Landsat-moderate resolution imaging spectroradiometer spatio-temporal fusion method based on double-layer cross attention mechanism

Abstract

The application discloses a Landsat-MODIS remote sensing image spatio-temporal fusion method based on a double-layer cross attention mechanism, and relates to the technical fields of remote sensing image processing and artificial intelligence.The application realizes the dual improvement of the time continuity of the ground reflectivity and the spatial resolution by the design of the DCAF model, the utilization of a series of operations of 'feature extraction-differential weighting-information enhancement', and the full use of the complementary information of Landsat and MODIS images, and has the following main advantages: 1) the number of input images is reduced, only two input images are needed to realize the high-quality fusion effect, the difficulty and cost of data acquisition are reduced, and the practicability of the model is significantly improved; 2) in the aspect of feature processing, the model can accurately capture the spatial-spectral correlation between the low-resolution and high-resolution images by integrating various attention mechanisms and deep convolution modules, thereby showing excellent fusion performance in complex scenes.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and artificial intelligence, specifically to a spatiotemporal fusion method for Landsat-MODIS remote sensing images based on a two-layer cross-attention mechanism. Background Technology

[0002] Satellite remote sensing imagery offers advantages such as wide-area coverage and periodic revisit capabilities, providing crucial data for continuous Earth observation. However, the spatial and temporal resolutions of remote sensing imagery are often mutually restrictive, making it difficult for a single remote sensing platform to achieve continuous, high-resolution monitoring of the Earth's surface. To address this, many researchers have proposed multi-sensor spatiotemporal fusion techniques. By integrating multi-source remote sensing data with complementary spatiotemporal characteristics, these techniques generate continuous surface observation data with both high spatiotemporal resolution, providing vital support for refined surface monitoring. Currently, spatiotemporal fusion methods can be broadly categorized into three types: weighted function-based methods, unmixing-based methods, and learning-based methods.

[0003] Weighted function-based methods estimate pixel values ​​in a target high-resolution image by weighting and combining information from the input image, and are currently the most widely used approach. The core idea is to distinguish low-resolution pixels into homogeneous and heterogeneous pixels, employing linear modeling and weighting with neighboring similar pixels respectively. This effectively reduces the impact of spectral mixing effects on the fusion result and improves the prediction accuracy of heterogeneous pixels. For example, spatiotemporal adaptive reflectance fusion models combine weighted functions with detailed pixel spectral variation information to obtain predicted values ​​for target pixels. These methods have advantages such as simple parameters and high computational efficiency; however, most weighting models are based on simple linear relationships, making it difficult to accurately describe instantaneous changes in the land surface, and the prediction performance for scenarios with abrupt changes in land surface features still needs improvement.

[0004] Unmixing-based methods utilize linear spectral mixing theory, assuming that coarse pixels are linear combinations of the mixed responses for each land cover category. These coarse pixels are then unmixed to predict pixel values ​​in high-resolution images. For example, Wu et al. proposed a spatiotemporal data fusion method that considers both spatial and temporal nonlinear information, solving for the average variation values ​​of various endmembers to improve the intra-class variability of the fusion results. Although these methods have a sound theoretical foundation, the mixed pixel decomposition process requires classification to estimate component abundance. Due to the significant uncertainty in the classification process, the fusion accuracy is significantly reduced.

[0005] In recent years, learning-based spatiotemporal fusion methods have developed rapidly. These methods achieve prediction of high-resolution temporal data by establishing complex nonlinear mapping relationships between multi-sensor data. Currently, various machine learning algorithms have been successfully applied to the field of spatiotemporal fusion, including models such as regression trees, random forests, artificial neural networks, convolutional neural networks, and generative adversarial networks. For example, Lin et al. introduced a cross-attention mechanism to achieve adaptive allocation of spatial and spectral domain feature weights, thereby capturing high-precision temporal and spatial variation information. Zhang et al. integrated wavelet transforms into a generative adversarial network (GAN), significantly improving the prediction accuracy of high-resolution temporal data by training high-frequency and low-frequency components of the image separately. Compared with traditional fusion methods, learning-based methods exhibit stronger robustness in complex land cover scenarios and can deeply mine multi-scale feature information of images, thereby generating high-precision fused data. Most existing spatiotemporal fusion models require three or more input images; however, due to cloud cover, temporal mismatch between coarse and fine resolution images, it is difficult to collect sufficient input images in practical applications.

