A near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic geometry cooperation

Abstract

The application discloses a kind of based on multiscale wavelet feature modulation and semantic geometry cooperation near-infrared remote sensing image super-resolution method, comprising: obtaining low-resolution near-infrared remote sensing image;According to low-resolution near-infrared remote sensing image, obtain the multi-channel feature map after pre-processing;According to multi-channel feature map, obtain multiscale wavelet transform modulation fusion feature;According to multiscale wavelet transform modulation fusion feature, obtain deep state representation feature, wherein, deep state representation feature is obtained by multiple iteration processing to semantic geometry joint perception feature, and semantic geometry joint perception feature is obtained based on the cooperative processing of semantic scanning branch and geometric scanning branch;According to the fusion result of deep state representation feature and multiscale wavelet transform modulation fusion feature, obtain high-resolution near-infrared remote sensing image.The present application solves the problems of insufficient detail texture recovery and insufficient frequency band information utilization in existing near-infrared remote sensing image super-resolution reconstruction.

Description

Technical Field

[0001] This invention belongs to the field of agricultural remote sensing image processing technology, and particularly relates to a near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy. Background Technology

[0002] With the rapid development of agricultural remote sensing imaging technology, agricultural near-infrared remote sensing images, due to their ability to reflect differences in the reflectance of ground objects in the near-infrared band and their strong penetration and separability of ground objects, have become an important data carrier in the field of remote sensing. They are widely used in tasks such as crop growth monitoring, vegetation cover analysis, land cover classification, change detection, disaster assessment, target identification, and fine mapping. However, in actual imaging and transmission, agricultural near-infrared remote sensing images are often affected by factors such as sensor spatial resolution limitations, long imaging distances, platform jitter, atmospheric scattering and noise interference, and compression coding. This results in problems such as insufficient spatial resolution, loss of detailed textures, and blurred edge structures in the acquired images. These problems not only reduce the visual quality of the images but also further affect the accuracy of quantitative analysis based on image data, the reliability of key ground object information extraction, and the effectiveness of final agricultural monitoring and decision-making applications.

[0003] Existing super-resolution methods for agricultural remote sensing images mainly fall into two categories: traditional methods and deep learning-based methods. Traditional super-resolution methods typically include interpolation-based reconstruction methods (such as bilinear interpolation and bicubic interpolation), methods based on sparse representation or dictionary learning, and methods based on reconstruction constraint optimization. These methods are simple to implement and have low computational costs, but they often rely on fixed mathematical models or prior assumptions, making it difficult to fully recover true high-frequency detail information. They are also prone to causing overly smoothed reconstructed images, texture distortion, or artifacts, making it difficult to meet the high requirements of detail clarity and structure preservation in near-infrared remote sensing images.

[0004] In recent years, the development of deep learning technology has provided new solutions for image super-resolution reconstruction. By constructing end-to-end neural networks to learn the mapping relationship from low-resolution to high-resolution images, the objective evaluation indicators and visual effects of reconstructed images can be improved to a certain extent. However, near-infrared remote sensing images are characterized by complex texture structures, significant scale variations, dense target distribution, and strong background interference, and existing deep learning super-resolution methods still have shortcomings in practical applications. On the one hand, many methods mainly rely on convolutional neural networks to extract local features. Although this can enhance local textures, its ability to model long-range dependencies is limited, resulting in defects in maintaining the structural consistency of the reconstruction results over a large area. On the other hand, some methods lack explicit modeling of frequency band information, making it difficult to effectively separate and enhance high-frequency detail features, easily leading to insufficient detail recovery or blurred edges. In addition, the texture and structure differences between different regions of near-infrared remote sensing images are significant. Fixed information propagation and fusion strategies are difficult to adapt to the feature requirements of different spatial locations, easily leading to artifacts, edge breaks, or unstable details in local areas. At the same time, using a single-path feature modeling approach can easily ignore the constraint of global semantic information on the reconstruction results, thus affecting the overall structural rationality and semantic consistency of the reconstructed image.

[0005] Therefore, there is an urgent need to design a near-infrared remote sensing image super-resolution method that can simultaneously take into account multi-scale detail enhancement, frequency band feature modulation, and semantic-geometric adaptive fusion capabilities, so as to further improve the detail clarity, structural consistency, and robustness of the reconstructed image and meet the needs of practical agricultural remote sensing applications. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy. This method solves the problems in existing near-infrared remote sensing image super-resolution reconstruction, such as insufficient restoration of detail and texture, inadequate utilization of frequency band information, limited long-range dependency modeling capabilities, and non-adaptive fusion of features from different spatial locations. Frequency band feature enhancement and modulation are achieved by combining multi-scale feature extraction with wavelet transform. A deep feature extraction network composed of multiple cascaded residual semantic-geometric co-engineering modules (RSMG) is constructed. The RSMG includes multiple semantic-geometric co-engineering blocks RSSD. The semantic-spatial coordinate dual-branch fusion module DSSM in RSSD includes a semantic scanning branch and a geometric scanning branch set in parallel. The semantic scanning branch performs selective aggregation modeling of global semantic features based on ASSM, and the geometric scanning branch performs multi-directional scanning modeling of two-dimensional features based on the Mamba state space model. At the same time, coordinate-aware gating is introduced to generate pixel-level directional weights, and the semantic scanning results and geometric scanning results are adaptively fused through a gating fusion mechanism to achieve synergistic enhancement of high-frequency details and global structural information of near-infrared remote sensing images. This ensures the recovery of high-frequency details while maintaining the rationality of global structure and improving the performance of super-resolution reconstruction.

[0007] To achieve the above objectives, this invention provides a near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy, comprising: Acquire low-resolution near-infrared remote sensing images; Based on the low-resolution near-infrared remote sensing image, obtain the preprocessed multi-channel feature map; Based on the multi-channel feature map, obtain the multi-scale wavelet transform modulation fusion features; Based on the multi-scale wavelet transform modulation fusion features, deep state representation features are obtained, wherein the deep state representation features are obtained by iteratively processing the semantic-geometric joint perception features, and the semantic-geometric joint perception features are obtained based on the collaborative processing of the semantic scanning branch and the geometric scanning branch. Based on the fusion result of the deep state characterization features and the multi-scale wavelet transform modulation fusion features, a high-resolution near-infrared remote sensing image is obtained.

