Remote sensing image ground object classification method, system and equipment based on deep learning, and medium

Abstract

The invention relates to a remote sensing image ground object classification method, system and device based on deep learning and a medium, and the method comprises the steps: carrying out the statistical normalization and overlapping cutting of an original remote sensing image and a label, and generating a subgraph data set; an encoder-decoder segmentation model integrated with an auxiliary supervision mechanism is constructed, the model captures long-distance spatial dependence by using an encoder based on sliding window attention, and multi-scale features are fused through a decoder; iteratively optimizing the model by adopting a composite loss function combining main loss and weighted auxiliary loss and a training strategy for optimizing specific parameters; and after same preprocessing is performed on images to be classified, reasoning is performed by using the trained model, and sub-graph prediction results are spliced into a complete ground feature classification graph according to coordinate information. According to the method, large-size remote sensing images can be effectively processed, the subdivision precision of complex landforms and the model training efficiency are improved, and an automatic process from preprocessing to classification result generation is realized.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image recognition technology, specifically to a method, system, device, and medium for classifying ground features in remote sensing images based on deep learning. Background Technology

[0002] Land cover classification in high-resolution remote sensing imagery refers to the process of automatically interpreting satellite or aerial images using computer technology, classifying each pixel in the image into predefined land cover categories such as buildings, water bodies, forests, and farmland. This technology is a core foundation for building smart cities, monitoring the ecological environment, and managing land resources. It can quickly and objectively extract valuable geographic information from massive amounts of image data, thereby assisting humans in planning and decision-making. With the rapid development of remote sensing satellite technology, the amount of image data available daily has increased dramatically, and its spatial and spectral resolution has also continuously improved, creating an urgent need for automated, high-precision land cover classification methods.

[0003] Currently, deep learning-based methods have become the mainstream technology in this field. By constructing complex neural network models, they can automatically learn the complex mapping relationship from image pixels to land cover categories. However, existing deep learning methods still face a series of challenges when applied to large-scale, high-resolution remote sensing imagery. First, remote sensing images are typically huge in size and contain multiple spectral bands. Directly inputting them into neural networks for processing leads to extremely high computational costs and memory consumption, while simple cropping can easily destroy the integrity and contextual relationships of land cover targets, affecting the consistency of classification. Second, although existing models are constantly improving in feature extraction capabilities, their accuracy and robustness are still insufficient when dealing with the intra-class differences and inter-class similarities unique to remote sensing images, caused by complex terrain and different imaging conditions. Furthermore, the complete process from model training to practical application often involves multiple loosely coupled links. The lack of integrated design in data preprocessing, model optimization strategies, and final result stitching and reconstruction results in low automation and efficiency in actual deployment, making it difficult to adapt to the rapid processing needs of images from different regions and different sensors. Summary of the Invention

[0004] Based on this, the purpose of the present invention is to provide a method, system, device and medium for classifying ground features in remote sensing images based on deep learning that can balance processing efficiency and classification accuracy, has a complete automated process and strong adaptability.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a method for classifying ground features in remote sensing images based on deep learning, comprising the following steps:

[0007] S1: Perform statistical normalization on the acquired original multispectral remote sensing images and corresponding RGB color label images based on preset quantiles, and perform overlapping grid cropping on the large-size images and label images to generate normalized image sub-sets and category index label sub-sets;

[0008] S2: Based on the sliding window attention mechanism and normalized image sub-sets, multi-scale features are extracted through a hierarchical Transformer encoder, and contextual information is fused using a pyramid pooling decoder. At the same time, an auxiliary supervision branch based on mid-level features is introduced to construct an initial semantic segmentation model with dual-path supervision.

[0009] S3: Input the normalized image sub-set and the category index label sub-set into the semantic segmentation initial model for iterative optimization. Calculate the composite loss of the main segmentation loss and the weighted auxiliary loss, and use an optimizer with weighted decay of location parameters and a piecewise changing learning rate strategy to backpropagate and update the network parameters, outputting the trained land cover classification model.

[0010] S4: Perform statistical normalization on the original remote sensing images to be classified based on preset quantiles and perform overlapping grid cropping, record the spatial starting coordinates of each cropped sub-image, and generate a test image sub-image set for model inference and a coordinate file that records the spatial location information of each sub-image;

[0011] S5: Input the test image sub-map set into the land cover classification model and perform forward propagation calculation to obtain the category prediction matrix of each sub-map. Combine the spatial location information recorded in the coordinate file to stitch and align the prediction results of all sub-maps to generate a high-resolution pixel-level land cover classification map that covers the entire original image range.

[0012] In one embodiment, S1 of the remote sensing image land cover classification method based on deep learning provided by the present invention specifically includes the following steps:

[0013] S11: The pixel values ​​of each band of the acquired original multispectral remote sensing image are truncated and linearly stretched based on preset high and low quantiles. The preset high and low quantiles of the pixel values ​​of each band of the image are calculated, and the pixel values ​​are mapped to the target value range based on the actual distribution range of the pixel values ​​of the bands to generate normalized data for each band.

[0014] S12: Perform color-to-category index mapping on the acquired RGB color label image. Based on the predefined RGB color values ​​of local object category standards, match the RGB vector of each pixel in the label image to the closest category and convert it into a single numerical index. Based on the preset unlabeled category identifier, assign an unlabeled identifier to the unmatched pixels to generate a single-channel category index image.

[0015] S13: Synchronously segment the normalized data and category index map for each band, and crop the original multispectral remote sensing image and label map with a set step size based on a sliding window of preset size to generate normalized image sub-atlas and category index label sub-atlas.

[0016] In one embodiment, the formula for calculating normalized data in a deep learning-based remote sensing image land cover classification method provided by the present invention is as follows:

[0017]

[0018] in, These are the normalized pixel values. For the original image in pixel coordinates The original pixel value at that location, and These represent the preset lower quantile and higher quantile cutoff boundaries to eliminate the effects of noise and extreme values, respectively. and These represent the minimum and maximum values ​​of the target numerical range after normalization, respectively.

[0019] In one embodiment, S2 of the remote sensing image land cover classification method based on deep learning provided by the present invention specifically includes the following steps:

[0020] S21: Based on a non-overlapping window of a preset size, each image sub-image in the normalized image sub-image set is divided into regular image blocks, and each two-dimensional image sub-image is converted into a series of one-dimensional image block embedding vector sequences to generate an image block sequence to be encoded;

[0021] S22: Based on the sliding window multi-head self-attention mechanism and the shift window multi-head self-attention mechanism, the image block sequence is hierarchically feature encoded, so that the image block sequence is sequentially passed through multiple hierarchical Transformer encoders composed of window multi-head self-attention layer, shift window multi-head self-attention layer and feedforward network layer to perform feature interaction and transformation, generating multi-scale encoded feature maps with decreasing resolution, increasing number of channels and containing long-distance contextual semantic information.

[0022] S23: Based on the feature pyramid network technology, multi-scale context information is aggregated for multi-scale encoded feature maps. The deepest feature map in the multi-scale encoded feature map is subjected to adaptive pooling at different scales in parallel. Then, the pooling results at different scales are upsampled and concatenated with the original feature map to generate an enhanced feature map rich in multi-scale global context.

[0023] S24: The enhanced feature map is progressively upsampled through convolutional layers and then channel-wise concatenated and convolutionally fused with the original encoded feature map of the multi-scale encoded feature map to gradually restore spatial details and generate a high-resolution decoded feature map for pixel-wise classification.

[0024] S25: Based on the intermediate layer feature map in the layer output of the hierarchical Transformer encoder in the multi-scale encoded feature map, construct the auxiliary supervision branch, select the feature output of a specific stage from the intermediate layer feature map, restore the spatial resolution of the intermediate layer feature map to the input image size through independent convolutional layers and upsampling operations, and generate an auxiliary prediction feature map for providing additional gradient supervision during the training stage.

[0025] S26: Based on the high-resolution decoded feature map, the main supervision branch is constructed. The high-resolution decoded feature map is linearly projected along the channel dimension to map the number of feature channels to be consistent with the total number of land cover categories, thereby generating the main prediction feature map used to output the final pixel-level classification result.