[0006] Therefore, a new solution is needed to address the above problems by constructing a spatiotemporal fusion method based on dual inputs. Summary of the Invention

[0007] The purpose of this invention is to provide a spatiotemporal fusion method for Landsat-MODIS remote sensing images based on a two-layer cross-attention mechanism, so as to solve the technical problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a spatiotemporal fusion method for Landsat-MODIS remote sensing images based on a two-layer cross-attention mechanism, comprising at least the following steps:

[0009] S1: Collect satellite remote sensing data from multiple data sources, including at least Landsat surface reflectance data, MOD09GA surface reflectance data, and MOD09GQ surface reflectance data;

[0010] S2: The data obtained in S1 is preprocessed by resampling all remote sensing images to 30 meters using the nearest neighbor interpolation method; Landsat is selected as a reference to ensure that the image sizes are exactly the same; in addition, the red and near-infrared bands of MOD09GQ are selected to rematch the band sequences with MOD09GA to enhance the spatial details of the fusion results.

[0011] S3: Construct a DCAF model, which consists of a feature extraction module, a differential weighting module, and a feature enhancement module;

[0012] S4: The DCAF model is trained by using a combined loss function. Once the overall loss converges, the trained DCAF model is used to predict surface reflectance data with high spatiotemporal resolution.

[0013] Furthermore, the feature extraction module includes a multi-channel cross-feature attention module, a full-feature joint cross-attention module, and a residual dense connection module with an embedded attention mechanism;

[0014] The application of the feature extraction module includes at least the following steps:

[0015] For the input Landsat reference day image (i.e., a high-resolution image):

[0016] First, feature maps of the Landsat image are initially extracted using convolutional layers;

[0017] Secondly, a residual dense connection module with an embedded attention mechanism is used to enhance the Landsat feature map and further mine the deeper information of the Landsat feature map.

[0018] For the input MODIS daytime image (i.e., a low-resolution image):

[0019] First, a multi-channel cross-feature attention module is used to calculate spatial and spectral weights based on the Landsat input image to achieve preliminary calibration of the MODIS image;

[0020] Secondly, a residual dense connection module with an embedded attention mechanism is used to extract deep features from the pre-calibrated MODIS feature map to obtain an enhanced MODIS feature map.

[0021] Finally, a joint cross-attention module for all features is designed to collaboratively process the Landsat feature map and the enhanced MODIS feature map by fitting the intrinsic correlation between multiple bands, thereby significantly enhancing the MODIS feature map.

[0022] Furthermore, the multi-channel cross-feature attention module is designed to recalibrate the spatial and spectral weights in the MODIS and Landsat spectral bands;

[0023] The application of the multi-channel cross-feature attention module includes at least the following steps:

[0024] First, features of the Landsat and MODIS input images are extracted through layer-by-layer convolution to generate Landsat and MODIS feature maps. Skip connections are then introduced to alleviate the gradient vanishing problem in deep networks while preserving the original information of the input features.

[0025] Secondly, different weights are assigned to the Landsat feature map and the MODIS feature map through an attention mechanism, and spectral weights and spatial weights are generated based on the cross-attention mechanism to adjust the high-resolution and low-resolution features.

[0026] Finally, by element-wise addition, the weighted low-resolution channel features and high-resolution spatial features are fused to achieve preliminary calibration of the MODIS image, which significantly enhances the feature representation capability of the module.

[0027] The calculation formula is as follows:

[0028]

[0029] In the formula: , and These represent the output feature maps of the spatial attention module, the channel attention module, and the cross attention module, respectively. This represents a convolutional layer with skip connections; and These represent low-resolution and high-resolution feature maps, respectively. and These represent convolution and element-wise multiplication operations, respectively. Indicates a cascading operation; It is the sigmoid function; These are weight parameters; Bias parameters for each attention module.