[0008] Optionally, obtaining the preprocessed multi-channel feature map based on the low-resolution near-infrared remote sensing image includes: Based on the single-channel image in the low-resolution near-infrared remote sensing image, a three-channel image is obtained, wherein each channel of the three-channel image is a near-infrared channel; Based on the three-channel image, an enhanced image is obtained after random horizontal flipping and rotation; based on the enhanced image, a multi-channel feature map is obtained after dimension adjustment and normalization.

[0009] Optionally, obtaining multi-scale wavelet transform modulation fusion features based on the multi-channel feature map includes: Based on the multi-channel feature map, first-scale features, second-scale features, and third-scale features are obtained, wherein the first-scale features, second-scale features, and third-scale features are extracted through different convolutional branches; Based on the first scale feature, a first wave domain feature map is obtained, wherein the first wave domain feature map is obtained by performing a discrete wavelet transform on the first scale feature. Based on the first wave domain feature map, obtain the spatial gating signal; The second scale feature and the third scale feature are dynamically modulated according to the spatial gating signal to obtain the modulated second scale feature and the modulated third scale feature; Based on the modulated second-scale features and the modulated third-scale features, an adaptive weighted fusion feature is obtained; The multi-scale wavelet transform modulation fusion feature is obtained by splicing the first scale feature, the second scale feature, the third scale feature and the adaptive weighted fusion feature.

[0010] Optionally, obtaining the spatially gated signal based on the first wave domain feature map includes: Based on the channel compression result of the first wave domain feature map, obtain the activated features; Based on the channel compression result of the activated feature, obtain the initial gate value after scaling the numerical range; The spatial gating signal is obtained by adding the initial gating value to the preset offset.

[0011] Optionally, obtaining the semantic-geometric joint sensing features based on the multi-scale wavelet transform modulation fusion features includes: Based on the multi-scale wavelet transform modulation fusion features, a normalized feature map is obtained. Based on the normalized feature map, geometric scan features are obtained, wherein the geometric scan features are obtained based on multi-directional scan modeling and coordinate-aware gating mechanism; Based on the normalized feature map, semantic scanning features are obtained, wherein the semantic scanning features are obtained based on semantic category selective aggregation modeling; Based on the geometric scanning features and the semantic scanning features, the semantic-geometric joint perception features are obtained.

[0012] Optionally, obtaining geometric scan features based on the normalized feature map includes: Based on the channel expansion result of the normalized feature map, the first sub-feature and the second sub-feature after segmentation are obtained; Based on the first sub-feature, obtain the features after convolution processing; Based on the features after convolution processing, multi-directional scanning features are obtained, wherein the multi-directional scanning features are obtained by performing multi-directional state space scanning on the features after convolution processing. Based on the features after convolution, coordinate-aware orientation weights are obtained, wherein the coordinate-aware orientation weights are obtained based on coordinate graph generation and convolution processing; The geometric scanning features are obtained based on the weighted fusion result of the multi-directional scanning features and the coordinate-aware direction weights.

[0013] Optionally, obtaining semantic scanning features based on the normalized feature map includes: Pixel sequence features are obtained based on the flattening result of the normalized feature map; Based on the dimensionality compression result of the pixel sequence features, obtain the semantic category score; Based on the semantic category scores, obtain the semantic selection matrix; Based on the semantic selection matrix and the semantic token, obtain the semantic hint vector; Based on the rearrangement result of the pixel sequence features and the semantic cue vector, the selective sequence scanning result is obtained; The semantic scanning features are obtained based on the dimensional adjustment and order recovery results of the selective sequence scanning results.

[0014] Optionally, obtaining the semantic-geometric joint sensing features based on the geometric scanning features and the semantic scanning features includes: Based on the block division results of the geometric scanning features and the semantic scanning features, obtain pixel-by-pixel channel fusion features; Based on the convolutional processing result of the pixel-by-pixel channel fusion features, the initial fusion features are obtained; Based on the segmentation results of the initial fusion features, the gate weights are obtained; The geometric scanning features and the semantic scanning features are weighted and fused according to the gating weights to obtain the gated fused features; Based on the residual connections and channel attention processing results of the gated fusion features, the semantic-geometric joint perception features are obtained.

[0015] Optionally, obtaining a high-resolution near-infrared remote sensing image based on the fusion result of the deep state characterization features and the multi-scale wavelet transform modulation fusion features includes: Based on the sum of the convolutional fine-tuning result of the deep state characterization features and the multi-scale wavelet transform modulation fusion feature, a fusion feature map is obtained; Based on the channel dimensionality reduction result of the fused feature map, preprocessed features are obtained; Based on the results of multiple upsampling processes of the preprocessed features, a high-resolution feature map is obtained, wherein each upsampling process includes channel expansion and pixel rearrangement operations; Based on the channel mapping results of the high-resolution feature map, a single-channel near-infrared intensity image is obtained; The high-resolution near-infrared remote sensing image is obtained based on the inverse normalization processing result of the single-channel near-infrared intensity image.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention effectively enhances the recovery of high-frequency texture details and edge structure information during super-resolution reconstruction of agricultural near-infrared remote sensing images by introducing a multi-scale wavelet transform feature modulation and semantic-geometric collaborative scanning mechanism. Specifically, the multi-scale wavelet transform separates and enhances information from different frequency bands, highlighting high-frequency details and suppressing redundant interference, thereby improving the detail clarity and texture fidelity of the reconstructed image. The semantic scanning branch aggregates and models global semantic information, enhancing the constraint on the overall structure and target consistency of the remote sensing scene, reducing structural misalignment and texture discontinuities. Simultaneously, the geometric scanning branch based on the Mamba state-space model effectively models long-range spatial dependencies, enabling cross-regional information propagation and structural compensation, improving the ability to maintain structural consistency in large-scale scenes. Furthermore, this invention introduces a coordinate-aware gating mechanism at the output of the geometric scanning branch, performing pixel-level adaptive weighted fusion of multi-directional scanning results, and achieving complementary enhancement of semantic and spatial features through a gating fusion module, thereby improving the robustness and generalization ability of the method in scenarios with complex terrain distribution and scale variations. The overall solution can effectively improve the visual quality and structural stability of super-resolution reconstruction of agricultural near-infrared remote sensing images, and has strong practical application value and promotion significance. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy according to an embodiment of the present invention. Figure 2Here are the framework diagrams of the embodiments of the present invention. (a) is the overall framework network diagram of the embodiment of the invention, and (b) is a structural schematic diagram of an important module in the framework network diagram. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] This embodiment proposes a near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy, such as... Figure 1 As shown, the specific steps include: Acquire low-resolution near-infrared remote sensing images; Based on the low-resolution near-infrared remote sensing image, obtain the preprocessed multi-channel feature map; Based on the multi-channel feature map, obtain the multi-scale wavelet transform modulation fusion features; Based on the multi-scale wavelet transform modulation fusion features, deep state representation features are obtained, wherein the deep state representation features are obtained by iteratively processing the semantic-geometric joint perception features, and the semantic-geometric joint perception features are obtained based on the collaborative processing of the semantic scanning branch and the geometric scanning branch. Based on the fusion result of the deep state characterization features and the multi-scale wavelet transform modulation fusion features, a high-resolution near-infrared remote sensing image is obtained.