[0026] S27: Based on the main prediction feature map and the auxiliary prediction feature map, a dual-path supervision model is integrated. The processing path corresponding to the main prediction feature map is defined as the main supervision branch, and the processing path corresponding to the auxiliary prediction feature map is defined as the auxiliary supervision branch. The two branches are structurally linked to form an initial semantic segmentation model with dual-path supervision.

[0027] In one embodiment, S3 of the remote sensing image land cover classification method based on deep learning provided by the present invention specifically includes the following steps:

[0028] S31: Input the normalized image sub-map into the semantic segmentation initial model, process it through the encoder-decoder backbone network path, output the main segmentation prediction map, and process it through the auxiliary supervision branch path to output the auxiliary prediction map, outputting the corresponding main prediction feature map and auxiliary prediction feature map.

[0029] S32: Based on the cross-entropy loss function, loss calculation is performed on the main predicted feature map, auxiliary predicted feature map and category index label sub-map set. The difference between the main predicted feature map and the corresponding real label of each sample is calculated as the main segmentation loss, and the difference between the auxiliary predicted feature map and the corresponding real label is calculated as the auxiliary loss. The loss is calculated by combining the preset positive decimal weights to generate the composite loss value of the current training batch.

[0030] S33: Based on the adaptive moment estimation algorithm, the composite loss value is backpropagated and updated to process the model parameters. The gradient of the composite loss value with respect to all trainable parameters of the model is calculated. The position encoding parameters and layer normalization parameters in the model are exempted from weight decay penalty. The gradient information is used to update the model weights to minimize the loss and generate a semantic segmentation model with updated parameters.

[0031] S34: Based on the polynomial decay strategy combined with linear preheating, the semantic segmentation model is dynamically scheduled. In the initial stage of training, the learner's learner rate is linearly increased from zero to a preset peak value, and then gradually decayed from the peak value according to the power rule during the training stage. The step size of parameter updates is dynamically adjusted to control the model convergence process and generate a trained land cover classification model.

[0032] In one embodiment, the formula for calculating the composite loss value of a deep learning-based remote sensing image land cover classification method provided by the present invention is as follows:

[0033]

[0034] in, This is the composite loss value. and Both represent the cross-entropy loss function; and These represent the prediction results output by the main supervision branch and the prediction results output by the auxiliary supervision branch of the model, respectively. A subgraph representing the actual category index labels; It is a preset auxiliary loss weight coefficient between 0 and 1, used to adjust the contribution of auxiliary supervision to the total loss.

[0035] In one embodiment, S5 of the remote sensing image land cover classification method based on deep learning provided by the present invention specifically includes the following steps:

[0036] S51: Input the test image sub-map set into the land cover classification model in batches for forward inference processing, so that each image sub-map passes through the encoder-decoder structure of the model and outputs the category prediction matrix of each sub-map. Each element of the category prediction matrix represents the predicted land cover category index of the corresponding pixel.

[0037] S52: Spatial positioning of the category prediction matrix of each sub-image based on the coordinate file; according to the coordinates of each sub-image recorded in the original image, place the category prediction matrix in the corresponding position of a blank canvas with all zeros and the same size as the original image to generate a complete prediction map.

[0038] S53: Arbitrate the pixel categories of the overlapping stitched areas in the complete prediction map. For pixel locations that are covered by the prediction results of multiple sub-images in the final stitched map due to overlapping cropping, calculate the mode of multiple prediction category indices corresponding to the pixel location, determine the final category of the pixels in the overlapping stitched areas, and generate a high-resolution pixel-level land cover classification map that covers the entire original image range.

[0039] Secondly, the present invention provides a remote sensing image land cover classification system based on deep learning, which is configured with the following modules:

[0040] The remote sensing image preprocessing and cropping module is used to perform statistical normalization on the acquired original multispectral remote sensing images and corresponding RGB color label images based on preset quantiles, and to perform overlapping grid cropping on the large-size images and label images to generate normalized image sub-atlases and category index label sub-atlases.

[0041] The dual-path supervised segmentation model construction module is used to extract multi-scale features through a hierarchical Transformer encoder based on the sliding window attention mechanism and normalized image sub-atlas, and to fuse contextual information using a pyramid pooling decoder. At the same time, an auxiliary supervision branch based on mid-level features is introduced to construct an initial semantic segmentation model with dual-path supervision.

[0042] The segmentation model iterative training optimization module is used to input the normalized image sub-map set and the category index label sub-map set into the semantic segmentation initial model for iterative optimization. It calculates the composite loss of the main segmentation loss and the weighted auxiliary loss, and uses an optimizer with weighted decay of location parameters and a segmented learning rate strategy to backpropagate and update the network parameters, outputting the trained land cover classification model.

[0043] The test image preprocessing coordinate recording module is used to perform statistical normalization on the original remote sensing images to be classified based on preset quantiles and to perform overlapping grid cropping, record the spatial starting coordinates of each cropped sub-image, and generate a test image sub-image set for model inference and a coordinate file that records the spatial location information of each sub-image.

[0044] The land cover classification result inference and stitching module is used to input the test image sub-map set into the land cover classification model for forward propagation calculation to obtain the category prediction matrix of each sub-map, and combine it with the spatial location information recorded in the coordinate file to stitch and align the prediction results of all sub-maps to generate a high-resolution pixel-level land cover classification map covering the entire original image range.

[0045] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned deep learning-based remote sensing image land cover classification methods.

[0046] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the aforementioned deep learning-based remote sensing image land cover classification methods.

[0047] In summary, the deep learning-based remote sensing image land cover classification method provided in this application effectively eliminates image noise and preserves land cover contextual information through statistical distribution-based normalization and overlapping cropping preprocessing, thus adapting to efficient and lossless processing of large-size images. By integrating a sliding window attention mechanism and a multi-scale feature fusion model architecture, the model's ability to model complex terrain structures and long-distance spatial dependencies in remote sensing images and its classification accuracy can be significantly improved. The composite loss function and refined optimization strategy for the Transformer structure used in the training phase effectively accelerate model convergence and enhance its generalization performance, improving the stability and efficiency of the training process. Finally, the coordinate-based automatic stitching mechanism ensures the integrity and geometric accuracy from local prediction to global mapping. Therefore, this method ultimately achieves an end-to-end automated solution from data preprocessing and model training to batch inference and reconstruction, enabling high-efficiency, high-accuracy, and robust large-scale remote sensing image land cover classification in practical applications.

[0048] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0049] Figure 1 A flowchart illustrating a deep learning-based method for classifying land cover in remote sensing images, provided for an embodiment of this application;

[0050] Figure 2 This is a schematic diagram of the structure of a remote sensing image land cover classification system based on deep learning, provided for another embodiment of this application. Detailed Implementation

[0051] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0052] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0053] In one embodiment, such as Figure 1 As shown, a deep learning-based method for classifying ground features in remote sensing images is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0054] S1: Perform statistical normalization on the acquired original multispectral remote sensing images and corresponding RGB color label images based on preset quantiles, and perform overlapping grid cropping on the large-size images and label images to generate normalized image sub-sets and category index label sub-sets.

[0055] Specifically, the system iterates through all pixels of the original multispectral remote sensing image, extracting the pixel value set for each spectral band. This set consists of the pixel values ​​of all pixels in the corresponding band. The system calculates the preset quantiles for each band's pixel value set, selecting the lower and upper quantiles, and performs normalization based on a linear transformation formula. The numerator of the linear transformation formula is the original pixel value minus the lower quantile of the corresponding band, and the denominator is the upper quantile minus the lower quantile. The result is the normalized pixel value. The system sets normalized pixel values ​​less than zero to zero and normalized pixel values ​​greater than one to one, eliminating the interference of illumination differences under different imaging conditions on model training. After normalization, the system performs overlapping grid cropping on the normalized multispectral remote sensing image and the RGB color label image.

[0056] Furthermore, the system sets the cropping grid size and the overlap ratio of adjacent cropped sub-images in the horizontal and vertical directions. The width of the overlapping area is calculated from the cropping grid size and the overlap ratio. During the cropping process, the system establishes a pixel coordinate system with the upper left corner as the origin and slides the cropping window sequentially from left to right and from top to bottom. The sliding step size is calculated from the cropping grid size and the width of the overlapping area to ensure that the cropping window completely covers the original image area. The system assigns a one-to-one correspondence number to the cropped image sub-images and their corresponding label sub-images, generating a normalized image sub-image set and a category index label sub-image set. The category index label sub-image set is obtained by converting the RGB color label image into a single-channel category index image, and the index value forms a one-to-one mapping relationship with the land cover category.

