibrated degradation metric d_norm. This stretches the effective interval [d_min, d_max] (in the preferred embodiment, [0.04, 0.4]) of d to the full scale [0,1], so that subsequent mapping can make full use of the entire time step interval and avoid d only occupying a small segment of the time step interval due to its narrow distribution. Among them, d_min and d_max are pre-calibrated according to the actual distribution of the output scores of the degradation estimation network on the target data. Extreme samples distributed below the lower bound or above the upper bound are truncated to correspond to the minimum and maximum time steps, respectively.

[0039] 3. Nonlinear Transformation. A nonlinear (gamma) transformation with an exponent less than 1 is applied to the calibrated degradation metric d_norm, i.e., d_norm is raised to the power of gamma. In a preferred embodiment, gamma is set to 0.33. Since this power function has a larger slope in the low-value region, a larger range of variation can be allocated to the time step in the less degraded region, thereby providing finer time step resolution and enabling more thorough differentiation between slightly degraded images.

[0040] 4. Map to time step interval and look up sigma. Linearly map the transformed degradation metric to the preset time step index interval [t_min, t_max] to obtain the time step index t_index (in the preferred embodiment, t_min and t_max are 100 and 300 respectively); then round the index and truncate it to the range [t_min, t_max], and then look up the corresponding sigma value from the index according to the time step sequence length of the diffusion model used, which is used as the starting noise level for the single-step diffusion reconstruction of the image.

[0041] In a preferred embodiment, the value at index n-1 - t_index is taken according to the time step sequence length n. The mapping as a whole satisfies a monotonic correspondence: the more severe the degradation, the larger the determined starting time step, thus introducing stronger generative repair; the less severe the degradation, the smaller the determined starting time step, thus being more inclined to preserve fidelity.

[0042] The mapping is performed sample-by-sample, meaning that different samples within a batch determine their respective starting time steps based on their degree of degradation. A small random jitter is superimposed on the time step index to enhance robustness to degradation estimation errors and avoid excessive concentration of time step values. In a preferred embodiment, the jitter amplitude is ±5 indices.

[0043] Single-step diffusion reconstruction guided by degradation estimation and adaptive mapping: Single-step diffusion reconstruction is based on a pre-trained diffusion model. In the preferred embodiment, it employs FLUX.1-dev based on flow matching, which includes a variational autoencoder (VAE) and a diffusion transformer (Transformer). The VAE encoder and diffusion transformer are fine-tuned using low-rank adaptation (LoRA) (rank 64 in the preferred embodiment), and the VAE decoder remains frozen throughout. The degradation estimation network is used as a degradation predictor in a frozen, single-precision (fp32) manner. Low-quality images are encoded into low-quality latent representations by the trainable VAE encoder (LoRA), serving as the input state for single-step diffusion prediction.

[0044] The time step sequence of the diffusion model is generated by a preset flow matching schedule. In the preferred embodiment, the length is 1000, which is denoted as the time step tensor and used for deg_to_sigma to look up the sigma value.

[0045] The single-step diffusion process specifically involves: inputting the low-quality latent representation into a trainable diffusion transformer, and injecting the sample-by-sample sigma obtained in S2 into the diffusion transformer to perform single-step prediction to obtain the flow field / velocity prediction; then, according to the sigma of each sample, restoring it step by step to the zero time step using "low-quality latent representation - sigma × prediction" to obtain the reconstructed latent representation, which is then decoded by a frozen VAE decoder to obtain the reconstructed image.

[0046] Furthermore, for large images exceeding the resolution of the video memory or base, they are divided into several overlapping image blocks in the latent space for prediction. Each image block shares the same sigma, and the overlapping areas are weighted and fused together using Gaussian weights. The encoding and decoding of VAEs also adopt a block-based approach accordingly.

[0047] In one possible implementation, loss and optimization are performed. Multiple losses are jointly supervised: including structure-aware (DISTS) perception loss based on DINOv3-ConvNeXt, pixel-level L1 loss, and generative adversarial loss provided by the discriminator. In a preferred embodiment, the discriminator is a DINOv3-ConvNeXt discriminator, and the losses are weighted according to preset weights. The generation side updates the LoRA parameters of the VAE encoder and diffusion transform; the discriminator side alternately updates the discriminator parameters using reconstructed images as negative samples and high-quality images as positive samples.

[0048] The specific configurations for each step in this embodiment are as follows: This embodiment uses FLUX.1-dev as a base, applying a rank-64 LoRA to its VAE encoder and diffusion transformer, and freezing the VAE decoder. The degradation estimation network SharedDEResNet is configured with 12 input channels, estimating two dimensions: ambiguity and noise, and using unsaturated bounded activation (SoftClamp) at the output. It is pre-trained using metric learning, frozen, and loaded for use in single-precision mode.