[0021] Specifically, (1) the input low-resolution agricultural near-infrared remote sensing image is acquired and preprocessed before being used as network input; (2) The preprocessed agricultural near-infrared remote sensing image obtained in step (1) is sent to the multi-scale wavelet transform feature modulation module to obtain the multi-scale wavelet transform modulation fusion feature. (3) The features output by the previous module are sent to the residual semantic geometry co-processing module (RSMG) to obtain semantic geometry joint perception features.

[0022] (4) Repeat step (3) n times to obtain the deep state representation features.

[0023] (5) The deep state representation features obtained in step (4) are passed through a convolutional layer and then added to the multi-scale wavelet transform modulation fusion features obtained in step (2) before being sent to the reconstruction module to output a high-resolution agricultural near-infrared remote sensing image, thereby achieving super-resolution reconstruction.

[0024] Further, based on the low-resolution near-infrared remote sensing image, obtaining the preprocessed multi-channel feature map includes: Based on the single-channel image in the low-resolution near-infrared remote sensing image, a three-channel image is obtained, wherein each channel of the three-channel image is a near-infrared channel; Based on the three-channel image, an enhanced image is obtained after random horizontal flipping and rotation; based on the enhanced image, a multi-channel feature map is obtained after dimension adjustment and normalization.

[0025] Specifically, the dataset consists of near-infrared single-channel agricultural remote sensing images. Image files are read from disk and preprocessed. First, each single-channel image is copied three times to obtain three near-infrared agricultural remote sensing images. Then, random horizontal flipping and rotation are performed to enhance the data. Next, the dimensional order is adjusted from [H, W, C] to [C, H, W] and normalized, then fed into the network as the final input for computation. Each data set has dimensions (batch_size, channels, height, width), where batch_size is the batch size (4 in this example), C is the number of channels, and W and H are the resolutions of the infrared remote sensing images, serving as inputs for subsequent model processing.

[0026] Furthermore, based on the multi-channel feature map, obtaining multi-scale wavelet transform modulation fusion features includes: Based on the multi-channel feature map, first-scale features, second-scale features, and third-scale features are obtained, wherein the first-scale features, second-scale features, and third-scale features are extracted through different convolutional branches; Based on the first scale feature, a first wave domain feature map is obtained, wherein the first wave domain feature map is obtained by performing a discrete wavelet transform on the first scale feature. Based on the first wave domain feature map, obtain the spatial gating signal; The second scale feature and the third scale feature are dynamically modulated according to the spatial gating signal to obtain the modulated second scale feature and the modulated third scale feature; Based on the modulated second-scale features and the modulated third-scale features, an adaptive weighted fusion feature is obtained; The multi-scale wavelet transform modulation fusion feature is obtained by splicing the first scale feature, the second scale feature, the third scale feature and the adaptive weighted fusion feature.

[0027] Further, based on the first wave domain feature map, obtaining the spatial gating signal includes: Based on the channel compression result of the first wave domain feature map, obtain the activated features; Based on the channel compression result of the activated feature, obtain the initial gate value after scaling the numerical range; The spatial gating signal is obtained by adding the initial gating value to the preset offset.

[0028] Specifically, the calculation process for step (2) is as follows: First, the preprocessed low-resolution agricultural near-infrared remote sensing image obtained in step (1) is used to extract features through three receptive fields (3×3, 5×5, 7×7) using multi-scale feature extraction branches to obtain local texture, edge structure, and target information at different scales. Then, wavelet transform decomposition is performed on the features extracted from the 3×3 receptive field to obtain the low-frequency approximate component cA and the horizontal, vertical, and diagonal high-frequency detail components cH, cV, and cD. The frequency band features are used to generate modulation parameters (gate), which are then used to adaptively modulate and enhance the features of different receptive field branches (5×5, 7×7), thereby highlighting key high-frequency detail information and suppressing redundant features. The enhanced receptive field branch features are then fed into an adaptive fusion module (Sknet) to generate a fused feature containing the common information of both. Finally, the multi-scale branch features and the fused feature are concatenated and fed into a 3×3 convolution to obtain a multi-scale wavelet transform modulated fusion feature. This feature integrates the original multi-scale information and the enhanced features modulated by wavelet information, providing high-quality input for subsequent deep processing.

[0029] Furthermore, obtaining the semantic-geometric joint sensing features based on the multi-scale wavelet transform modulation fusion features includes: Based on the multi-scale wavelet transform modulation fusion features, a normalized feature map is obtained. Based on the normalized feature map, geometric scan features are obtained, wherein the geometric scan features are obtained based on multi-directional scan modeling and coordinate-aware gating mechanism; Based on the normalized feature map, semantic scanning features are obtained, wherein the semantic scanning features are obtained based on semantic category selective aggregation modeling; Based on the geometric scanning features and the semantic scanning features, the semantic-geometric joint perception features are obtained.