[0057] S2: Based on the sliding window attention mechanism and normalized image sub-sets, multi-scale features are extracted through a hierarchical Transformer encoder, and contextual information is fused using a pyramid pooling decoder. At the same time, an auxiliary supervision branch based on mid-level features is introduced to construct an initial semantic segmentation model with dual-path supervision.

[0058] Specifically, the system constructs an initial semantic segmentation model with dual-path supervision based on a sliding window attention mechanism and a normalized image sub-map set. The model consists of three parts: a hierarchical Transformer encoder, a pyramid pooling decoder, and a mid-level feature-assisted supervision branch. Preferably, the system constructs a hierarchical Transformer encoder with a hierarchical cascaded structure. The number of feature extraction stages is determined according to actual needs, and each stage consists of a sliding window attention module and a multilayer perceptron module. In the sliding window attention module, the system divides the input image sub-map into multiple non-overlapping windows. The module only calculates attention weights within the windows. Compared to the global attention mechanism, this approach reduces computational complexity and effectively adapts to feature extraction from large-size remote sensing images. Each feature extraction stage outputs feature maps of different scales sequentially, with the scaling ratios of the feature maps increasing sequentially to achieve multi-scale hierarchical feature extraction.

[0059] Furthermore, the system constructs a pyramid pooling decoder. The decoder takes the smallest-scale feature map output by the encoder as input and uses a pyramid pooling module to fuse multi-scale contextual information. The pyramid pooling module contains multiple parallel pooling branches, each with a different pooling kernel size. Each branch performs a global average pooling operation on the input feature map, then adjusts the number of channels through a one-dimensional convolutional layer, and restores it to the same size as the input feature map through an upsampling operation. Finally, the output features of each branch are concatenated with the original input features by channel dimension to achieve the fusion of features from different receptive fields. An auxiliary supervision branch is connected to the output of the encoder at the specified feature extraction stage. This branch consists of a one-dimensional convolutional layer and a Softmax activation function, used for class prediction of mid-level features. The main segmentation branch is set at the decoder output. The main segmentation branch and the auxiliary supervision branch form a dual-path supervision structure, improving the model's ability to recognize fine-grained ground features through supervision constraints on features at different levels. After completing the parameter initialization operations of each module, the system outputs an initial semantic segmentation model. The model input is a normalized image sub-map set, and the output is the prediction results of the main branch and the prediction results of the auxiliary branch.

[0060] S3: Input the normalized image sub-set and the category index label sub-set into the initial semantic segmentation model for iterative optimization. Calculate the composite loss of the main segmentation loss and the weighted auxiliary loss, and use an optimizer with weighted decay of location parameters and a segmented learning rate strategy for backpropagation to update the network parameters, outputting the trained land cover classification model.

[0061] Specifically, the system inputs a normalized image sub-set and a category index label sub-set into the initial semantic segmentation model for iterative optimization training, outputting a trained land cover classification model. The system constructs a composite loss function consisting of a main segmentation loss and a weighted auxiliary loss. The composite loss function is calculated by adding the main segmentation loss to the product of the auxiliary loss and the weight coefficients. The main segmentation loss uses the cross-entropy loss function to calculate the difference between the decoder output and the category index labels. The auxiliary loss is calculated using the same method as the main segmentation loss. The weight coefficients are used to balance the loss contributions of the main branch and the auxiliary branch. The system is configured with an optimizer that disables weight decay for location parameters. During optimization, the system does not impose weight decay constraints on the location encoding parameters in the Transformer encoder to avoid destroying location information during parameter updates. The system applies weight decay to the remaining network parameters.

[0062] The system employs a segmented learning rate strategy, dividing the training process into two phases: a warm-up phase where the learning rate linearly increases from its initial value to a preset learning rate, and a cosine annealing phase where the learning rate periodically decays according to a cosine function. The system inputs normalized image sub-sets and category index label sub-sets into the model in batches, calculates the composite loss function value through forward propagation, and updates the network parameters through backpropagation. After each preset number of iterations, the system uses a validation set to evaluate the model's accuracy. Training terminates when the classification accuracy on the validation set fails to improve for the preset number of iterations. Upon completion of training, the system saves the optimal model parameters, which correspond to the highest classification accuracy on the validation set. The system then outputs the trained ground cover classification model.

[0063] S4: Perform statistical normalization on the original remote sensing images to be classified based on preset quantiles and perform overlapping grid cropping. Record the spatial starting coordinates of each cropped sub-image to generate a test image sub-image set for model inference and a coordinate file that records the spatial location information of each sub-image.

[0064] Specifically, the system performs preprocessing operations on the raw remote sensing images to be classified, generating test image sub-maps and coordinate files for model inference. The system employs a statistical normalization method based on preset quantiles, identical to the normalization and overlapping grid clipping steps, to normalize the raw remote sensing images, ensuring consistency in the preprocessing workflow. The system uses the same grid size and overlap ratio as the normalization and overlapping grid clipping steps to perform overlapping grid clipping on the normalized remote sensing images to be classified. During the clipping process, the system records the spatial starting coordinates of each clipped sub-map in real time. The spatial starting coordinates are the x and y coordinates of the top-left pixel of the sub-map in the pixel coordinate system of the original remote sensing image.

[0065] Furthermore, the system assigns a number to all cropped test image sub-images using a sequential numbering method, generating a test image sub-image set. The system also stores the sub-image number, spatial starting coordinates, and sub-image size in a coordinate file. The coordinate file is stored in text format, with each record containing all relevant information for a single sub-image, and the information fields are arranged in a preset order. By maintaining consistency between the preprocessing steps and the model training phase, the system ensures that the test data input to the model has the same data distribution characteristics as the training data.

[0066] S5: Input the test image sub-map set into the land cover classification model and perform forward propagation calculation to obtain the category prediction matrix of each sub-map. Combine the spatial location information recorded in the coordinate file to stitch and align the prediction results of all sub-maps to generate a high-resolution pixel-level land cover classification map that covers the entire original image range.

[0067] Specifically, the system inputs test image sub-maps into the trained land cover classification model to perform inference operations, and combines the results with coordinate files to stitch together and generate a complete land cover classification map. The system inputs test image sub-maps into the trained land cover classification model in batches, and calculates the category prediction matrix for each sub-map through forward propagation. The dimensions of the category prediction matrix consist of the sub-map height, sub-map width, and the total number of land cover categories. The system performs an argmax operation on the prediction matrix to obtain the category index for each pixel, generating pixel-level classification results for each sub-map. The system reads the spatial location information from the coordinate file, and using the pixel coordinate system of the original remote sensing image as a reference, maps the classification result of each sub-map to the corresponding position in the original image according to its spatial starting coordinates. For overlapping areas of adjacent sub-maps, the system uses a majority voting method to determine the final category. The majority voting method involves statistically analyzing multiple category prediction results for the same pixel location within the overlapping area and selecting the category with the most occurrences as the final category for that pixel. After completing the stitching of all sub-maps and processing of overlapping areas, the system generates a high-resolution pixel-level land cover classification map covering the entire original image area. In the classification map, each pixel is assigned the category index of the corresponding land feature. The system can convert the classification map into an RGB color classification map based on the mapping relationship between the category index and the land feature category. The conversion method is to match each category index with the corresponding RGB color value and assign it to the corresponding pixel.

[0068] In summary, the deep learning-based remote sensing image land cover classification method provided in this application effectively eliminates image noise and preserves land cover contextual information through statistical distribution-based normalization and overlapping cropping preprocessing, thus adapting to efficient and lossless processing of large-size images. By integrating a sliding window attention mechanism and a multi-scale feature fusion model architecture, the model's ability to model complex terrain structures and long-distance spatial dependencies in remote sensing images and its classification accuracy can be significantly improved. The composite loss function and refined optimization strategy for the Transformer structure used in the training phase effectively accelerate model convergence and enhance its generalization performance, improving the stability and efficiency of the training process. Finally, the coordinate-based automatic stitching mechanism ensures the integrity and geometric accuracy from local prediction to global mapping. Therefore, this method ultimately achieves an end-to-end automated solution from data preprocessing and model training to batch inference and reconstruction, enabling high-efficiency, high-accuracy, and robust large-scale remote sensing image land cover classification in practical applications.