[0030] Furthermore, the residual dense connection module is used to further extract and enhance deep feature information after recalibrating the feature map by employing a residual dense connection module with an embedded attention mechanism;

[0031] The residual dense connection module combines dense connections with an attention mechanism to effectively capture local detail information;

[0032] The residual dense connection module contains six dense connection layers, which transmit features through local skip connections and connect input and output features along the channel dimension to enhance the reusability of feature maps, thereby fully extracting detailed information;

[0033] Secondly, the residual dense connection module embeds a residual attention structure to further enhance the feature map and effectively extract high-level information from the feature map, as shown below:

[0034]

[0035] In the formula, and These represent the input and output feature maps of the residual attention, respectively. and These represent the output feature maps of the channel attention and spatial attention modules, respectively. Represents the weight parameters of the residual attention module; This represents the bias term of the residual attention module.

[0036] Furthermore, the full-feature joint cross-attention module is used to collaboratively process the Landsat feature map and the enhanced MODIS feature map to further fit multi-band information;

[0037] By enhancing the feature weights at the spatial and spectral levels and combining them with a cross-attention mechanism, a recalibrated feature map is obtained, thereby enhancing the spatial detail information of the output image.

[0038] Furthermore, the differential weighting module is used to capture local change information in the time dimension;

[0039] The application of the differential weighting module includes at least the following steps:

[0040] First, weights are learned from the enhanced Landsat feature map and the significantly enhanced MODIS feature map, and batch normalization is used during the learning process to stabilize the training process and accelerate model convergence.

[0041] Secondly, the sigmoid function is used to limit the weight ratio to the range of [0, 1] to generate a spatial weight mask, thereby dynamically assigning weights to each spatial location of the input feature map;

[0042] Finally, the input features are weighted by element-wise multiplication to obtain a preliminary fusion result, which significantly enhances the ability to focus on important regions. The expression is as follows:

[0043]

[0044]

[0045]

[0046] In the formula, and These represent the significantly enhanced MODIS feature map and the enhanced Landsat feature map, respectively. This indicates the preliminary fusion results; This is a feature map of the difference. For convolution kernel; This is a convolution operation; For batch normalization and the sigmoid function; For spatial attention weights.

[0047] Furthermore, the feature enhancement module is used to combine the preliminary fusion result with the enhanced Landsat feature map to obtain a more accurate high-resolution fused image;

[0048] The feature enhancement module includes at least the following steps:

[0049] First, the output feature map of the differential weighting module and the enhanced Landsat feature map are fused through residual connection;

[0050] Secondly, the residual dense connection module is used to fully extract multi-level semantic information;

[0051] Finally, the high-dimensional features are mapped to the number of channels in the original image through convolution operations, generating the final high-resolution fusion result, as shown in the following expression:

[0052]

[0053] In the formula, This indicates the final output result; and These represent the initial fused feature map and the enhanced Landsat feature map, respectively. This is the output feature map of the residual dense connection module; This is a convolution operation; This indicates a convolution operation that adjusts the number of channels to the original input channel number.

[0054] Furthermore, the combined loss function consists of the Charbonnier loss function, the edge loss function, and the structural similarity loss function, and its expression is as follows:

[0055]

[0056]

[0057]

[0058]

[0059] In the formula, , and For regularization parameters, , and The Charbonnier loss function, edge loss function, and structural similarity loss function are respectively determined by the following formulas:

[0060]

[0061] In the formula, Indicates training image pairs Quantity, For label data, This represents the output of the residual fusion network. For constant terms, they are usually set as ; Represents the average structural similarity index; Represents the structural similarity index; Represents the Laplace operator;

[0062] By combining the three loss functions, pixel-level accuracy, edge information, and structural similarity can be optimized simultaneously.

[0063] Furthermore, a dynamic weighting mechanism is employed to enable the DCAF model to adaptively balance the importance of different loss functions, ensuring that each sub-loss function plays its maximum role at different stages of training and improving the robustness of training.

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] This invention, through the design of a DCAF model, utilizes a series of operations—feature extraction, differential weighting, and information enhancement—to fully leverage the complementary information from Landsat and MODIS images, achieving a dual improvement in the temporal continuity and spatial resolution of surface reflectance. The main advantages are as follows:

[0066] 1) The number of input images is reduced, and only two input images are needed to achieve high-quality fusion results, which reduces the difficulty and cost of data acquisition and significantly improves the practicality of the model;

[0067] 2) In terms of feature processing, the model integrates multiple attention mechanisms and deep convolution modules, which can accurately capture the spatial-spectral correlation between low-resolution and high-resolution images, thus demonstrating excellent fusion performance in complex scenes. Attached Figure Description

[0068] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is a schematic diagram of the entire invention. Detailed Implementation

[0070] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0071] This invention utilizes deep learning technology to fuse multi-source satellite data, and can be applied to fields such as environmental monitoring and ecological research.