[0030] Further, obtaining geometric scan features based on the normalized feature map includes: Based on the channel expansion result of the normalized feature map, the first sub-feature and the second sub-feature after segmentation are obtained; Based on the first sub-feature, obtain the features after convolution processing; Based on the features after convolution processing, multi-directional scanning features are obtained, wherein the multi-directional scanning features are obtained by performing multi-directional state space scanning on the features after convolution processing. Based on the features after convolution, coordinate-aware orientation weights are obtained, wherein the coordinate-aware orientation weights are obtained based on coordinate graph generation and convolution processing; The geometric scanning features are obtained based on the weighted fusion result of the multi-directional scanning features and the coordinate-aware direction weights.

[0031] Further, based on the normalized feature map, obtaining semantic scanning features includes: Pixel sequence features are obtained based on the flattening result of the normalized feature map; Based on the dimensionality compression result of the pixel sequence features, obtain the semantic category score; Based on the semantic category scores, obtain the semantic selection matrix; Based on the semantic selection matrix and the semantic token, obtain the semantic hint vector; Based on the rearrangement result of the pixel sequence features and the semantic cue vector, the selective sequence scanning result is obtained; The semantic scanning features are obtained based on the dimensional adjustment and order recovery results of the selective sequence scanning results.

[0032] Further, obtaining the semantic-geometric joint perception features based on the geometric scanning features and the semantic scanning features includes: Based on the block division results of the geometric scanning features and the semantic scanning features, obtain pixel-by-pixel channel fusion features; Based on the convolutional processing result of the pixel-by-pixel channel fusion features, the initial fusion features are obtained; Based on the segmentation results of the initial fusion features, the gate weights are obtained; The geometric scanning features and the semantic scanning features are weighted and fused according to the gating weights to obtain the gated fused features; Based on the residual connections and channel attention processing results of the gated fusion features, the semantic-geometric joint perception features are obtained.

[0033] Specifically, the calculation process for step (3) is as follows: First, the features output from the previous module are input into the Residual Semantic-Geometric Co-engineering Module (RSMG), which consists of multiple Semantic-Geometric Co-engineering Modules (RSSD). Within each RSSD module, the input RSSDInput is first normalized by an infeed layer, and then fed into the Semantic-Spatial Coordinate Dual-Branch Fusion Module (DSSM). The DSSM is constructed by parallel modeling of the Semantic Scanning Branch (ASSM) and the Geometric Scanning Branch (VSSM). The input is fed into each of the two branches. In the VSSM branch, the features are first expanded through a linear layer and then divided into two parts along the channel dimension. One part serves as the main feature branch for subsequent spatial modeling, while the other part is retained as a gated branch for output modulation. The main feature branch is then processed through depthwise separable convolution and nonlinear activation for local feature extraction before being input into the 2D-SSM. Parallel scanning in four directions is performed to aggregate multi-directional global contextual information. Simultaneously, a coordinate-aware mechanism is introduced. By constructing a normalized spatial coordinate map and fusing it with the features to generate directional weights, the four-directional scanning results are adaptively weighted and summed pixel-by-pixel to obtain the four-directional spatial fusion features. Subsequently, the feature is normalized and multiplied element-wise with the retained gated branch, then projected back to the original channel dimension output through a linear layer. The ASSM branch flattens the feature into a sequence, which is fed into a routing network composed of fully connected layers, non-linear activation functions, and fully connected layers to predict the probability of each pixel location in a predefined set of semantic categories and form discrete selections using the Gumbel-Softmax activation function. Based on this, the corresponding semantic cue vector is extracted from the semantic token (the semantic token is not extracted from the image, but is a set of learnable parameters registered during the initialization phase), and the sequence is rearranged according to the semantic label. The rearranged sequence is then position-modulated by depthwise convolution and fed into the Selective Sequence Scanner (ASE) operator. This operator first performs a linear mapping on the input sequence and then cuts the last dimension into three parameter vectors, including Δt (time step control), B4 (input matrix), and C (output matrix). Subsequently, for each sequence position, the C branch is added element-wise with the semantic cue vector to obtain the semantically modulated vector. Based on this, forward and reverse state recursion scans are performed along the sequence direction, and the bidirectional scan results are fused at each position to finally restore the original spatial order and obtain the semantically enhanced feature map.Subsequently, the outputs from VSSM and ASSM are adaptively weighted and fused through a Dynamic Per-Pixel Channel Fusion Module (DPCF). This module divides the feature maps from the spatial scan branch and the geometric scan branch into four blocks on average along the channel dimension. Internally, the module holds a learnable scalar d (a single-valued vector of shape [1,1,1,1], randomly initialized with the network and updated via backpropagation during training). This d is gating with 0 to 1 using a Sigmoid function and is broadcast-expanded to the spatial size of each pair of blocks during runtime. Each block is weighted and summed to obtain the fused block. After weighting each of the four pairs of blocks, they are concatenated along the channel dimension to reconstruct the complete fused feature map. Finally, this feature is processed by a lightweight network consisting of a 1×1 convolution (128→128), batch normalization, and ReLU activation to obtain the joint enhanced feature. A learnable channel scaling vector, scale1, is then used to multiply the backbone input feature RSSDInput channel-wise by a scaling factor. This scaled feature is then added element-wise with the joint enhanced feature to form the initial fused feature S. Subsequently, feature S is normalized and fed into a lightweight convolutional network consisting of a 1×1 convolution, a non-linear activation function (GELU), and another 1×1 convolution for feature transformation, yielding feature Smiddle. Next, a channel attention module is introduced, using global average pooling to obtain channel statistics. Channel weights w are generated through two convolutional layers, the ReLU activation function, and a sigmoid function. These weights w are then multiplied back into feature Smiddle by channel to obtain feature H, thus highlighting channels more critical for reconstruction and suppressing redundant information. S is then multiplied by a second channel scaling vector (scale2) and added element-wise to H to form the second residual. This yields the final block-level output representation. This feature is then flattened into a sequence form and passed as output to the next RSSD. After multiple RSSD iterations, a sequence reshaping and a 3×3 convolution are performed, followed by element-wise addition to the RSMG module input to obtain the semantic-geometric joint perceptual feature.