[0069] For example, the deep learning-based remote sensing image land cover classification method described in this application can be adapted to the high-resolution remote sensing image land cover classification needs of the Daqingshan National Nature Reserve. Based on the landform characteristics and management requirements of the reserve, a complete application chain is constructed, from multispectral image processing to accurate land cover identification, and from dynamic monitoring to decision support. Specific application scenarios are as follows:

[0070] In the dynamic monitoring of the ecosystem in the Daqingshan Nature Reserve, this method can effectively balance image contrast and retain key spectral information by truncating and linearly stretching 16-bit multi-channel TIFF remote sensing images containing B, G, R, and Nir bands acquired by the Gaofen-2 satellite through truncation based on preset high and low quantiles and large-size image data cropping. Combined with a dual-path supervised semantic segmentation model, it accurately identifies the pixel-level distribution of seven land cover types: man-made structures, farmland, water bodies, forests, grasslands, mountains, and other landforms. By periodically processing remote sensing images from different time phases, the system can continuously track changes in forest cover, grassland degradation trends, water area fluctuations, and mountain topographic integrity within the reserve, providing standardized data support for ecosystem health assessment and biodiversity conservation effectiveness analysis, and helping the reserve understand the evolutionary patterns of the ecosystem.

[0071] In the refined management of natural resources in protected areas, the method can achieve quantitative statistics on various natural resources within the protected area through precise mapping between RGB color-labeled maps and category indexes. For key resource information such as farmland distribution, forest land ownership boundaries, and watershed extent within the protected area, the system utilizes multi-scale features extracted by a hierarchical Transformer encoder and contextual information fused by pyramid pooling to improve the accuracy and robustness of land cover classification, accurately delineating the spatial boundaries and area scale of different resource types. The generated high-resolution pixel-level land cover classification map can directly provide data support for land resource planning, forest protection red line delineation, and rational water resource allocation in protected areas. Simultaneously, the standardized output of the category index map adapts to the needs of digital archiving and cross-departmental data sharing in natural resource management.

[0072] In the planning and law enforcement supervision of protected areas, this method can quickly process large-scale remote sensing images. Through test image preprocessing, coordinate recording, and prediction result stitching, a complete land feature classification map covering the entire Daqingshan Nature Reserve is generated, clearly showing the distribution of man-made features such as illegal buildings and road expansions within the reserve. Law enforcement personnel can quickly locate the specific location and scope of illegal development and illegal encroachment on protected area land based on the classification results, providing precise and targeted guidance for law enforcement inspections. At the same time, in the process of compiling the overall plan for the protected area, the classification results can help identify ecologically sensitive areas and suitable development areas, providing scientific support for planning decisions such as functional zoning and ecological corridor construction, ensuring that the planning scheme is compatible with the natural landform and ecological pattern of the protected area.

[0073] In emergency response and risk prevention in protected areas, this method can efficiently process remote sensing image data in emergency scenarios. When natural disasters such as forest fires and landslides occur within the protected area, the system can quickly classify and process post-disaster remote sensing images. Through forward inference and result stitching, it can quickly identify the types of land features in the affected area, such as the extent of burned forest land, the area of ​​mountains and grasslands affected by landslides, and quantify the degree of damage. The generated classification results can provide timely disaster assessment data to emergency command departments, assisting in the formulation of disaster relief plans and post-disaster recovery and reconstruction plans. At the same time, by monitoring changes in key land features such as mountains and water bodies over a long period of time, it can provide early warnings of potential geological disaster sites and water resource safety risks, thereby improving the risk prevention and control capabilities of the protected area. This method is compatible with the joint training mode of the GID dataset and the Daqingshan Gaofen-2 satellite imagery. Through transfer learning and iterative optimization, it can continuously improve the classification accuracy for the unique landforms of the protected area. The standardized land feature classification data generated can be integrated with the existing ecological monitoring platform and resource management system of the protected area, forming a closed loop from data collection, processing, analysis to application, providing core technical support for the scientific, refined, and digital management of the Daqingshan National Nature Reserve.

[0074] In one embodiment, S1 of the remote sensing image land cover classification method based on deep learning provided by the present invention specifically includes the following steps:

[0075] S11: The pixel values ​​of each band of the acquired original multispectral remote sensing image are truncated and linearly stretched based on preset high and low quantiles. The preset high and low quantiles of the pixel values ​​of each band of the image are calculated, and the pixel values ​​are mapped to the target value range based on the actual distribution range of the pixel values ​​of the bands to generate normalized data for each band.

[0076] Specifically, the system performs truncation and linear stretching processing on the pixel values ​​of each band of the acquired raw multispectral remote sensing image based on preset high and low quantiles, generating normalized data for each band. The system first reads all band data from the raw multispectral remote sensing image, and independently performs quantile statistical analysis of the pixel values ​​for each spectral band to determine preset low and high quantiles as truncation boundaries. To adapt to the differences in pixel value distribution across different bands, the system adaptively corrects the preset quantiles using a newly created formula:

[0077]

[0078]

[0079] in, To truncate the boundary for the corrected lower quantiles, To cut off the boundary for the corrected higher quantiles, This is the quantile correction factor. This is the mean of all original pixel values ​​in this band. The system performs a truncation operation on the original pixel values ​​of this band based on the corrected truncation boundary, adjusting pixel values ​​below the corrected lower quantile to the value corresponding to the corrected lower quantile, and adjusting pixel values ​​above the corrected higher quantile to the value corresponding to the corrected higher quantile.

[0080] After truncation, the system performs linear stretching calculations on the truncated pixel values, mapping them to a preset target value range. The formula for normalizing the data is as follows:

[0081]

[0082] in, These are the normalized pixel values. For the original image in pixel coordinates The original pixel value at that location, and These represent the preset lower quantile and higher quantile cutoff boundaries to eliminate the effects of noise and extreme values, respectively. and These represent the minimum and maximum values ​​of the target numerical range after normalization, respectively.

[0083] S12: Perform color-to-category index mapping on the acquired RGB color label image. Based on the predefined RGB color values ​​of local object category standards, match the RGB vector of each pixel in the label image to the closest category and convert it into a single numerical index. Based on the preset unlabeled category identifier, assign an unlabeled identifier to the unmatched pixels to generate a single-channel category index image.

[0084] Specifically, the system reads all pixel data from the RGB color label image, obtains the RGB vector corresponding to each pixel, and then calls a pre-stored set of standard RGB color values ​​for land cover categories. This set contains standard RGB vectors corresponding to all land cover categories to be classified, with each standard RGB vector corresponding to a unique numerical index of the land cover category. The system determines the matching relationship between the pixel RGB vector and the standard RGB vector using a newly created similarity calculation standard. The similarity calculation formula is as follows:

[0085]

[0086] in, The similarity between the pixel RGB vector and the standard RGB vector. Represents the three color channels of RGB. This represents the value of the pixel's RGB vector in the corresponding channel. This refers to the numerical value of the standard RGB vector in the corresponding channel. For each pixel, the system calculates the similarity between its RGB vector and each standard RGB vector in the set, and determines the land feature category index corresponding to the standard RGB vector with the highest similarity as the category index of that pixel. During the mapping process, the system simultaneously identifies pixels that cannot be matched. The similarity of the RGB vector of such pixels with all standard RGB vectors in the set does not meet the preset matching requirements. The system assigns a corresponding label value to such pixels based on preset unlabeled category identifiers. The system performs the above operations sequentially on all pixels, and finally generates a single-channel category index map, the size of which is the same as the original RGB color label. Figure 1 To.

[0087] S13: Synchronously segment the normalized data and category index map for each band, and crop the original multispectral remote sensing image and label map with a set step size based on a sliding window of preset size to generate normalized image sub-atlas and category index label sub-atlas.

