High-resolution remote sensing image accurate classification method based on image frequency domain features

Abstract

This invention relates to the field of image classification technology, specifically to a method for accurate classification of high-resolution remote sensing images based on image frequency domain features. The method includes: acquiring high-resolution remote sensing image data; building a classification model using an encoder-decoder architecture, with a multi-scale feature enhancement module connected after each output scale of the backbone network; the decoder employing a U-Net structure, using bilinear interpolation for step-by-step upsampling, and concatenating the enhanced feature maps of the corresponding scales with the upsampled features through skip connections to output a prediction result consistent with the input image size; setting a dual-segmentation head to assist in the training of the classification model, obtaining a trained classification model; and using the trained classification model to obtain the classification result from the high-resolution remote sensing image data to be classified. This invention incorporates a frequency domain modeling layer into the neural network to address classification and recognition errors caused by the lack of frequency domain features.

Description

Technical Field

[0001] This invention relates to the field of image classification technology, and specifically to a method for accurate classification of high-resolution remote sensing images based on image frequency domain features. Background Technology

[0002] With the rapid development of remote sensing technology and unmanned aerial vehicle (UAV) platforms, acquiring high-resolution remote sensing imagery has become increasingly convenient, with spatial resolution reaching sub-meter or even centimeter levels. This provides a rich data foundation for detailed identification of ground features and extraction of surface information. Compared to medium- and low-resolution imagery, high-resolution imagery can clearly reflect the shape, texture, boundaries, and spatial structure characteristics of ground features. However, correspondingly, the processing methods for high-resolution imagery are more complex than those for medium- and low-resolution imagery.

[0003] Currently, there are some problems with the classification of high-resolution, especially ultra-high-resolution, remote sensing images. High-resolution images have very rich texture information, which easily leads to local repetition and high-frequency noise at the edges. Existing remote sensing image classification methods are mainly based on image features in the spectral, spatial, and temporal domains. However, frequency domain features are also fundamental features of digital images. Some researchers have used Fourier transform to transform the spectral signal of each pixel in both the spectral and frequency domains, extracting its amplitude spectrum as frequency domain features. However, Fourier transform is a global operation, lacking spatial local perception capability, and its transformation parameters are fixed, making it unlearnable and difficult to adapt to different task characteristics. Summary of the Invention

[0004] To address the problems existing in the background technology, the present invention provides a high-resolution remote sensing image accurate classification method based on image frequency domain features. This method incorporates a frequency domain modeling layer into a neural network to solve the classification and recognition errors caused by the lack of frequency domain features.

[0005] The first objective of this invention is to provide a high-resolution remote sensing image accurate classification method based on image frequency domain features, comprising:

[0006] Acquire high-resolution remote sensing image data, normalize the image data, slice the data with a fixed window size and retain the overlapping area of ​​adjacent slices, and at the same time create an integer mask label corresponding to the image so that the label is aligned with the image through nearest neighbor interpolation.

[0007] A classification model is built using an encoder-decoder architecture. The encoder uses a pre-trained ResNet-50 network as the backbone network and extracts feature maps at different scales through four-stage convolution and downsampling operations. A multi-scale feature enhancement module is connected after each output scale of the backbone network to enhance the feature maps at each scale and obtain enhanced feature maps at each scale. The decoder adopts a U-Net structure and upsamples step by step through bilinear interpolation. The enhanced feature maps at the corresponding scale are concatenated and fused with the upsampled features through skip connections to output a prediction result with the same size as the input image.

[0008] Set up a dual-segmentation head to assist in the training of the classification model and obtain the trained classification model;

[0009] The high-resolution remote sensing image data to be classified is processed by a trained classification model to obtain the classification result;

[0010] The multi-scale feature enhancement module includes two parallel large kernel convolution branches and a frequency domain interaction branch; the feature maps of the backbone network are processed by the large kernel convolution branches and the frequency domain interaction branches respectively, and then spliced ​​and fused to obtain the enhanced feature maps.