[0034] Furthermore, based on the fusion result of the deep state characterization features and the multi-scale wavelet transform modulation fusion features, obtaining a high-resolution near-infrared remote sensing image includes: Based on the sum of the convolutional fine-tuning result of the deep state characterization features and the multi-scale wavelet transform modulation fusion feature, a fusion feature map is obtained; Based on the channel dimensionality reduction result of the fused feature map, preprocessed features are obtained; Based on the results of multiple upsampling processes of the preprocessed features, a high-resolution feature map is obtained, wherein each upsampling process includes channel expansion and pixel rearrangement operations; Based on the channel mapping results of the high-resolution feature map, a single-channel near-infrared intensity image is obtained; The high-resolution near-infrared remote sensing image is obtained based on the inverse normalization processing result of the single-channel near-infrared intensity image.

[0035] Specifically, the calculation process for step (5) is as follows: The deep state representation features obtained in step (4) are passed through a 3×3 convolutional layer and then globally element-wise added to the multi-scale wavelet transform modulation fusion features obtained in step (2). Subsequently, an upsampling operation is performed to enlarge the spatial resolution of the feature map, resulting in a high-resolution image. For a 4x super-resolution, this operation is performed twice consecutively. Then, a final 3×3 convolutional layer is applied to the upsampled feature map, compressing its channel count from 64 back to 3 channels. Channel compression and weighted fusion are then performed to convert it into a single-channel form. Finally, the pixel values ​​are denormalized to restore the network output values ​​to the pixel range of a standard image, generating the final high-resolution image.

[0036] The following is a detailed description of this embodiment with reference to the accompanying drawings: like Figure 2 As shown in (a)-(b), the specific process of this embodiment is as follows: (1) The training dataset consists of near-infrared single-channel agricultural remote sensing images, where the low-resolution images are 16×16 in size and the high-resolution images are 64×64 in size. The input low-resolution near-infrared remote sensing images are acquired and preprocessed before being used as network input. Specifically, the agricultural remote sensing near-infrared single-channel images are first read from the dataset, with dimensions of [4,1,16,16]. This channel is then copied three times to obtain a three-channel image with dimensions of [4,3,16,16]. Next, data augmentation is performed on the images, including random horizontal flipping and rotations of 90 / 180 / 270 degrees. At the same time, the feature values ​​are normalized to the floating-point range of [0.0,1.0].

[0037] (2) The preprocessed features from step (1) are input in parallel into three convolutional branches: a 3×3 small-scale branch, a 5×5 medium-scale branch, and a 7×7 large-scale branch, to extract feature maps f3, f5, and f7 with different receptive fields, each with dimensions [4, 32, 16, 16]. Subsequently, only the small-scale feature f3 is fed into the wavelet transform module (WTFMBlock). This module performs discrete wavelet transform on f3, decomposing it into low-frequency approximate components cA and high-frequency detail components cH, cV, and cD in the horizontal, vertical, and diagonal directions. These four sub-band features are concatenated along the channel dimension to form a wavelet domain feature map f_wave with dimensions [4, 128, 16, 16]. f_wave is passed through a gated generation network (gate_gen), which first passes through a 1×1 convolution to compress its channel number from 128 to 64, changing the dimension to [4, 64, 16, 16]. The result is then activated with SiLU, without changing the dimension. Another 1×1 convolution further compresses the number of channels from 64 to 32. The Sigmoid function is applied to the output, scaling its numerical range to [0,1], and then adding 1 generates a spatially gated signal (dimensional [4,32,16,16]), resulting in a final range of [1,2]. This structure is a stable gating mechanism, ensuring that the gating retains at least the original features (multiplied by 1) while learning where enhancements should be applied (multiplied by values ​​greater than 1). This gate is multiplied pixel-wise onto the mesoscale feature f5 and the large-scale feature f7, achieving dynamic modulation of them to obtain modulated features f5m and f7m, both with dimensions [4,32,16,16]. Finally, f5m and f7m are adaptively fused through a SkNet module. First, f5m and f7m are added element-wise to generate a fused feature u containing information shared by both, with dimensions still [4,32,16,16]. Next, in the Select phase, the module compresses the spatial information of u into a dimension of [4,32,1,1] using global average pooling. This compressed information is then fed into a mini-network consisting of two 1×1 convolutions. The first convolution reduces the number of channels from 32 to 2, and the second convolution expands it to 64, generating a weight vector of dimension [4,64,1,1]. This vector is split into two parts, corresponding to the weights of the two branches, and normalized using the Softmax function to obtain two weight features of dimension [4,32,1,1]. Finally, these two weight features are element-wise multiplied onto the original input feature maps f5m and f7m, respectively, and the results are summed to output a dynamically weighted and fused feature map f_fuse with dimensions [4,32,16,16].The original f3, f5, f7 and the fused f_fuse are concatenated along the channel dimension, and then subjected to a 3×3 convolution to finally output the multi-scale wavelet transform modulation fusion feature d, which has dimensions [4, 128, 16, 16].