[0088] Specifically, the system stitches and integrates the normalized data of all bands along their band dimensions to form multi-channel normalized image data. The spatial dimensions of this data are consistent with the original multispectral remote sensing image, and the number of channels is the same as the number of bands in the original image. The system calls the single-channel category index map to confirm that its spatial dimensions completely match the multi-channel normalized image data. Based on a preset sliding window size and a set step size, the system performs simultaneous cropping processing on both types of data. To ensure the spatial uniqueness and traceability of the cropped sub-images, the system generates a unique identifier for each sub-image using a preset formula. The identifier generation formula is:

[0089]

[0090] in, Used as a unique identifier for the subgraph. This represents the total number of horizontal pixels in the multi-channel normalized image data. y is the horizontal coordinate of the top-left corner of the sliding window, and y is the vertical coordinate of the top-left corner of the sliding window. The step size of the sliding window. This is a rounding down function. During the cropping process, the sliding window simultaneously covers the same spatial region of both types of data, ensuring that the image sub-image and the label sub-image correspond in spatial position. The system stores the generated multi-channel normalized image sub-images in order of their labels to form a normalized image sub-image set, and stores the category index label sub-images in order of their same labels to form a category index label sub-image set.

[0091] In one embodiment, S2 of the remote sensing image land cover classification method based on deep learning provided by the present invention specifically includes the following steps:

[0092] S21: Based on a non-overlapping window of a preset size, each image sub-image in the normalized image sub-image set is divided into regular image blocks, and each two-dimensional image sub-image is converted into a series of one-dimensional image block embedding vector sequences to generate an image block sequence to be encoded.

[0093] Specifically, based on a non-overlapping window of a preset size, the system performs regular image block partitioning on each image sub-image in the normalized image sub-image set, converting each two-dimensional image sub-image into a series of one-dimensional image block embedding vector sequences, generating an image block sequence to be encoded. The system first reads each image sub-image in the normalized image sub-image set, obtaining the spatial dimensions and number of channels for each sub-image. It then calls the preset non-overlapping window size parameters and uniformly partitions the two-dimensional image sub-image according to the window size, ensuring that the windows do not overlap and completely cover the entire image sub-image, with no unpartitioned pixel regions. To achieve the conversion from two-dimensional image blocks to one-dimensional embedding vectors, the system completes dimensionality transformation and feature enhancement through a newly created embedding mapping formula. The embedding mapping formula is as follows:

[0094]

[0095] in, Let be the one-dimensional embedding vector of the i-th image patch. To embed the weight matrix, Let be the one-dimensional vector formed by flattening the pixel value matrix of the i-th two-dimensional image patch. For embedding bias vectors, the system sequentially performs flattening and embedding mapping operations on all two-dimensional image blocks after each image sub-image is divided. All generated one-dimensional embedding vectors are arranged according to the spatial position of the image blocks within the sub-image, forming an image block embedding vector sequence corresponding to each image sub-image. The system sequentially performs the above division, flattening, and embedding mapping operations on all image sub-images in the normalized image sub-image set, ultimately generating a sequence of image blocks to be encoded corresponding to all image sub-images.

[0096] S22: Based on the sliding window multi-head self-attention mechanism and the shift window multi-head self-attention mechanism, the image block sequence is hierarchically feature encoded, so that the image block sequence is sequentially passed through multiple hierarchical Transformer encoders composed of window multi-head self-attention layer, shift window multi-head self-attention layer and feedforward network layer to perform feature interaction and transformation, generating multi-scale encoded feature maps with decreasing resolution, increasing number of channels and containing long-distance contextual semantic information.

[0097] Specifically, the system performs hierarchical feature encoding on the generated image patch sequences based on sliding window multi-head self-attention and shift window multi-head self-attention mechanisms, generating multi-scale encoded feature maps with decreasing resolution, increasing channel number, and containing long-range contextual semantic information. Specifically, the system calls multiple pre-built hierarchical Transformer encoders, each consisting of a window multi-head self-attention layer, a shift window multi-head self-attention layer, and a feedforward network layer connected in series, with each encoder's parameters configured independently. To achieve feature complementarity between the two attention mechanisms, the system integrates the output features of both mechanisms using a newly created attention weight fusion formula, which is:

[0098]

[0099] in, The fused feature map For attention fusion weights, The output features of the sliding window multi-head self-attention layer, The output features of the multi-head self-attention layer in the shift window are used. Image patch sequences are sequentially input into multiple hierarchical Transformer encoders. The output features of the previous encoder serve as the input features for the next. Each encoder performs feature interaction and transformation on the input features, outputting a feature map with reduced resolution and increased channel count. Through the progressive processing of each hierarchical Transformer encoder, the system gradually extracts shallow detail features and deep semantic features from the image patch sequence, while simultaneously capturing long-distance contextual relationships between image patches. Ultimately, it generates multiple encoded feature maps at different scales, forming a multi-scale encoded feature map set.

[0100] S23: Based on the feature pyramid network technology, multi-scale context information is aggregated for multi-scale encoded feature maps. The deepest feature map in the multi-scale encoded feature map is subjected to adaptive pooling at different scales in parallel. Then, the pooling results at different scales are upsampled and concatenated with the original feature map to generate an enhanced feature map rich in multi-scale global context.

[0101] Specifically, the system extracts the deepest feature map from a multi-scale encoded feature map set. This feature map contains the richest deep semantic information but has the lowest resolution. The system performs adaptive pooling operations at different scales in parallel on the deepest feature map. Each pooling branch corresponds to a preset pooling scale, and the pooling kernel size is adaptively adjusted according to the feature map size during the adaptive pooling process. To achieve effective fusion of pooling results at different scales with the original feature map, the system adjusts the dimensions of each pooling result using a newly created pooling feature alignment formula. The alignment formula is as follows:

[0102]

[0103] in, These are pooled features after dimension alignment. To align the weight matrix, This represents the original pooling characteristics. To align the bias vector, the system performs upsampling operations on the pooling results at different scales after aligning all dimensions, restoring their resolution to the same level as the original features of the deepest layer. Figure 1 Then, all the upsampled pooled features are concatenated with the deepest original feature map. After concatenation, features are fused through convolutional layers to eliminate redundant information between features of different scales, enhance effective features, and finally generate an enhanced feature map rich in multi-scale global context information.

[0104] S24: The enhanced feature map is progressively upsampled through convolutional layers and then fused with the original encoded feature map of the multi-scale encoded feature map through channel concatenation and convolution to gradually restore spatial details and generate a high-resolution decoded feature map for pixel-by-pixel classification.

[0105] Specifically, the system performs upsampling on the enhanced feature map. This upsampling process is implemented through a transposed convolutional layer. Each upsampling doubles the resolution of the feature map, while simultaneously adjusting the number of feature channels to reduce feature redundancy. To achieve effective fusion of the upsampled features and the original encoded features, the system determines the fusion weights using a newly created feature fusion coefficient calculation standard. The fusion coefficient formula is as follows:

[0106]

[0107] in, The fusion coefficient is... This is the variance calculation function. This is the feature map after upsampling. This corresponds to the original encoded feature map. The system performs channel concatenation between the upsampled feature map and the original encoded feature map of the corresponding resolution in the multi-scale encoded feature map. After concatenation, feature fusion is performed through a convolutional layer using fusion coefficients to strengthen the association between shallow detail features and deep semantic features. The system repeats the above upsampling, channel concatenation, and convolutional fusion operations to gradually restore the spatial resolution of the feature map until the feature map resolution matches the size of the input image sub-image, generating a high-resolution decoded feature map for pixel-by-pixel classification.

[0108] S25: Based on the intermediate layer feature map in the layer output of the hierarchical Transformer encoder in the multi-scale encoded feature map, construct the auxiliary supervision branch, select the feature output of a specific stage from the intermediate layer feature map, and restore the spatial resolution of the intermediate layer feature map to the input image size through independent convolutional layers and upsampling operations to generate an auxiliary prediction feature map for providing additional gradient supervision during the training stage.

[0109] Specifically, the system extracts intermediate layer feature maps from the multi-scale encoded feature map set, derived from the outputs of the hierarchical Transformer encoder. These feature maps contain both semantic features and retain some spatial detail features, making them suitable as feature inputs for auxiliary supervision. The system selects feature outputs for specific stages using a newly created intermediate layer feature selection formula:

[0110]

[0111] in, This is the filtered intermediate layer feature map. Weights are used to select features. This is the intermediate layer feature map output by the kth intermediate layer. The system configures independent convolutional layers and upsampling layers for the auxiliary supervision branch. The filtered intermediate layer feature maps are input into the independent convolutional layer, the number of feature channels is adjusted, and then the feature map resolution is gradually increased through the upsampling layer, using the same upsampling method as in S24. Through the coordinated processing of the independent convolutional layer and the upsampling operation, the system restores the spatial resolution of the intermediate layer feature maps to the input image size while ensuring the feature integrity of the feature maps. Finally, an auxiliary prediction feature map is generated, which provides additional gradient supervision during the training phase, alleviating the gradient vanishing problem during model training.