[0011] In one embodiment, training a dual-segmentation-head-assisted classification model includes:

[0012] The first segmentation head calculates the first cross-entropy loss on the feature map after skip connections and multiple upsampling. The second segmentation head calculates the second cross-entropy loss for the deep fusion features. By controlling the proportion of the second cross-entropy loss with a weight coefficient of 0.6, the final loss function is obtained as follows: .

[0013] In one embodiment, the formula for calculating the first cross-entropy loss or the second cross-entropy loss is as follows:

[0014]

[0015] In the formula, The actual results of image classification; The predicted results output by the model; Image height; Image width; Number of image categories; To predict the first in the image i Line number j Column corresponding to the first The pixel value of the class; For the first in the tagged image i Line number j Column corresponding to the first The pixel value of the class.

[0016] In one embodiment, the large kernel convolution branch includes:

[0017] The feature maps of the backbone network are processed through two large-kernel convolutional layers of 5×5 and 7×7 to extract multi-scale spatial features. These two features are then concatenated along the channel dimension and subjected to max pooling and average pooling, respectively.

[0018] The pooled features are then passed through a 7×7 convolution to generate two spatial attention masks;

[0019] Two spatial attention masks are applied to the corresponding 5×5 and 7×7 convolution outputs, respectively. Spatial weights are then enhanced by element-wise multiplication, and the enhanced features are then fused for output.

[0020] In one embodiment, the frequency domain interaction branch includes:

[0021] High-frequency features are extracted from the feature maps of the backbone network using 1×1 convolution, while low-frequency features are obtained by combining average pooling downsampling with 1×1 convolution.

[0022] The high-frequency and low-frequency features are combined and fused using octave convolution to output the result.

[0023] In one embodiment, when acquiring high-resolution remote sensing image data, high-resolution remote sensing images containing three bands are included.

[0024] In one embodiment, extracting feature maps at different scales includes extracting feature maps at scales of 1 / 2, 1 / 4, 1 / 8, and 1 / 16.

[0025] A second objective of this invention is to provide a system for accurate classification of high-resolution remote sensing images based on image frequency domain features, comprising:

[0026] The data acquisition module is used to acquire high-resolution remote sensing image data, normalize the image data, slice it with a fixed window size and retain the overlapping area of ​​adjacent slices, and at the same time create an integer mask label corresponding to the image so that the label is aligned with the image through nearest neighbor interpolation.

[0027] The model building module is used to build a classification model using an encoder-decoder architecture. The encoder uses a pre-trained ResNet-50 network as the backbone network, extracting feature maps at different scales through four-stage convolution and downsampling operations. A multi-scale feature enhancement module is connected after each output scale of the backbone network to enhance the feature maps at each scale, obtaining enhanced feature maps at each scale. The decoder adopts a U-Net structure, upsampling step by step through bilinear interpolation, and concatenating and fusing the enhanced feature maps at the corresponding scales with the upsampled features through skip connections, outputting a prediction result with the same size as the input image. A dual segmentation head is set to assist in the training of the classification model, obtaining a trained classification model.

[0028] The classification module is used to obtain classification results from the high-resolution remote sensing image data to be classified using a trained classification model.

[0029] The present invention has at least the following beneficial effects:

[0030] This invention provides a high-resolution remote sensing image classification method based on image frequency domain features. The method utilizes a multi-scale feature enhancement module (MFEM) to process the four-layer features output from the ResNet-50 backbone network. This module progressively upsamples the lower-level features to match the original feature scale, and after each sampling, skip connections are used to compensate for spatial detail lost during sampling. The feature maps extracted from the backbone network are further enhanced using the MFEM module. Large-kernel convolution branches compensate for insufficient contextual information caused by a small receptive field, while frequency domain interaction branches enhance the network's perception of frequency domain features, ultimately improving the network's classification accuracy. Attached Figure Description

[0031] Figure 1 The flowchart shows the high-resolution remote sensing image accurate classification method based on image frequency domain features provided by the present invention.

[0032] Figure 2 This is the dataset used for model training.

[0033] Figure 3 This is a schematic diagram of the MFEM module.