[0038] (3) The multi-scale wavelet transform modulation fusion feature d obtained in step (2) is first flattened and reshaped into a feature map with dimensions [batch_size,H,W,C], i.e. [4,16,16,128]. Then, it is input into a deep feature extraction network composed of 6 cascaded Semantic Geometric Coordination Groups (RSMGs), each of which contains 8 Semantic Geometric Coordination Modules (RSSDs). Inside each core RSSD, the feature map first enters LayerNorm (layer normalization) to normalize the features. It calculates the mean and variance of each sample feature and then uses them to adjust the distribution of the features so that the mean is 0 and the variance is 1, resulting in feature m with dimensions [4,16,16,128]. Then, it is fed into the Semantic-Spatial Coordinate Dual-Branch Fusion Module (DSSM), which contains two branches: the Geometric Scanning Branch (VSSM) and the Semantic Scanning Branch (ASSM). First, the Geometric Scan branch (VSSM) passes feature m through a linear layer, expanding the channel dimension from C(128) to 512, i.e., the dimension change from [4,16,16,128] to [4,16,16,512]. The expanded feature is split into two parts, i and z, from the last dimension. z is temporarily reserved for subsequent gating operations. The dimensions of i and z are both [4,16,16,256]. i is adjusted to the [batch_size,C,H,W] format ([4,256,16,16]) and then passed through a 3×3 depthwise separable convolution (conv2d) and the SiLU activation function to obtain feature i1, whose dimension remains unchanged. Feature i1 is then fed into the core 2D-SSM module. This function performs four independent scans in parallel: from top left to bottom right, from top right to bottom left, from bottom left to top right, and from bottom right to top left. Four independent output feature maps are obtained, all with the same dimensions as the input i1, i.e., [4, 256, 16, 16]. Each output contains global information aggregated from a specific direction. The convolutionally processed feature map i1 is then fed in parallel into the GeoCoordGate4Map module to generate orientation-aware weights. Internally, the GeoCoordGate4Map module first calls the CoordMap module to generate a coordinate tensor with the same resolution as the input feature map. This module writes the x-coordinate and y-coordinate normalized to [-1, 1] for each position, and then calculates the radius r = √(x² + y²) from that pixel to the center of the image, generating a coordinate map of the same size as i1 with 3 channels. Channel 0 stores the x-value, channel 1 stores the y-value, and channel 2 stores the r-value. The coordinate map dimensions are [4, 3, 16, 16]. The feature map i1 and the coordinate map are concatenated along the channel dimension. After concatenation, the dimension changes from [4,256,16,16] to [4,259,16,16].Then, a 1×1 convolution is used to fuse the channels, resulting in a fused feature map with dimensions [4, 256, 16, 16]. The fused features are then activated by a 3×3 depthwise convolution followed by a SiLU activation function, and then compressed to 4 channels by another 1×1 convolution, ultimately outputting a 4-channel feature map with dimensions [4, 4, 16, 16]. A Softmax function is applied to these 4 channels to generate orientation-aware weights w_map with dimensions [4, 4, 16, 16]. These 4 channels correspond to the four scanning directions of the 2D-SSM. At each pixel location (h, w), the sum of the values ​​of these 4 channels equals 1, representing the extent to which the features at that location depend on information from the four different directions. The four channels of w_map (dimension [4,4,16,16]) are separated and used as weights for the four directions. These weights are then summed element-wise with the four independent output feature maps from the 2D-SSM module to output a four-directional spatial fusion feature k with dimensions [4,256,16,16]. The fused k is then processed by LayerNorm, multiplied by the previously retained z, and projected back to the original number of channels C (128) through a linear layer, finally outputting dimensions [4,128,16,16]. Then, the semantic scanning branch (ASSM) flattens the input feature map m in row-major order into a pixel sequence m1 with dimensions [B,L,C] (B=4, L=H×W=256, C=128). A fully connected layer then compresses the feature dimension from 128 to 42 ([4,256,128]→[4,256,42]). Next, the non-linear activation function GELU is used to increase non-linearity, and another fully connected layer maps the 42 dimensions to 64-dimensional category scores ([4,256,42]→[4,256,64]). Finally, LogSoftmax is applied to the category dimension to obtain a score [4,256,64]—this generates a log probability score for each pixel location in the batch against 64 semantic categories, resulting in a score for each location against 64 semantic categories. Gumbel-Softmax then discretizes the score into a one-hot selection matrix [4,256,64]. This step is equivalent to selecting a "semantic label" for each of the 256 pixels. Next, the corresponding row vector is extracted from the semantic token and multiplied by the one-hot selection matrix to generate a pixel-wise semantic cue vector P[4,256,16]. To ensure that pixels with the same label are adjacent in the sequence, m1 is rearranged into m2 according to the semantic label index; at the same time, the inversion index is recorded for later restoration of the order. m2 is first reshaped into a 4D feature map with dimensions [4,128,16,16], then raised to 256 dimensions [4,256,16,16] by 1×1 convolution, then flattened back to the sequence dimension [4,256,256] and position modulated by depthwise convolution to obtain m3[4,256,256].Then, m3 and P are fed into the ASE (Selective Sequence Scan Operator). Internally, the input sequence m3 is first linearly mapped to [4,256,48] using a 1×1 method. Then, the last dimension is cut into three parameter vectors: Δt (time step control) with dimensions [4,256,16], B4 (input matrix) [4,256,16], and C (output matrix) [4,256,16]. Adding the learnable D (jump coefficient, dimension 16) and the learnable A (state decay, diagonal length 16), there are a total of 5 parameter vectors. Subsequently, for each sequence position, the C branch is element-wise added to the semantic cue vector P to obtain the semantically modulated vector. After this preparation, a forward and a reverse recursive scan are performed along the 256-length sequence axis; the forward scan runs from the 1st position to the 256th position, and the reverse scan runs from the 256th position back to the 1st position. Two scanning trajectories each generate a 256-dimensional output vector at each position. These two vectors are then added element-wise to obtain the fused result y_core, with the overall dimensions remaining [4,256,256]. After scanning, the dimensions of y_core are rearranged. First, the channel dimension is moved to the center, and then the 256 sequence positions are folded back into a 16×16 spatial grid, maintaining the numerical order as before scanning. This yields y′, still in shape [4,256,256], but arranged in a manner consistent with the format required by subsequent networks. y′ is reduced to 128 dimensions using LayerNorm and a linear layer to obtain g[4,256,128]. The original pixel order is then restored using the previously saved inverse index, and the resulting spatial feature map [4,128,16,16] is returned to the main network. Finally, Dynamic Per-Pixel Channel Fusion (DPCF) is performed. First, the feature maps output by the spatial scan branch and the geometric scan branch, both with dimensions [4, 128, 16, 16], are divided into four blocks along the channel dimension, each with a size of [4, 32, 16, 16]. The module internally holds a learnable scalar d (a single-valued vector of shape [1, 1, 1, 1], randomly initialized with the network and updated via backpropagation during training). This d is then gated with 0-1 using a sigmoid function and broadcast to the spatial size of each pair of blocks during runtime. The weighted summation of each block yields the fused block, still with a shape of [4, 32, 16, 16]. After weighting each of the four pairs of blocks, they are concatenated along the channel dimension to reconstruct the complete fused feature map, restoring it to [4, 128, 16, 16]. Finally, this feature is processed through a lightweight network—a single-layer 1×1 convolution (128→128), batch normalization, and ReLU activation—to obtain the joint enhancement feature y_fused with dimensions [4,128,16,16]. This output is then split into two parts, gate_vssm and gate_assm, each with dimensions [4,128,16,16].Finally, the values ​​of these two gates are normalized using a Sigmoid function, making them weights. The learned gate weights are applied pixel-by-pixel to the feature maps output by the spatial scan branch and the geometric scan branch, and then weighted and summed to obtain the joint enhancement feature y_fused with dimensions [4, 128, 16, 16]. Then, a learnable channel scaling vector scale1 (all-1 learnable channel weights registered directly using the constructor during RSSD initialization, automatically updated by gradients during training, and after a long training period, some channels become >1 (amplification), and some channels become <1 (suppression), thus realizing learnable residual branch strength) is broadcast to expand to the dimension [4, 16, 16, 128]. Each channel is multiplied by the corresponding coefficient, and the 128 channels of the RSSD module input feature map RSSDInput are multiplied by independent scaling factors respectively, resulting in RSSDInput × scale1 with dimensions still [4, 16, 16, 128]. Then, it is added element-wise with the joint enhancement feature y_fused to obtain S. Then, S is passed through a LayerNorm layer (pure normalization, size unchanged) and fed into a lightweight convolutional network, specifically composed of a 1×1 convolution, a GELU activation function, and another 1×1 convolution, to obtain the feature Smiddle. Next, it enters the channel attention module, where global average pooling is performed on the spatial dimension to obtain the channel description vector [4,1,1,128]. Then, two 1×1 convolution layers are used; the first layer compresses the channels from 128 to 4, resulting in [4,1,1,4]. After a ReLU activation function, the second layer maps it back to 128, obtaining [4,1,1,128]. Subsequently, a Sigmoid function is used to normalize it to channel weights w, ranging from 0 to 1. Finally, these weights are multiplied back into Smiddle by channel, and the output H remains [4,16,16,128], achieving enhancement of important channels and suppression of ineffective channels. Then, S is multiplied by the second learnable channel scaling vector scale2 and added element-wise to H to form the second residual. Finally, the features are flattened into a sequence along the space to obtain the semantic-geometric joint sensing features with dimensions [B,H×W,C]=[4,256,128], which then proceed to the next RSSD. One RSMG contains 8 RSSDs, and their output dimensions are [4,256,128]. The group tail reshapes the sequence back into the feature map [4,128,16,16], fine-tunes it through a 3×3 convolution, then folds it back to [4,256,128] and adds it to the input sequence of the RSMG module to form a group-level residual loop, thus obtaining the semantic-geometric joint sensing features.