[0112] S26: Based on the high-resolution decoded feature map, the main supervision branch is constructed. The high-resolution decoded feature map is linearly projected along the channel dimension to map the number of feature channels to be consistent with the total number of land cover categories, thereby generating the main prediction feature map used to output the final pixel-level classification result.

[0113] Specifically, the system inputs the high-resolution decoded feature map into the linear projection layer of the main supervision branch. This linear projection layer performs a channel-dimensional linear projection on the high-resolution decoded feature map, mapping the number of feature channels to match the total number of land cover categories, ensuring that the output features correspond to the predicted probabilities of each land cover category. To achieve accurate channel-dimensional mapping, the system completes the feature transformation using a newly created linear projection formula, which is:

[0114]

[0115] in, Main prediction feature map, This is a linear projection weight matrix, where the number of columns equals the total number of land cover categories, and the number of rows equals the number of channels in the high-resolution decoded feature map. For high-resolution decoding feature maps, This is the linear projection bias vector. The system performs channel transformation on each pixel feature of the high-resolution decoded feature map through a linear projection layer, mapping the high-dimensional feature vector of each pixel to a low-dimensional feature vector corresponding to the total number of land cover categories. Each element of the low-dimensional feature vector corresponds to the feature response value of the pixel belonging to the corresponding land cover category. After linear projection, the system outputs the master prediction feature map, the size of which is proportional to the input image sub-image. Figure 1 The number of channels is consistent with the total number of land feature categories.

[0116] S27: Based on the main prediction feature map and the auxiliary prediction feature map, a dual-path supervision model is integrated. The processing path corresponding to the main prediction feature map is defined as the main supervision branch, and the processing path corresponding to the auxiliary prediction feature map is defined as the auxiliary supervision branch. The two branches are structurally linked to form an initial semantic segmentation model with dual-path supervision.

[0117] Specifically, the system defines the processing path corresponding to the main predicted feature map as the main supervision branch. The input of the main supervision branch is the high-resolution decoded feature map, and the output is the main predicted feature map, responsible for outputting the final pixel-level classification result. The processing path corresponding to the auxiliary predicted feature map is defined as the auxiliary supervision branch. The input of the auxiliary supervision branch is the filtered intermediate layer feature map, and the output is the auxiliary predicted feature map, responsible for providing additional gradient supervision during the training phase. To achieve structural correlation and feature collaboration between the two branches, the system establishes a connection between them through a newly created branch correlation formula:

[0118]

[0119] in, The fusion feature after branch association, The main supervisory branch correlation weight matrix, To assist in supervising the branch association weight matrix, To assist in predicting feature maps, the system uses this association formula to achieve feature interaction between the main and auxiliary supervision branches, ensuring that the two branches optimize collaboratively during training and improving the model's training performance and classification accuracy. The system integrates the main supervision branch, auxiliary supervision branch, and branch association structure to form an initial semantic segmentation model with dual-path supervision. This model includes a complete processing flow from image input to predicted output, providing a basic model architecture for subsequent model training and optimization.

[0120] In one embodiment, S3 of the remote sensing image land cover classification method based on deep learning provided by the present invention specifically includes the following steps:

[0121] S31: Input the normalized image sub-map set into the initial semantic segmentation model, process it through the encoder-decoder backbone network path, output the main segmentation prediction map, and process it through the auxiliary supervision branch path to output the auxiliary prediction map, outputting the corresponding main prediction feature map and auxiliary prediction feature map.

[0122] Specifically, the system inputs a normalized image sub-map set into the initial semantic segmentation model. This sub-map is processed through the encoder-decoder backbone network path, outputting a main segmentation prediction map. An auxiliary prediction map is then output through the auxiliary supervision branch path, resulting in corresponding main and auxiliary prediction feature maps. The system first reads the normalized image sub-map set and performs batch segmentation on all sub-maps according to preset batch parameters. After segmentation, each batch of sub-maps is sequentially input into the initial semantic segmentation model, ensuring an orderly input process and no sample omissions.

[0123] After the image sub-image enters the encoder-decoder backbone network path, it sequentially undergoes feature encoding operations through a pre-defined hierarchical Transformer encoder to extract multi-scale features from the image sub-image. Then, through feature aggregation and upsampling operations, the spatial resolution of the feature map is gradually restored, ultimately outputting a high-resolution master prediction feature map. To achieve the conversion from the master prediction feature map to the master segmentation prediction map, the system uses a newly created feature transformation formula, which is as follows:

[0124]

[0125] in, Main segmentation prediction map For normalized exponential functions, For the feature transformation weight matrix, Main prediction feature map, This serves as the feature transformation bias vector. Simultaneously, the system inputs the mid-layer features from the encoder into the auxiliary supervision branch path. After processing through the branch's built-in convolution and upsampling operations, it outputs an auxiliary prediction feature map, which is then generated using the same feature transformation method. The system sequentially performs the above forward propagation processing on all samples in each training batch, ensuring that the main segmentation prediction map, auxiliary prediction map, and corresponding main prediction feature map and auxiliary prediction feature map are aligned. Figure 1 One-to-one correspondence.

[0126] S32: Based on the cross-entropy loss function, loss calculation is performed on the main predicted feature map, auxiliary predicted feature map and category index label sub-map set. The difference between the main predicted feature map and the corresponding real label of each sample is calculated as the main segmentation loss, and the difference between the auxiliary predicted feature map and the corresponding real label is calculated as the auxiliary loss. The composite loss value of the current training batch is generated by combining the preset positive decimal weights.

[0127] Specifically, the system first reads the category index label sub-map and accurately matches the true label of each sample with the corresponding main prediction feature map and auxiliary prediction feature map according to the sample identifier. This ensures that the predicted data and the true label are completely consistent in spatial location and sample attribution, avoiding inaccurate loss calculation due to matching bias. To optimize the accuracy of loss calculation and accommodate unlabeled pixels that may exist in the category index labels, the system uses a newly created cross-entropy loss optimization formula. The optimized formula is as follows:

[0128]

[0129] in, This is the optimized cross-entropy loss value. This represents the total number of pixels in a single sample. Let the loss weight be the i-th pixel. Let i be the prediction result for the i-th pixel. Let i be the true label of the i-th pixel. This is the basic cross-entropy loss calculation function. The system calculates the main segmentation loss and auxiliary loss for each sample using this optimized formula. The main segmentation loss reflects the difference between the main predicted feature map and the true label, while the auxiliary loss reflects the difference between the auxiliary predicted feature map and the true label. The system calls preset positive decimal weights and performs a composite loss calculation according to a preset formula:

[0130]

[0131] in, This is the composite loss value. and Both represent the cross-entropy loss function; and These represent the prediction results output by the main supervision branch and the prediction results output by the auxiliary supervision branch of the model, respectively. A subgraph representing the actual category index labels; It is a preset auxiliary loss weight coefficient between 0 and 1, used to adjust the contribution of auxiliary supervision to the total loss.

[0132] S33: Based on the adaptive moment estimation algorithm, the composite loss value is backpropagated and updated to process the model parameters. The gradient of the composite loss value with respect to all trainable parameters of the model is calculated. The position encoding parameters and layer normalization parameters in the model are exempted from weight decay penalty. The gradient information is used to update the model weights to minimize the loss and generate a semantic segmentation model with updated parameters.