[0072] Please see Figure 1 The Landsat-MODIS remote sensing image spatiotemporal fusion method based on a two-layer cross-attention mechanism includes at least the following steps:

[0073] S1: Collect satellite remote sensing data from multiple data sources, including at least Landsat surface reflectance data, MOD09GA surface reflectance data, and MOD09GQ surface reflectance data;

[0074] The specific data sources for this embodiment are shown in the table below:

[0075] Enter variable name Data source Spatiotemporal resolution Landsat surface reflectance https: / / earthexplorer.usgs.gov / 30 meters, 16 days MOD09GA Surface Reflectance https: / / lpdaac.usgs.gov / products / mod09gav006 / 500 meters, 1 day MOD09GQ Surface Reflectivity https: / / lpdaac.usgs.gov / products / mod09gqv006 / 250 meters, 1 day

[0076] S2: The data obtained in S1 is preprocessed by resampling all remote sensing images to 30 meters using the nearest neighbor interpolation method; Landsat is selected as a reference to ensure that the image sizes are exactly the same; in addition, the red and near-infrared bands of MOD09GQ are selected to rematch the band sequences with MOD09GA to enhance the spatial details of the fusion results.

[0077] S3: Construct the DCAF model, which consists of a feature extraction module, a difference weighting module, and a feature enhancement module;

[0078] S4: The DCAF model is trained by using a combined loss function. Once the overall loss converges, the trained DCAF model is used to predict surface reflectance data with high spatiotemporal resolution.

[0079] The feature extraction module includes a multi-channel cross-feature attention module, a full-feature joint cross-attention module, and a residual dense connection module with an embedded attention mechanism;

[0080] The application of the feature extraction module includes at least the following steps:

[0081] For the input Landsat reference day image (i.e., a high-resolution image):

[0082] First, feature maps of the Landsat image are initially extracted using convolutional layers;

[0083] Secondly, a residual dense connection module with an embedded attention mechanism is used to enhance the Landsat feature map and further mine the deeper information of the Landsat feature map.

[0084] For the input MODIS daytime image (i.e., a low-resolution image):

[0085] First, a multi-channel cross-feature attention module is used to calculate spatial and spectral weights based on the Landsat input image to achieve preliminary calibration of the MODIS image;

[0086] Secondly, a residual dense connection module with an embedded attention mechanism is used to extract deep features from the pre-calibrated MODIS feature map to obtain an enhanced MODIS feature map.

[0087] Finally, a joint cross-attention module for all features is designed to collaboratively process the Landsat feature map and the enhanced MODIS feature map by fitting the intrinsic correlation between multiple bands, thereby significantly enhancing the MODIS feature map.

[0088] The multi-channel cross-feature attention module is designed to recalibrate the spatial and spectral weights in the MODIS and Landsat spectral bands;

[0089] The application of a multi-channel cross-feature attention module includes at least the following steps:

[0090] First, features of the Landsat and MODIS input images are extracted through layer-by-layer convolution to generate Landsat and MODIS feature maps. Skip connections are then introduced to alleviate the gradient vanishing problem in deep networks while preserving the original information of the input features.

[0091] Secondly, different weights are assigned to the Landsat feature map and the MODIS feature map through an attention mechanism, and spectral weights and spatial weights are generated based on the cross-attention mechanism to adjust the high-resolution and low-resolution features.

[0092] Finally, by element-wise addition, the weighted low-resolution channel features and high-resolution spatial features are fused to achieve preliminary calibration of the MODIS image, which significantly enhances the feature representation capability of the module.

[0093] The calculation formula is as follows:

[0094]

[0095] In the formula: , and These represent the output feature maps of the spatial attention module, the channel attention module, and the cross attention module, respectively. This represents a convolutional layer with skip connections; and These represent low-resolution and high-resolution feature maps, respectively. and These represent convolution and element-wise multiplication operations, respectively. Indicates a cascading operation; It is the sigmoid function; These are weight parameters; Bias parameters for each attention module.