[0039] (4) Repeat step (3) 6 times with the semantic-geometric joint perception features obtained in step (3) as input. After all RSMGs are cascaded, the overall deep sequence is summarized as [4,256,128]. Then, through LayerNorm and reshaping back to two-dimensional feature map, the deep state representation features [4,128,16,16] are finally obtained.

[0040] (5) The deep state representation features obtained in step (4) are fine-tuned by a 3×3 convolutional layer, keeping their dimensions [4,128,16,16]. Then, it is added pixel-by-pixel to the multi-scale wavelet transform modulation fusion feature d, which also has dimensions [4,128,16,16], to obtain a fused feature map with dimensions [4,128,16,16]. Next, this fused feature map enters the upsampling reconstruction module. First, a preprocessing convolutional layer reduces the number of feature channels from 128 to 64, resulting in a feature map with dimensions [4,64,16,16]. Subsequently, a first 2x upsampling is performed: the feature map [4,64,16,16] is expanded by a factor of 4 in the channel dimension using a 3×3 convolution, generating a feature map [4,256,16,16]. Then, the PixelShuffle operation is used to spatially rearrange this tensor, mapping the channel information to the spatial dimension, resulting in a feature map of dimension [4,64,32,32]. Next, the above process is repeated for a second 2x upsampling, where the [4,64,32,32] feature map is first expanded to [4,256,32,32] using a convolution, and then rearranged using PixelShuffle to obtain a high-resolution feature map of dimension [4,64,64,64]. Subsequently, a 3×3 convolutional layer maps the feature map [4,64,64,64] into a three-channel form, obtaining an intermediate reconstruction result with dimensions [4,3,64,64]. An output mapping convolutional layer then performs channel compression and weighted fusion on the three-channel features, converting them into a single-channel form, resulting in a near-infrared intensity image with dimensions [4,1,64,64]. Finally, the output result undergoes inverse normalization, i.e., the mean subtracted in the preprocessing stage is added, and the numerical range is restored. The final output is a high-resolution near-infrared remote sensing image with numerical values ​​in the range [0.0,1.0] and dimensions [4,1,64,64], completing the super-resolution reconstruction task.

[0041] In this embodiment, the input low-resolution agricultural near-infrared remote sensing image is first normalized and preprocessed, then input into a multi-scale feature extraction module. This module extracts multi-scale shallow features through different receptive field branches to capture more detailed information. Simultaneously, wavelet transform is used to perform frequency band decomposition and modulation enhancement on the features to highlight high-frequency texture details and suppress redundant information. Next, the modulated features and features from different receptive fields are fused to obtain multi-scale wavelet transform modulated fused features, which are then input into a semantic-geometric synergy module. This module first inputs a semantic-spatial coordinate dual-branch fusion module for deep modeling. The semantic scanning branch is used to aggregate and model global semantic information, while the geometric scanning branch performs multi-directional spatial scanning based on the Mamba state space model to enhance long-range dependencies and structural consistency. Subsequently, a coordinate-aware gating mechanism is introduced at the output of the spatial scanning branch to generate pixel-level directional weights. The multi-directional scanning results are adaptively weighted and fused, and the spatial scanning features and semantic scanning features are complementaryly fused through a gating fusion module to obtain jointly enhanced features. Based on this, the jointly enhanced features are stabilized through residual modulation, channel attention enhancement, and group-level residual closure, and then input into the reconstruction module for upsampling and mapping output to generate high-resolution near-infrared remote sensing images, thus achieving super-resolution reconstruction. This method can effectively improve the detail clarity and structural consistency of agricultural near-infrared remote sensing image super-resolution reconstruction, and has strong robustness and practical application value.