[0133] Specifically, the system calculates the gradient of the composite loss value with respect to all trainable parameters of the model based on the chain rule of differentiation. The gradient value corresponding to each trainable parameter reflects the degree to which the change of that parameter affects the composite loss value. The gradient calculation process covers all trainable parameters of the encoder, decoder, main supervision branch, and auxiliary supervision branch of the model. To avoid abnormal gradient fluctuations affecting the model update performance, the system processes the calculated gradients using a newly created gradient constraint formula, which is as follows:

[0134]

[0135] in, The gradient vector after constraints, This is the original gradient vector. The L2 norm of the original gradient vector. The system uses a preset gradient constraint threshold. It exempts positional encoding parameters and layer normalization parameters from weight decay penalties, only applying weight decay to other trainable parameters such as convolutional and fully connected layers. This avoids over-adjustment of positional encoding information and layer normalization parameters, ensuring the stability of model feature extraction. The system inputs the constrained gradient information into an adaptive moment estimation algorithm. This algorithm adaptively adjusts the update step size of each trainable parameter, dynamically optimizing the update strategy based on the first and second moment estimates of the parameter gradient, iterating the model parameters towards minimizing the composite loss value. The system processes all gradient information in the current training batch sequentially, updating all trainable parameters of the model and generating a semantic segmentation model with updated parameters, gradually improving the model's prediction accuracy.

[0136] S34: Based on the polynomial decay strategy combined with linear preheating, the semantic segmentation model is dynamically scheduled. In the initial stage of training, the learner's learner rate is linearly increased from zero to a preset peak value, and then gradually decayed from the peak value according to the power rule during the training stage. The step size of parameter updates is dynamically adjusted to control the model convergence process and generate a trained land cover classification model.

[0137] Specifically, the system first calls preset learning rate-related parameters, including peak learning rate, number of linear warm-up iterations, total number of training iterations, and polynomial decay power. Based on these parameters, a dynamic learning rate scheduler is constructed to achieve automated dynamic adjustment of the learning rate. To precisely control the change process of the learning rate, the system unfolds the scheduling logic through a newly created refined learning rate calculation formula, which is as follows:

[0138]

[0139] in, Let be the learning rate for the t-th iteration. Preset a peak value for the learning rate. This represents the current training iteration number. For the number of linear preheating iterations, This represents the total number of training iterations. The learning rate is determined by a polynomial decay power rule. In the initial training phase, the system linearly increases the optimizer's learning rate from zero to a preset peak value according to this formula, avoiding problems such as non-convergence and parameter oscillations caused by an excessively high learning rate in the early stages of training. In the later stages of training, the learning rate gradually decays from the peak value according to a preset power rule, dynamically reducing the step size of parameter updates, allowing the model to gradually converge to the optimal parameter state in the later stages of training. The system continuously and iteratively executes learning rate scheduling, parameter backpropagation, and update operations until the model iterations reach the preset total number of training iterations, or the composite loss value stabilizes and no longer decreases significantly. At this point, all training operations are stopped, generating a trained land cover classification model capable of high-precision land cover classification of remote sensing imagery.

[0140] In one embodiment, S5 of the remote sensing image land cover classification method based on deep learning provided by the present invention specifically includes the following steps:

[0141] S51: Input the test image sub-maps into the land cover classification model in batches for forward inference processing, so that each image sub-map passes through the encoder-decoder structure of the model and outputs the category prediction matrix of each sub-map. Each element of the category prediction matrix represents the predicted land cover category index of the corresponding pixel.

[0142] Specifically, the system reads test image sub-maps and batch-processes all test image sub-maps to ensure that the size and number of channels of the sub-maps are consistent with the input specifications during model training, avoiding inference failures or accuracy deviations due to input specification mismatches. The system then batch-inputs the processed test image sub-maps into the land cover classification model. After each image sub-map enters the model, it sequentially performs feature encoding through a pre-set hierarchical Transformer encoder to extract multi-scale semantic and detail features of the sub-map. Then, it uses a pyramid pooling decoder for feature fusion and upsampling to gradually restore the feature map resolution.

[0143] To verify the reliability of the prediction results, the system uses a preset prediction confidence verification formula, which is as follows:

[0144]

[0145] in, Let be the prediction confidence score for the i-th pixel. This is the class probability distribution vector for this pixel. It is the maximum value in the probability distribution. This represents the total number of land cover categories. This represents the probability that the pixel belongs to the k-th land cover class. The system filters valid prediction results based on confidence levels and finally outputs a class prediction matrix for each sub-image. The matrix size is the same as the input sub-image. Figure 1 This corresponds to the predicted land cover category index for each element's pixel.

[0146] S52: Spatial positioning of the category prediction matrix of each sub-image is performed based on the coordinate file. According to the coordinates recorded in the original image for each sub-image, the category prediction matrix is ​​placed at the corresponding position on a blank canvas with all zeros and the same size as the original image to generate a complete prediction map.

[0147] Specifically, the system reads the coordinate file, extracts the unique identifier and corresponding spatial starting coordinates of each test image sub-image, and establishes a one-to-one correspondence between the sub-image identifier and the spatial coordinates to ensure accurate coordinate matching. The system creates a blank canvas with all zeros, whose spatial dimensions are exactly the same as the original remote sensing image to be classified. Each pixel on the canvas has an initial value of zero and is used to carry the category prediction results for each sub-image. To achieve accurate spatial positioning of the category prediction matrix, the system uses a newly created coordinate mapping formula:

[0148]

[0149] in, This represents the target coordinates of a pixel in the category prediction matrix on the blank canvas. These are the spatial starting coordinates of the subgraph. This represents the relative coordinates of the pixel within the category prediction matrix. For each sub-image, the system, based on the coordinate mapping formula, precisely places each pixel in its category prediction matrix into the corresponding position on the blank canvas according to the mapping relationship between its relative coordinates and the starting coordinates, covering the corresponding zero-value pixels on the canvas. The system sequentially performs the above spatial positioning and placement operations on the category prediction matrices of all test image sub-images to ensure that the prediction result of each sub-image accurately corresponds to the spatial position of the original image. After placement, a preliminary complete prediction map is generated, which covers the entire range of the original image, but multiple prediction results exist in the overlapping stitching area.

[0150] S53: Arbitrate the pixel categories of the overlapping stitched areas in the complete prediction map. For pixel locations that are covered by the prediction results of multiple sub-images in the final stitched map due to overlapping cropping, calculate the mode of multiple prediction category indices corresponding to the pixel location, determine the final category of the pixels in the overlapping stitched areas, and generate a high-resolution pixel-level land cover classification map that covers the entire original image range.

[0151] Specifically, the system arbitrates the pixel categories of overlapping stitched areas in the complete prediction map. For pixel locations covered by predictions from multiple sub-images in the final stitched map due to overlapping cropping, the system calculates the mode of multiple prediction category indices corresponding to the pixel location to determine the final category of the pixels in the overlapping stitched area, generating a high-resolution pixel-level land cover classification map covering the entire original image range. The system first scans the initially generated complete prediction map to identify all overlapping stitched areas. Pixels in these areas are covered by predictions from multiple sub-images and have multiple different category indices, requiring arbitration to determine the final category.

[0152] To enhance the reasonableness of the arbitration result, the system uses a preset weighted mode arbitration formula, which is as follows:

[0153]

[0154] in, For candidate category indexing, The number of sub-images covering this pixel. Let be the prediction confidence of the pixel corresponding to the m-th sub-image. This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0155] Furthermore, for each overlapping pixel, the system extracts all predicted class indices and corresponding confidence scores covering that pixel. A weighted mode is calculated using this formula, and the class index corresponding to the weighted mode is determined as the final class of that pixel. For pixels in non-overlapping areas, the system directly retains their original predicted class indices. The system sequentially performs arbitration processing on all overlapping pixels in the complete prediction map to eliminate class conflicts in overlapping areas. After processing, a high-resolution pixel-level land cover classification map covering the entire original image area is generated. Each pixel in this map corresponds to a unique land cover class index, clearly reflecting the land cover distribution in the original image.

[0156] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0157] Based on the same inventive concept, this application also provides a deep learning-based remote sensing image land cover classification system for implementing the aforementioned deep learning-based remote sensing image land cover classification method. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more deep learning-based remote sensing image land cover classification system embodiments provided below can be found in the limitations of the deep learning-based remote sensing image land cover classification method described above, and will not be repeated here.

[0158] Preferably, such as Figure 2 As shown, this invention provides a remote sensing image land cover classification system 600 based on deep learning, which is configured with the following modules:

[0159] The remote sensing image preprocessing and cropping module 610 is used to perform statistical normalization on the acquired original multispectral remote sensing images and corresponding RGB color label images based on preset quantiles, and to perform overlapping grid cropping on the large-size images and label images to generate normalized image sub-atlases and category index label sub-atlases.