[0049] The training data for the pre-trained diffusion model consists of low-quality to high-quality image pairs synthesized online from high-resolution remote sensing images via the Real-ESRGAN second-order degradation pipeline; the training resolution is 1024×1024, the batch size is 2, AdamW is used, and the learning rate is 5×10. -5 Constant scheduling, 500 warm-up steps, weight decay of 1×10 -4 The training was performed with bf16 mixed precision and gradient clipping norm of 1.0, for a total of approximately 50,000 steps. The loss weights were set as DINOv3-ConvNeXt (large)DISTS loss 2, L1 loss 5, and adversarial loss 0.01.

[0050] The parameters for the adaptive time-step mapping are set as follows: blur and noise weights are each 0.5, the effective degradation interval d∈[0.04,0.4], gamma=0.33, and the time-step index interval [t_min, t_max]=[100, 300]. Random jitter of ±5 indices is superimposed during training, and no jitter is applied during inference. Degradation estimation is performed after the image is restored from [-1, 1] to [0, 1], resampled to 512×512, and subjected to pixel inverse shuffling with a factor of 2.

[0051] During inference, the input remote sensing image is first divided into blocks of 512 for degradation estimation before magnification and padding, and the scores of each block are averaged. The starting time step of the entire image is obtained by adaptive mapping. Then, it is magnified and padded to the size that meets the constraints (base padding 128, minimum side length 1024, aligned to multiples of 16) by copying the boundary. The latent space is divided into overlapping blocks, each block shares the same time step and is fused with Gaussian weights. VAE encoding and decoding adopt a block-based approach. After a single-step prediction, it is restored and decoded in one step by "latent representation - sigma × prediction". After the padding boundary is removed, it is saved as a TIF file.

[0052] The following section illustrates the technical effectiveness of this embodiment by performing single-step super-resolution reconstruction on low-resolution remote sensing images with different degrees of degradation using three strategies: a fixed large time step, a fixed small time step, and the adaptive time step of this embodiment. Figure 3 As shown, for images with mild degradation, using a fixed, relatively large time step (e.g., t=300) introduces over-generation due to the excessively high initial step size, resulting in false textures and "scratching" in the reconstruction results, and even altering the original orientation of linear features, thus destroying the authenticity of the features. For images with severe degradation, using a fixed, relatively small time step (e.g., t=100) results in insufficient restoration due to insufficient generation intensity, and the reconstruction results remain blurry with limited detail recovery. This invention adaptively determines the starting time step based on the degree of degradation of each image—a smaller time step is used for mild degradation to prioritize fidelity and avoid over-generation, while a larger time step is used for severe degradation to enhance generation and fully restore texture. Therefore, faithful and clear reconstruction results can be obtained in both mild and severe degradation cases, verifying the advantage of adaptive time steps over fixed time steps.

[0053] like Figure 4 As shown, to verify the improvement effect of the proposed degradation estimation network SharedDEResNet compared with the existing degradation estimation network DEResNet, the inter-class separability of the degradation metrics output by the two networks is compared and evaluated under the same settings.

[0054] The results show that the improved degradation estimation module significantly outperforms the original module in inter-class separability: the average inter-class Euclidean distance is improved by more than 50%, and the Fisher discrimination ratio is improved by more than 60%; among them, the intra-class joint standard deviation of the mildly degraded mild class decreases by more than 50%, indicating a significant improvement in the estimation consistency for low-degraded images. These results demonstrate that the degradation estimation network of this invention can more stably and accurately distinguish remote sensing images with different degrees of degradation, especially improving the estimation consistency for mildly degraded images, thus providing a more reliable degradation metric basis for the adaptive mapping from degradation degree to time step.

[0055] In summary, all contents not described in detail in this specification are existing technologies known to those skilled in the art. Furthermore, for those skilled in the art, there will be changes in specific implementation methods and application scope based on the ideas of this invention. Therefore, the contents of this specification should not be construed as limiting the invention.

Claims

1. A method for super-resolution reconstruction of single-step diffusion remote sensing images guided by degradation-aware time steps, characterized in that, Includes the following steps: S1. Construct a degradation estimation network, which consists of a first-layer convolution, a shared feature extraction backbone, a channel projection neck, a self-attention bottleneck, a degradation estimation head, and a corresponding regression output layer. It takes the low-resolution remote sensing image to be reconstructed as input, estimates and outputs several scalar scores that characterize the degree of degradation of the image, and uses them to estimate the degree of degradation of the input remote sensing image on an image-by-image basis. S2. The scalar score output by the degradation estimation network in S1 is converted into the starting time step injected by the single-step diffusion reconstruction. An adaptive mapping from the degree of degradation to the diffusion time step is performed to determine the single-step diffusion starting time step that matches the degree of degradation for each image. S3. Using the adaptive time step determined by S2 as the starting point for single-step diffusion reconstruction, the remote sensing image reconstruction from low resolution to high resolution is completed.