[0096] The residual dense connection module is used to further extract and enhance deep feature information after the feature map is recalibrated, using a residual dense connection module with an embedded attention mechanism.

[0097] The residual dense connection module combines dense connections with attention mechanisms to effectively capture local detail information;

[0098] The residual dense connection module contains six dense connection layers. It passes features through local skip connections and connects input and output features along the channel dimension, which enhances the reusability of feature maps and thus fully extracts detailed information.

[0099] Secondly, the residual dense connection module embeds a residual attention structure to further enhance the feature map and effectively extract high-level information from the feature map, as shown below:

[0100]

[0101] In the formula, and These represent the input and output feature maps of the residual attention, respectively. and These represent the output feature maps of the channel attention and spatial attention modules, respectively. Represents the weight parameters of the residual attention module; This represents the bias term of the residual attention module.

[0102] The full-feature joint cross-attention module is used to collaboratively process the Landsat feature map and the enhanced MODIS feature map to further fit multi-band information;

[0103] By enhancing the feature weights at the spatial and spectral levels and combining them with a cross-attention mechanism, a recalibrated feature map is obtained, thereby enhancing the spatial detail information of the output image.

[0104] The differential weighting module is used to capture local change information in the time dimension;

[0105] The application of the differential weighting module includes at least the following steps:

[0106] First, weights are learned from the enhanced Landsat feature map and the significantly enhanced MODIS feature map, and batch normalization is used during the learning process to stabilize the training process and accelerate model convergence.

[0107] Secondly, the sigmoid function is used to limit the weight ratio to the range of [0, 1] to generate a spatial weight mask, thereby dynamically assigning weights to each spatial location of the input feature map;

[0108] Finally, the input features are weighted by element-wise multiplication to obtain a preliminary fusion result, which significantly enhances the ability to focus on important regions. The expression is as follows:

[0109]

[0110]

[0111]

[0112] In the formula, and These represent the significantly enhanced MODIS feature map and the enhanced Landsat feature map, respectively. This indicates the preliminary fusion results; This is a feature map of the difference. For convolution kernel; This is a convolution operation; For batch normalization and the sigmoid function; Spatial attention weights;

[0113] In summary, the differential weighting module can capture significant differences between input feature maps, thereby extracting and fusing key information from the input data.

[0114] The feature enhancement module is used to combine the preliminary fusion result with the enhanced Landsat feature map to obtain a more accurate high-resolution fused image;

[0115] The feature enhancement module includes at least the following steps:

[0116] First, the output feature map of the differential weighting module and the enhanced Landsat feature map are fused through residual connection;

[0117] Secondly, the residual dense connection module is used to fully extract multi-level semantic information;

[0118] Finally, the high-dimensional features are mapped to the number of channels in the original image through convolution operations, generating the final high-resolution fusion result, as shown in the following expression:

[0119]

[0120] In the formula, This indicates the final output result; and These represent the initial fused feature map and the enhanced Landsat feature map, respectively. This is the output feature map of the residual dense connection module; This is a convolution operation; This indicates a convolution operation that adjusts the number of channels to the original input channel number.

[0121] The feature enhancement module significantly enhances the spatial details and semantic information of the initial fused image.

[0122] The combined loss function consists of the Charbonnier loss function, the marginal loss function, and the structural similarity loss function, and its expression is as follows:

[0123]

[0124]

[0125]

[0126]

[0127] In the formula, , and For regularization parameters, , and The Charbonnier loss function, edge loss function, and structural similarity loss function are respectively determined by the following formulas:

[0128]

[0129] In the formula, Indicates training image pairs Quantity, For label data, This represents the output of the residual fusion network. For constant terms, they are usually set as ; Represents the average structural similarity index; Represents the structural similarity index; Represents the Laplace operator;

[0130] By combining the three loss functions, pixel-level accuracy, edge information, and structural similarity can be optimized simultaneously.

[0131] Furthermore, a dynamic weighting mechanism is employed to enable the DCAF model to adaptively balance the importance of different loss functions, ensuring that each sub-loss function plays its maximum role at different stages of training and improving the robustness of training.