[0042] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy, characterized in that, include: Acquire low-resolution near-infrared remote sensing images; Based on the low-resolution near-infrared remote sensing image, obtain the preprocessed multi-channel feature map; Based on the multi-channel feature map, obtain the multi-scale wavelet transform modulation fusion features; Based on the multi-scale wavelet transform modulation fusion features, deep state representation features are obtained, wherein the deep state representation features are obtained by iteratively processing the semantic-geometric joint perception features, and the semantic-geometric joint perception features are obtained based on the collaborative processing of the semantic scanning branch and the geometric scanning branch. Based on the fusion result of the deep state characterization features and the multi-scale wavelet transform modulation fusion features, a high-resolution near-infrared remote sensing image is obtained.

2. The near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy according to claim 1, characterized in that, Based on the low-resolution near-infrared remote sensing image, the preprocessed multi-channel feature map is obtained, including: Based on the single-channel image in the low-resolution near-infrared remote sensing image, a three-channel image is obtained, wherein each channel of the three-channel image is a near-infrared channel; Based on the three-channel image, an enhanced image is obtained after random horizontal flipping and rotation; based on the enhanced image, a multi-channel feature map is obtained after dimension adjustment and normalization.

3. The near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy as described in claim 1, characterized in that, Based on the multi-channel feature map, the multi-scale wavelet transform modulation fusion features are obtained, including: Based on the multi-channel feature map, first-scale features, second-scale features, and third-scale features are obtained, wherein the first-scale features, second-scale features, and third-scale features are extracted through different convolutional branches; Based on the first scale feature, a first wave domain feature map is obtained, wherein the first wave domain feature map is obtained by performing a discrete wavelet transform on the first scale feature. Based on the first wave domain feature map, obtain the spatial gating signal; The second scale feature and the third scale feature are dynamically modulated according to the spatial gating signal to obtain the modulated second scale feature and the modulated third scale feature; Based on the modulated second-scale features and the modulated third-scale features, an adaptive weighted fusion feature is obtained; The multi-scale wavelet transform modulation fusion feature is obtained by splicing the first scale feature, the second scale feature, the third scale feature and the adaptive weighted fusion feature.

4. The near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy according to claim 3, characterized in that, Based on the first wave domain feature map, obtaining the spatially gated signal includes: Based on the channel compression result of the first wave domain feature map, obtain the activated features; Based on the channel compression result of the activated feature, obtain the initial gate value after scaling the numerical range; The spatial gating signal is obtained by adding the initial gating value to the preset offset.

5. The near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy according to claim 1, characterized in that, Based on the multi-scale wavelet transform modulation fusion features, the semantic-geometric joint sensing features are obtained as follows: Based on the multi-scale wavelet transform modulation fusion features, a normalized feature map is obtained. Based on the normalized feature map, geometric scan features are obtained, wherein the geometric scan features are obtained based on multi-directional scan modeling and coordinate-aware gating mechanism; Based on the normalized feature map, semantic scanning features are obtained, wherein the semantic scanning features are obtained based on semantic category selective aggregation modeling; Based on the geometric scanning features and the semantic scanning features, the semantic-geometric joint perception features are obtained.

6. The near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy according to claim 5, characterized in that, Obtaining geometric scan features based on the normalized feature map includes: Based on the channel expansion result of the normalized feature map, the first sub-feature and the second sub-feature after segmentation are obtained; Based on the first sub-feature, obtain the features after convolution processing; Based on the features after convolution processing, multi-directional scanning features are obtained, wherein the multi-directional scanning features are obtained by performing multi-directional state space scanning on the features after convolution processing. Based on the features after convolution, coordinate-aware orientation weights are obtained, wherein the coordinate-aware orientation weights are obtained based on coordinate graph generation and convolution processing; The geometric scanning features are obtained based on the weighted fusion result of the multi-directional scanning features and the coordinate-aware direction weights.

7. The near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy according to claim 5, characterized in that, Obtaining semantic scanning features based on the normalized feature map includes: Pixel sequence features are obtained based on the flattening result of the normalized feature map; Based on the dimensionality compression result of the pixel sequence features, obtain the semantic category score; Based on the semantic category scores, obtain the semantic selection matrix; Based on the semantic selection matrix and the semantic token, obtain the semantic hint vector; Based on the rearrangement result of the pixel sequence features and the semantic cue vector, the selective sequence scanning result is obtained; The semantic scanning features are obtained based on the dimensional adjustment and order recovery results of the selective sequence scanning results.

8. A near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy as described in claim 5, characterized in that, Obtaining the semantic-geometric joint perception features based on the geometric scanning features and the semantic scanning features includes: Based on the block division results of the geometric scanning features and the semantic scanning features, obtain pixel-by-pixel channel fusion features; Based on the convolutional processing result of the pixel-by-pixel channel fusion features, the initial fusion features are obtained; Based on the segmentation results of the initial fusion features, the gate weights are obtained; The geometric scanning features and the semantic scanning features are weighted and fused according to the gating weights to obtain the gated fused features; Based on the residual connections and channel attention processing results of the gated fusion features, the semantic-geometric joint perception features are obtained.

9. A near-infrared remote sensing image super-resolution method based on multi-scale wavelet feature modulation and semantic-geometric synergy as described in claim 1, characterized in that, Based on the fusion result of the deep state characterization features and the multi-scale wavelet transform modulation fusion features, high-resolution near-infrared remote sensing images are obtained, including: Based on the sum of the convolutional fine-tuning result of the deep state characterization features and the multi-scale wavelet transform modulation fusion feature, a fusion feature map is obtained; Based on the channel dimensionality reduction result of the fused feature map, preprocessed features are obtained; Based on the results of multiple upsampling processes of the preprocessed features, a high-resolution feature map is obtained, wherein each upsampling process includes channel expansion and pixel rearrangement operations; Based on the channel mapping results of the high-resolution feature map, a single-channel near-infrared intensity image is obtained; The high-resolution near-infrared remote sensing image is obtained based on the inverse normalization processing result of the single-channel near-infrared intensity image.

Images