[0034] Figure 4 This is a schematic diagram of the network model.

[0035] Figure 5 A visual comparison chart showing the recognition performance of different models. Detailed Implementation

[0036] In order to illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description is provided in conjunction with the embodiments.

[0037] This invention addresses the issue that when Fourier transform transforms the spectral signal of each pixel in both the "spectral domain" and the "frequency domain" to extract its amplitude spectrum as a frequency domain feature, the Fourier transform is a global operation, lacking spatial local perception capability, and its transformation parameters are fixed, making it unlearnable and difficult to adapt to different task features. This invention addresses this by designing a dedicated frequency domain modeling layer or frequency domain interaction module within the neural network, enabling the network to automatically learn the importance and fusion method of frequency domain features during training. Based on these considerations, the inventors incorporate the frequency domain feature information of high-resolution images into the interpretation labels for classification, achieving the model's capture of frequency domain information through the interaction of feature maps of different resolutions within the deep learning model.

[0038] The present invention aims to provide a high-resolution remote sensing image classification method based on image frequency domain features. This invention utilizes a deep learning algorithm that takes frequency domain features into account. Its structure is based on an encoder-decoder architecture, with a ResNet-50 backbone network capable of extracting rich features from UAV imagery. Considering the ultra-high spatial resolution of UAV imagery, U-Net skip connections are incorporated into the decoder to fully preserve spatial detail features.

[0039] To achieve the above objectives, see Figure 1 As shown, a high-resolution remote sensing image accurate classification method based on image frequency domain features includes:

[0040] S1. Acquire high-resolution remote sensing image data, normalize the image data, slice the data with a fixed window size and retain the overlapping area of ​​adjacent slices, and at the same time create an integer mask label corresponding to the image so that the label is aligned with the image through nearest neighbor interpolation.

[0041] When acquiring high-resolution remote sensing image data, the high-resolution remote sensing image includes three bands. This invention slices the image using a fixed window size of 256×256 pixels.

[0042] S2. A classification model is built using an encoder-decoder architecture. The encoder uses a pre-trained ResNet-50 network as the backbone network. It extracts feature maps at different scales through four-stage convolution and downsampling operations. A multi-scale feature enhancement module is connected after each output scale of the backbone network to enhance the feature maps at each scale and obtain enhanced feature maps at each scale. The decoder adopts a U-Net structure and upsamples step by step through bilinear interpolation. It concatenates and fuses the enhanced feature maps at the corresponding scales with the upsampled features through skip connections to output a prediction result with the same size as the input image.

[0043] When extracting feature maps at different scales, including feature maps at scales of 1 / 2, 1 / 4, 1 / 8, and 1 / 16.

[0044] The multi-scale feature enhancement module includes two parallel large kernel convolution branches and a frequency domain interaction branch; the feature maps of the backbone network are processed by the large kernel convolution branches and the frequency domain interaction branches respectively, and then spliced ​​and fused to obtain the enhanced feature maps.

[0045] The large kernel convolution branch includes:

[0046] The feature maps of the backbone network are processed through two large-kernel convolutional layers of 5×5 and 7×7 to extract multi-scale spatial features. These two features are then concatenated along the channel dimension and subjected to max pooling and average pooling, respectively.

[0047] The pooled features are then passed through a 7×7 convolution to generate two spatial attention masks;

[0048] Two spatial attention masks are applied to the corresponding 5×5 and 7×7 convolution outputs, respectively. Spatial weights are then enhanced by element-wise multiplication, and the enhanced features are then fused for output.

[0049] The frequency domain interaction branch includes:

[0050] High-frequency features are extracted from the feature maps of the backbone network using 1×1 convolution, while low-frequency features are obtained by combining average pooling downsampling with 1×1 convolution.

[0051] The high-frequency and low-frequency features are combined and fused using octave convolution to output the result.