[0132] In summary:

[0133] This invention addresses the limitations of image data quality in spatiotemporal fusion by designing a dual cross-attention based spatiotemporal fusion (DCAF) method based on cross-paired images. This method fully leverages the complex characteristics of multi-source data to generate high-precision fused data. First, using a fine-resolution image of the reference date and a coarse-resolution image of the predicted date, a multi-channel cross-feature attention module is used to calibrate the spatial and spectral feature weights of Moderate-resolution Imaging Spectroradiometer (MODIS) and Landsat images band-by-band. Second, a full-feature joint cross-attention structure is introduced to establish the intrinsic relationships between bands. Subsequently, a residual dense connection structure with an embedded attention mechanism is used to extract deep features from the enhanced MODIS and Landsat feature maps, further optimizing feature representation capabilities. Furthermore, a differential weighting module is used to adaptively allocate weights to the MODIS and Landsat feature maps, thereby learning temporal variation information. Finally, a combined loss function is designed to constrain the network training process, including Charbonnier loss, edge loss, and structural similarity loss. This method can fully utilize the complementary spatiotemporal information of Landsat and MODIS remote sensing images to generate fused images with high spatial and temporal resolution, providing technical support for the spatiotemporal fine-grained monitoring of Earth's surface elements.

[0134] Compared to existing spatiotemporal fusion algorithms such as the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM), Cross-Attention Based Adaptive Weighting Fusion Network (CAFE), and Enhanced Cross-paired Wavelet based Spatiotemporal Fusion Networks (ECPW-STFN), the DCAF model can fully utilize the spatiotemporal advantages of Landsat and MODIS remote sensing images to achieve high spatiotemporal resolution surface reflectance data fusion. Under different surface conditions, the DCAF method exhibits superior robustness and adaptability compared to traditional methods. Furthermore, the DCAF model proposed in this invention effectively expands the potential application scenarios of spatiotemporal fusion technology, providing high-quality data support for Earth environmental monitoring.

[0135] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A Landsat-MODIS remote sensing image spatio-temporal fusion method based on a double-layer cross-attention mechanism, characterized in that: At least comprising the following steps: S1: Collecting satellite remote sensing data from multiple data sources, the satellite remote sensing data at least comprising Landsat surface reflectance data, MOD09GA surface reflectance data and MOD09GQ surface reflectance data; S2: Preprocessing the data obtained in S1, using the nearest neighbor interpolation method to resample all remote sensing images to 30 meters; selecting Landsat as the reference to ensure that the image sizes are exactly the same; in addition, re-matching the red and near-infrared bands of MOD09GQ with MOD09GA to enhance the spatial details of the fusion results; S3: Building a DCAF model, the DCAF model comprising a feature extraction module, a differential weighting module and a feature enhancement module; The feature extraction module comprises a multi-channel cross-feature attention module, a full-feature joint cross-attention module and a residual dense connection module with embedded attention mechanism; The application of the feature extraction module at least comprises the following steps: For the input Landsat reference daily image, i.e. the high-resolution image: Firstly, using a convolutional layer to preliminarily extract the feature map of the Landsat image; Secondly, using the residual dense connection module with embedded attention mechanism to enhance the Landsat feature map, further mining the deep information of the Landsat feature map; For the input MODIS daily image, i.e. the low-resolution image: Firstly, using the multi-channel cross-feature attention module to calculate the spatial and spectral weights based on the Landsat input image, realizing the preliminary calibration of the MODIS image; Secondly, using the residual dense connection module with embedded attention mechanism to perform deep feature extraction on the preliminarily calibrated MODIS feature map, obtaining an enhanced MODIS feature map; Finally, designing a full-feature joint cross-attention module to fit the internal correlation among multiple bands, cooperatively processing the Landsat feature map and the enhanced MODIS feature map, and significantly enhancing the MODIS feature map; The differential weighting module is used to capture the local change information in the time dimension; The application of the differential weighting module at least comprises the following steps: Firstly, learning the weights from the enhanced Landsat feature map and the significantly enhanced MODIS feature map, and using batch normalization in the learning process to stabilize the training process and accelerate model convergence; Secondly, using a sigmoid function to limit the weight ratio to the range of [0, 1] to generate a spatial weight mask, thereby dynamically assigning weights to each spatial position of the input feature map; Finally, weighting the input features by element-wise multiplication to obtain a preliminary fusion result, significantly enhancing the attention ability to important areas, and the expression is as follows: wherein, and represent the significantly enhanced MODIS feature map and the enhanced Landsat feature map, respectively; represents the preliminary fusion result; is a difference feature map; is a convolution kernel; is a convolution operation; is a batch normalization and sigmoid function; is a spatial attention weight; The feature enhancement module is used to combine the preliminary fusion result with the enhanced Landsat feature map to obtain a more accurate high-resolution fusion image; The feature enhancement module at least comprises the following steps: Firstly, fusing the output feature map of the differential weighting module with the enhanced Landsat feature map through residual connection; Secondly, using the residual dense connection module to fully extract multi-level semantic information; Finally, the high-dimensional features are mapped to the channel number of the original image through convolution operation to generate the final high-resolution fusion result, which is expressed as follows: In the formula, represents the final output result; and respectively represent the preliminary fusion feature map and the enhanced Landsat feature map; is the output feature map of the residual dense connection module; is a convolution operation; represents a convolution operation for adjusting the number of channels to the original input channel number; S4: The DCAF model is trained by using a combined loss function, and when the overall loss converges, the trained DCAF model is used to predict high-spatial and temporal resolution surface reflectance data; The combined loss function is composed of a Charbonnier loss function, an edge loss function and a structural similarity loss function, and is expressed as follows: wherein , and are regularization parameters, , and are the Charbonnier loss function, the edge loss function and the structural similarity loss function, respectively, determined by the following calculation formulas: In the formula, denotes the number of training image pairs , is a MODIS feature map, is a Landsat feature map, is label data, denotes the output of the residual fusion network, is a constant term, usually set to ; denotes the average structural similarity index; denotes the structural similarity index; denotes the Laplacian operator.