2. The degradation-sensing time-step guided single-step diffusion remote sensing image super-resolution reconstruction method according to claim 1, characterized in that, The shared feature extraction backbone consists of several cascaded downsampling stages, which are shared across all degenerate dimensions. Each downsampling stage contains several residual blocks for local feature extraction, as well as a convolution to change the spatial resolution or number of channels, and outputs a feature map. The self-attention bottleneck applies one or more self-attention modules to the preserved feature map, introducing global context modeling for the entire image while preserving mid-frequency texture information. The degradation estimation head consists of a residual block, followed by global average pooling to compress the feature map into a feature vector, and then outputs the scalar score of the degradation dimension through a regression output layer consisting of a fully connected layer, ReLU, a fully connected layer, and bounded activation.

3. The degradation-sensing time-step guided single-step diffusion remote sensing image super-resolution reconstruction method according to claim 1, characterized in that, The adaptive mapping described in S2 is specifically as follows: First, the scalar scores output in S1 are checked for validity and truncated to between 0 and 1. Then, they are weighted and summed according to preset weights to obtain the comprehensive degradation metric d. Linear rescaling is applied to the overall degradation metric d, and (d-d_min) / (d_max-d_min) is calculated and truncated to between 0 and 1 to obtain the calibrated degradation metric. A nonlinear transformation with a power exponent less than 1 is applied to the degradation metric; The transformed degradation metric is linearly mapped to the preset time step index interval [t_min, t_max] to obtain the time step index t_index; this index is rounded and truncated to the range [t_min, t_max], and then the corresponding sigma is obtained by looking up the index according to the length of the time step sequence. sigma is the starting noise level of single-step diffusion reconstruction.

4. The degradation-sensing time-step guided single-step diffusion remote sensing image super-resolution reconstruction method according to claim 3, characterized in that, The single-step diffusion reconstruction in S3 is specifically as follows: Using a pre-trained diffusion model as a base, low-quality images are encoded into low-quality latent representations. The sample-by-sample sigma obtained by S2 is used as the initial noise level injected into the trainable diffusion transformer. Single-step prediction is performed to obtain flow field / velocity predictions. Based on the initial noise level of each sample, the low-quality latent representation is predicted by -sigma × , and the image is restored to the zero time step one sample at a time step to obtain the reconstructed latent representation. The image is then decoded by a frozen VAE decoder to obtain the reconstructed image.

5. The degradation-sensing time-step guided single-step diffusion remote sensing image super-resolution reconstruction method according to claim 1, characterized in that, The degradation dimension includes two dimensions: blur and noise. The degradation estimation network outputs the blur and noise scores of the corresponding image blocks, and averages the scores of all image blocks to obtain the blur and noise scores of the entire image.

6. The degradation-sensing time-step guided single-step diffusion remote sensing image super-resolution reconstruction method according to claim 2, characterized in that, The shared feature extraction backbone is set up in four stages, with the number of channels in each stage being 64, 64, 64, and 128, the number of residual blocks being 1, 1, 2, and 2, and the downsampling factors being 2, 2, 2, and 1, respectively. The total spatial downsampling factor is 8 times.

7. The degradation-sensing time-step guided single-step diffusion remote sensing image super-resolution reconstruction method according to claim 4, characterized in that, The single-step diffusion reconstruction uses FLUX.1-dev based on flow matching as the pre-trained diffusion model, which includes a variational autoencoder (VAE) and a diffusion transformer (Transformer). The encoder (VAE) and the diffusion transformer are fine-tuned using a low-rank adaptation method, and the VAE decoder is kept frozen throughout. The degradation estimation network is used as a degradation degree predictor with frozen and single precision.

8. The degradation-sensing time-step guided single-step diffusion remote sensing image super-resolution reconstruction method according to claim 7, characterized in that, The training data for the pre-trained diffusion model is generated by online synthesis of low-quality and high-quality image pairs from high-resolution remote sensing images via the Real-ESRGAN second-order degradation pipeline; Training resolution 1024×1024, batch size 2, using AdamW, learning rate 5×10 -5 Constant scheduling, 500 warm-up steps, weight decay of 1×10 -4 bf16 blend precision, gradient clipping norm 1.0.

Images