[0052] S3. Set up dual-segmentation head auxiliary classification model training to obtain the trained classification model;

[0053] Setting up a dual-segmentation head-assisted classification model for training includes:

[0054] The first segmentation head calculates the first cross-entropy loss on the feature map after skip connections and multiple upsampling. The second segmentation head calculates the second cross-entropy loss for the deep fusion features. By controlling the proportion of the second cross-entropy loss with a weight coefficient of 0.6, the final loss function is obtained as follows: .

[0055] The formulas for calculating the first cross-entropy loss or the second cross-entropy loss are as follows:

[0056]

[0057] In the formula, The actual results of image classification; The predicted results output by the model; Image height; Image width; Number of image categories; To predict the first in the image i Line number j Column corresponding to the first The pixel value of the class; For the first in the tagged image i Line number j Column corresponding to the first The pixel value of the class. It should be noted that, corresponding to the... In semantic segmentation tasks, each pixel needs to be classified into one of L categories, such as "Suaeda salsa, weeds, Spartina alterniflora, water, and reeds", where L=5.

[0058] S4. The high-resolution remote sensing image data to be classified is processed by a trained classification model to obtain the classification results.

[0059] To illustrate the high-resolution remote sensing image accurate classification method based on image frequency domain features provided by this invention, specific examples are used for explanation.

[0060] This embodiment is implemented using the Python programming language, and uses PyTorch 2.5.1 and CUDA 12.6 to build and train the deep learning algorithm model. It mainly includes five stages: data preprocessing, feature encoding, multi-scale feature enhancement, feature decoding, and result output.

[0061] During the data processing phase, the input data consists of multispectral or high-resolution images acquired by UAVs, with each image containing three bands. To ensure the stability of model training, the image data is first normalized and then sliced ​​into fixed window sizes of 256×256 pixels, while retaining some overlapping areas between adjacent slices to avoid boundary effects. The label data is an integer mask image, aligned with the image using nearest-neighbor interpolation.

[0062] In this embodiment, the aerial photography location selected was a wetland in the coastal area of ​​Dongying City, Shandong Province, China. It should be noted that Dongying City is located at the mouth of the Yellow River, a typical estuary delta wetland region. The Dongying wetland exhibits typical zonal distribution characteristics of estuarine wetland vegetation. Among them, *Suaeda salsa* mainly grows in low-lying saline-alkali land in the intertidal zone and is the most representative halophyte community in the Yellow River estuary wetlands; *Phragmites australis* is widely distributed in freshwater and brackish water wetlands and is a dominant native species in the region; while the invasive alien species *Spartina alterniflora* has rapidly expanded along the coastal zone, forming a complex competitive and symbiotic pattern with *Phragmites australis* in some areas. The final classification of species includes *Suaeda salsa*, *Spartina alterniflora*, weeds, *Phragmites australis*, water bodies, and others.

[0063] The data used in this embodiment is orthophotos acquired by an airborne drone manufactured by Beijing Pegasus Robotics in China. The spatial resolution is 0.02m, and the images were taken with an RGB camera. After preprocessing such as stitching and mosaicking, the map projection is WGS-84. The modeling dataset is shown in Figure 2. The selected data is sufficient to cover all vegetation types in the entire aerial photography location. Labels were drawn using manual visual interpretation. Finally, 160 image-label pairs of 256×256 pixels were created, and the data was divided into training and test sets at a 9:1 ratio. Both images and labels were saved in TIFF format.

[0064] The model employs an encoder-decoder architecture. The encoder uses a ResNet-50 network pre-trained on ImageNet as its backbone, extracting multi-scale feature maps through four stages of convolution and downsampling operations at scales of 1 / 2, 1 / 4, 1 / 8, and 1 / 16. A multi-scale feature enhancement module (MSFE) is appended after each output scale of the backbone network. This embodiment introduces a multi-scale feature enhancement module to enhance the features of each layer of the backbone network, including large-kernel convolution and frequency domain interaction of features at different resolutions.

[0065] See Figure 3 As shown, the multi-scale feature enhancement module enhances the feature maps extracted by the backbone network. This module includes two branches: a large kernel convolution branch to compensate for insufficient contextual information caused by a small receptive field, and a frequency domain interaction branch to enhance the network's perception of frequency domain features, ultimately improving the network's classification accuracy.