2. The Landsat-MODIS remote sensing image spatio-temporal fusion method based on a double-layer cross-attention mechanism according to claim 1, characterized in that: The multi-channel cross-feature attention module aims to recalibrate the spatial and spectral weights of MODIS and Landsat spectral bands. The application of the multi-channel cross-feature attention module at least includes the following steps: First, the features of Landsat input images and MODIS input images are extracted through layer-by-layer convolution to generate Landsat feature maps and MODIS feature maps, and a skip connection is introduced to retain the original information of the input features while alleviating the gradient vanishing problem in the deep network; Second, the Landsat feature maps and MODIS feature maps are given different weights through attention mechanism, and spectral weights and spatial weights are generated based on cross-attention mechanism to adjust the high-resolution and low-resolution features; Finally, the weighted low-resolution channel features and high-resolution spatial features are fused through element addition to realize the preliminary calibration of MODIS images, which significantly enhances the feature representation capability of the module; The calculation formula is as follows: In the formula: , and respectively represent the output feature maps of the spatial attention module, the channel attention module and the cross attention module; represents a layer-by-layer convolution layer of a skip connection; and respectively represent low-resolution and high-resolution feature maps; and respectively represent convolution operations and element multiplication operations; represents a concatenation operation; is a sigmoid function; is a weight parameter; is a bias parameter of each attention module.

3. The Landsat-MODIS remote sensing image spatio-temporal fusion method based on a double-layer cross-attention mechanism according to claim 2, characterized in that: The residual dense connection module is used to further extract and enhance deep feature information after recalibrating the feature maps by using a residual dense connection module with embedded attention mechanism; The residual dense connection module combines dense connection and attention mechanism to effectively capture local detail information. The residual dense connection module contains six dense connection layers, which pass features through local skip connection and connect input and output features along the channel dimension to enhance the reusability of the feature maps, thereby fully extracting detailed information. Second, the embedded residual attention structure of the residual dense connection module further enhances the feature maps and effectively extracts high-level information in the feature maps, which is expressed as follows: wherein, and denote the input and output feature maps of the residual attention, respectively; and denote the output feature maps of the channel attention and spatial attention modules, respectively; denote the weight parameters of the residual attention module; denote the bias terms of the residual attention module.

4. The Landsat-MODIS remote sensing image spatio-temporal fusion method based on a double-layer cross-attention mechanism according to claim 3, characterized in that: The full-feature joint cross-attention module is used to cooperatively process Landsat feature maps and enhanced MODIS feature maps to further fit multi-band information. By enhancing the feature weights in the spatial and spectral dimensions and combining the cross-attention mechanism, the recalibrated feature maps are obtained, and the spatial detail information of the output image is further enhanced.

Images