[0066] The large-kernel convolution branch expands the receptive field to capture richer spatial relationships, enhancing the model's contextual modeling capabilities. Specifically, feature maps from the backbone network are first processed through two large-kernel convolutional layers (5×5 and 7×7) to extract multi-scale spatial features. These two feature paths are then concatenated along the channel dimension and subjected to max pooling and average pooling respectively to obtain statistical features at two scales. The pooled features are then passed through a 7×7 convolution to generate two spatial attention masks. Finally, these two attention masks are applied to the corresponding 5×5 and 7×7 convolution outputs, respectively, achieving spatial weight enhancement through element-wise multiplication. The enhanced features are then fused and output. Furthermore, to preserve the high spatial resolution details of the UAV imagery, this branch maintains the same feature map size throughout the process. Its mathematical process can be expressed as:

[0067]

[0068] In the formula, This represents the output of the large kernel convolution branch. This represents a feature map derived from the backbone network; This represents the spatial attention mask obtained through average pooling. This represents the spatial attention mask obtained through max pooling; This represents a 5×5 convolutional layer; This represents a 7×7 convolutional layer.

[0069] The frequency domain interaction branch aims to enhance the network's representational capabilities from a frequency domain perspective. This branch also takes the feature map output from the backbone network as input and achieves multi-frequency feature modeling through explicit high-frequency and low-frequency decomposition. Specifically, a 1×1 convolution that maintains spatial resolution is first used to extract high-frequency feature components; simultaneously, average pooling downsampling combined with a 1×1 convolution is used to obtain the corresponding low-frequency features, thus achieving frequency domain partitioning of the input features. Based on this, an octave convolution is used to perform bidirectional interaction and fusion of high-frequency and low-frequency features. As a network layer with learnable parameters, the octave convolution can adaptively learn the weight allocation and fusion strategies for different frequency components during training, achieving more effective frequency domain information integration. The overall mathematical process of this frequency domain interaction branch can be expressed as:

[0070]

[0071] In the formula, This indicates the output of the frequency domain interactive branch; This represents a feature map derived from the backbone network; Indicates an octave convolutional layer; This represents a 1×1 convolutional layer; This indicates downsampling.

[0072] Finally, the comprehensive features extracted by the multi-scale feature enhancement module can be represented as:

[0073]

[0074]

[0075] In the formula, This represents the characteristics of each encoder layer; Indicates the first i The output of the large kernel convolution branch of the layer; Indicates the first i The output of the frequency domain interactive branch of the layer; Indicates the first 0 Features of the layer encoder; Indicates the first 0 The output of the large kernel convolution branch of the layer; Indicates the first 0 The output of the frequency domain interactive branch of the layer.

[0076] The advantage of this invention by introducing a multi-scale feature enhancement module is that it enables the model to capture frequency domain features through the interaction of feature maps at different resolutions, thereby enhancing the model's recognition performance.

[0077] The decoder uses a U-Net structure, which performs stepwise upsampling through bilinear interpolation, and then concatenates and fuses the enhanced feature maps of the corresponding scale with the upsampled features through skip connections, finally outputting a prediction result with the same size as the input image.

[0078] To avoid errors in recognition results caused by the network over-focusing on shallow features, this invention employs a cross-entropy loss function to calculate model error and sets up a dual-segmentation head to assist in training the classification model. The first segmentation head calculates the cross-entropy loss on the feature maps after skip connections and all upsampling, denoted as... The cross-entropy loss is calculated using the feature map after skip connections and one upsampling of the second segmentation head. That is, calculating the cross-entropy loss for deep fusion features. Then, a weighting coefficient is used to control it. The proportion of [amount] is 0.6 in this invention. The final loss function is:

[0079]

[0080] The cross-entropy loss function used is calculated using the following formula:

[0081]

[0082] In this invention, during the training of the classification model, the optimizer used is Adam, with an initial learning rate set to 0.01, and a CosineAnnealingLR learning rate scheduling strategy is employed. During model training, the Adam optimizer first calculates the gradient of the current parameters through backpropagation, and then performs an exponentially weighted average of the first moment (mean) and second moment (squared mean) of the gradient, thereby obtaining a more stable and adaptive gradient estimate. CosineAnnealingLR is a cosine function-based learning rate scheduling strategy that gradually decreases the learning rate according to a cosine curve during training.

[0083] During the training of the classification model, the normalized image is first input into the model, and the model will output two integer single-channel prediction results. The cross-entropy loss is calculated between these two results and the corresponding true classification results of the image. Then, the parameters in the model are updated through the backpropagation process. This process is repeated 100 times to finally obtain a model for high-resolution image classification.

[0084] In subsequent prediction, the image data to be classified is predicted using the same window slicing strategy as during training. The predictions involving overlapping areas are then fused using a weighted average. The output is a floating-point probability map or an integer category mask, saved in GeoTIFF format to retain spatial reference information for subsequent georeferencing and analysis.

[0085] See Figure 4 As shown, the four-layer features output by the backbone network are processed by a multi-scale feature enhancement module (MFEM) to progressively upsample the bottom-level features to the same scale as the original features. After each sampling, skip connections are used to compensate for the spatial detail information lost due to sampling.

[0086] To illustrate the performance of the high-resolution remote sensing image classification method based on image frequency domain features provided in this invention, six commonly used deep learning models for natural image semantic segmentation and remote sensing image classification tasks were selected for comparison. The recognition results are shown in [link to relevant documentation]. Figure 5 As shown.

[0087] To objectively evaluate the model's performance, as shown in Table 1, four metrics were selected to evaluate the model's accuracy: accuracy, Kappa coefficient, recall, and semantic segmentation accuracy (mIoU). Based on the visualization results and accuracy evaluation results, it can be seen that the present invention has excellent performance.

[0088] Among them, Table 1 and Figure 5 In this document, Our is the model of the present invention; Label, Unet, Transunet, Linknet, Dplabv3, MCCA, and Segformer are all existing models; Image is a test image.

[0089] Table 1 Accuracy results for all models

[0090]

[0091] From Table 1 and Figure 5 As can be seen, this invention demonstrates superior performance, exhibiting a clear advantage over other mainstream deep learning methods in accuracy, Kappa, and recall. Furthermore, Transunet, which combines transform and CNN, also demonstrates good performance, dominating in mIoU. Specifically, the method provided in this invention outperforms Transunet by 0.52% in accuracy, 0.79% in Kappa, and 0.52% in mIoU.

[0092] This invention provides a system for accurate classification of high-resolution remote sensing images based on image frequency domain features, comprising:

[0093] The data acquisition module is used to acquire high-resolution remote sensing image data, normalize the image data, slice it with a fixed window size and retain the overlapping area of ​​adjacent slices, and at the same time create an integer mask label corresponding to the image so that the label is aligned with the image through nearest neighbor interpolation.

[0094] The model building module is used to build a classification model using an encoder-decoder architecture. The encoder uses a pre-trained ResNet-50 network as the backbone network, extracting feature maps at different scales through four-stage convolution and downsampling operations. A multi-scale feature enhancement module is connected after each output scale of the backbone network to enhance the feature maps at each scale, obtaining enhanced feature maps at each scale. The decoder adopts a U-Net structure, upsampling step by step through bilinear interpolation, and concatenating and fusing the enhanced feature maps at the corresponding scales with the upsampled features through skip connections, outputting a prediction result with the same size as the input image. A dual segmentation head is set to assist in the training of the classification model, obtaining a trained classification model.

[0095] The classification module is used to obtain classification results from the high-resolution remote sensing image data to be classified using a trained classification model.

[0096] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An image frequency domain feature-based high-resolution remote sensing image accurate classification method, characterized in that, The application relates to a high-resolution remote sensing image classification method and device. The application comprises the following steps: An encoder-decoder architecture is adopted to build a classification model, the encoder adopts a pre-trained ResNet-50 network as a backbone network, different scale feature maps are extracted through four-stage convolution and down-sampling operations, and a multi-scale feature enhancement module is connected after each output scale of the backbone network to enhance the features of the feature maps of different scales and obtain enhanced feature maps of different scales; the decoder adopts a U-Net structure, the enhanced feature maps of corresponding scales are fused through jump connection and up-sampling, and the prediction result consistent with the input image size is outputted; A double-segmentation-head auxiliary classification model training is set, and a trained classification model is obtained; High-resolution remote sensing image data to be classified is subjected to the trained classification model to obtain a classification result; The multi-scale feature enhancement module comprises two parallel large-kernel convolution branches and a frequency domain interaction branch; the feature maps of the backbone network are processed through the large-kernel convolution branch and the frequency domain interaction branch, and then the enhanced feature maps are obtained through splicing and fusion; The large-kernel convolution branch comprises the following steps: The feature maps of the backbone network are extracted through 5*5 and 7*7 large-kernel convolution layers to obtain multi-scale spatial features, and then the two features are spliced along the channel dimension and subjected to maximum pooling and average pooling respectively; The pooled features are subjected to a 7*7 convolution to generate two spatial attention masks; The two spatial attention masks are respectively applied to the outputs of the corresponding 5*5 and 7*7 convolutions, and then the spatial weight enhancement is carried out through element-by-element multiplication, and then the enhanced features are fused and outputted; The frequency domain interaction branch comprises the following steps: The feature maps of the backbone network are extracted through 1*1 convolution to obtain high-frequency features, and the corresponding low-frequency features are obtained through average pooling down-sampling and 1*1 convolution; The high-frequency features and the low-frequency features are outputted through bidirectional interaction and fusion through an eight-degree convolution.

2. The method according to claim 1, wherein, When the double-segmentation-head auxiliary classification model training is set, the following steps are included: The first segmentation head calculates a first cross-entropy loss on the feature map after the jump connection and the multiple up-sampling The second segmentation head calculates a second cross-entropy loss on the deep fusion feature The proportion of the second cross-entropy loss is controlled by a weight coefficient 0.6, and a final loss function is obtained as .

3. The method according to claim 2, wherein, The calculation formula of the first cross-entropy loss or the second cross-entropy loss is as follows: wherein, is the true result of the image classification; is the predicted result output by the model; is the height of the image; is the width of the image; is the number of image classes; is the pixel value of the i row and the j column of the predicted image corresponding to the class; is the pixel value of the i row and the j column of the label image corresponding to the class.

4. The method of claim 1, wherein the method further comprises: When the high-resolution remote sensing image data is obtained, the high-resolution remote sensing image contains three bands.

5. The method of claim 1, wherein, When the different scale feature maps are extracted, the feature maps of 1 / 2, 1 / 4, 1 / 8 and 1 / 16 scales are extracted.

6. A system for the image frequency domain feature-based high-resolution remote sensing image accurate classification method of claim 1, characterized in that, The application relates to a high-resolution remote sensing image classification method and device. The application comprises the following steps: A data acquisition module is used to acquire high-resolution remote sensing image data, normalize the image data, slice the image data with a fixed window size and retain the overlapping areas of adjacent slices, and meanwhile, a whole mask label corresponding to the image is made, and the label is aligned with the image through nearest neighbor interpolation. The model construction module is configured to construct a classification model using an encoder-decoder architecture, the encoder adopts a pre-trained ResNet-50 network as a backbone network, extracts different scale feature maps through four-stage convolution and down-sampling operations, and connects a multi-scale feature enhancement module after each output scale of the backbone network to perform feature enhancement on the scale feature maps and obtain enhanced feature maps of each scale; the decoder adopts a U-Net structure, performs step-by-step up-sampling through bilinear interpolation, splices and fuses the enhanced feature maps of the corresponding scale through jump connection and up-sampling features, and outputs a prediction result consistent with the size of the input image; and a double segmentation head auxiliary classification model is set to train the classification model to obtain a trained classification model. The classification module is configured to obtain a classification result of high-resolution remote sensing image data to be classified through the trained classification model.

Images