[0160] The dual-path supervised segmentation model building module 620 is used to extract multi-scale features through a hierarchical Transformer encoder based on a sliding window attention mechanism and a normalized image sub-atlas, and to fuse contextual information using a pyramid pooling decoder. At the same time, an auxiliary supervision branch based on mid-level features is introduced to build an initial semantic segmentation model with dual-path supervision.

[0161] The segmentation model iterative training optimization module 630 is used to input the normalized image sub-map set and the category index label sub-map set into the semantic segmentation initial model for iterative optimization. It calculates the composite loss of the main segmentation loss and the weighted auxiliary loss, and uses an optimizer with weighted decay of location parameters and a segmented change learning rate strategy to backpropagate and update the network parameters, and outputs the trained land cover classification model.

[0162] The test image preprocessing coordinate recording module 640 is used to perform statistical normalization on the original remote sensing images to be classified based on preset quantiles and to perform overlapping grid clipping, record the spatial starting coordinates of each clipped sub-image, and generate a test image sub-image set for model inference and a coordinate file that records the spatial position information of each sub-image.

[0163] The ground cover classification result inference and stitching module 650 is used to input the test image sub-map set into the ground cover classification model for forward propagation calculation to obtain the category prediction matrix of each sub-map, and stitch and align all the sub-map prediction results together with the spatial location information recorded in the coordinate file to generate a high-resolution pixel-level ground cover classification map covering the entire original image range.

[0164] Preferably, the remote sensing image preprocessing and cropping module 610 provided in this application is configured with the following units:

[0165] The image band normalization processing unit is used to truncate and linearly stretch the pixel values ​​of each band of the acquired original multispectral remote sensing image based on preset high and low quantiles, calculate the preset high and low quantiles of the pixel values ​​of each band of the image, and map the pixel values ​​to the target value range based on the actual distribution range of the band pixel values ​​to generate normalized data for each band.

[0166] The label image category index conversion unit is used to perform color-to-category index mapping processing on the acquired RGB color label image. Based on the predefined RGB color values ​​of local object category standards, it matches the RGB vector of each pixel in the label image to the closest category and converts it into a single numerical index. Based on the preset unlabeled category identifier, it assigns an unlabeled identifier to the unmatched pixels and generates a single-channel category index image.

[0167] The image label synchronous cropping unit is used to synchronously segment the normalized data and category index map of each band. Based on a sliding window of preset size, it crops the original multispectral remote sensing image and label map with a set step size to generate normalized image sub-atlas and category index label sub-atlas.

[0168] Preferably, the dual-path supervised segmentation model construction module 620 provided in this application is configured with the following units:

[0169] The image patch embedding sequence generation unit is used to divide each image sub-image in the normalized image sub-image set into regular image patches based on a non-overlapping window of a preset size, and convert each two-dimensional image sub-image into a series of one-dimensional image patch embedding vector sequences to generate an image patch sequence to be encoded.

[0170] The hierarchical Transformer feature encoding unit is used to perform hierarchical feature encoding on image patch sequences based on sliding window multi-head self-attention mechanism and shift window multi-head self-attention mechanism. The image patch sequence is sequentially passed through multiple hierarchical Transformer encoders composed of window multi-head self-attention layer, shift window multi-head self-attention layer and feedforward network layer to perform feature interaction and transformation, generating multi-scale encoded feature maps with decreasing resolution, increasing number of channels and containing long-distance contextual semantic information.

[0171] The multi-scale feature aggregation enhancement unit is used to aggregate multi-scale context information of multi-scale encoded feature maps based on feature pyramid network technology. It performs adaptive pooling at different scales in parallel on the deepest feature map of the multi-scale encoded feature map, and then upsamples the pooling results at different scales and concatenates and fuses them with the original feature map to generate an enhanced feature map rich in multi-scale global context.

[0172] The decoding feature map resolution restoration unit is used to progressively upsample the enhanced feature map through convolutional layers and perform channel concatenation and convolutional fusion with the original encoded feature map of the multi-scale encoded feature map to progressively restore spatial details and generate a high-resolution decoded feature map for pixel-by-pixel classification.

[0173] The auxiliary supervision branch construction unit is used to construct auxiliary supervision branches based on the intermediate layer feature map of the multi-scale encoded feature map in the layer output of the hierarchical Transformer encoder. It selects the feature output of a specific stage from the intermediate layer feature map, restores the spatial resolution of the intermediate layer feature map to the input image size through independent convolutional layers and upsampling operations, and generates an auxiliary prediction feature map to provide additional gradient supervision during the training phase.

[0174] The main supervision branch construction unit is used to construct the main supervision branch based on the high-resolution decoded feature map. It performs linear projection of the channel dimension on the high-resolution decoded feature map, maps the number of feature channels to be consistent with the total number of land cover categories, and generates the main prediction feature map for outputting the final pixel-level classification result.

[0175] The dual-path supervised model integration unit is used to integrate the main predicted feature map and the auxiliary predicted feature map for dual-path supervised model integration. The processing path corresponding to the main predicted feature map is defined as the main supervised branch, and the processing path corresponding to the auxiliary predicted feature map is defined as the auxiliary supervised branch. The two branches are structurally linked to form a semantic segmentation initial model with dual-path supervision.

[0176] Preferably, the segmentation model iterative training and optimization module 630 provided in this application is configured with the following units:

[0177] The model prediction feature map output unit is used to input the normalized image sub-map set into the semantic segmentation initial model, process it through the encoder-decoder backbone network path, output the main segmentation prediction map, and process it through the auxiliary supervision branch path to output the auxiliary prediction map, outputting the corresponding main prediction feature map and auxiliary prediction feature map.

[0178] The composite loss calculation unit is used to perform loss calculation on the main predicted feature map, auxiliary predicted feature map and category index label sub-map based on the cross-entropy loss function. It calculates the difference between the main predicted feature map and the corresponding real label of each sample as the main segmentation loss, and the difference between the auxiliary predicted feature map and the corresponding real label as the auxiliary loss. The composite loss value of the current training batch is generated by combining the preset positive decimal weights.

[0179] The model parameter gradient update unit is used to perform backpropagation and update of model parameters on the composite loss value based on the adaptive moment estimation algorithm. It calculates the gradient of the composite loss value with respect to all trainable parameters of the model, exempts the position encoding parameters and layer normalization parameters in the model from weight decay penalty, updates the model weights using gradient information to minimize loss, and generates a semantic segmentation model with updated parameters.

[0180] The dynamic learning rate scheduling and optimization unit is used to perform dynamic learning rate scheduling on the semantic segmentation model based on a polynomial decay strategy combined with linear preheating. In the initial stage of training, the learning rate of the optimizer is linearly increased from zero to a preset peak value, and then gradually decayed from the peak value according to the power rule during the training stage. The step size of parameter updates is dynamically adjusted to control the model convergence process and generate a trained land cover classification model.

[0181] Preferably, the land cover classification result reasoning and splicing module 650 provided in this application is configured with the following units:

[0182] The test sub-image forward inference unit is used to input the test image sub-image set into the land cover classification model in batches for forward inference processing, so that each image sub-image passes through the encoder-decoder structure of the model and outputs the category prediction matrix of each sub-image. Each element of the category prediction matrix represents the predicted land cover category index of the corresponding pixel.

[0183] The prediction matrix spatial positioning unit is used to spatially locate the category prediction matrix of each sub-image according to the coordinate file. Based on the coordinates of each sub-image recorded in the original image, the category prediction matrix is ​​placed at the corresponding position of a blank canvas with all zeros and the same size as the original image to generate a complete prediction map.

[0184] The overlapping region pixel category arbitration unit is used to arbitrate the pixel category of the overlapping stitched region in the complete prediction map. For pixel positions that are covered by the prediction results of multiple sub-images in the final stitched map due to overlapping cropping, the mode of multiple prediction category indices corresponding to the pixel position is calculated to determine the final category of the pixels in the overlapping stitched region and generate a high-resolution pixel-level land cover classification map covering the entire original image range.

[0185] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described deep learning-based remote sensing image land cover classification method.

[0186] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described deep learning-based remote sensing image land cover classification method.

[0187] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0188] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0189] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all 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 method for classifying ground features in remote sensing images based on deep learning, characterized in that, Includes the following steps: S1: Perform statistical normalization on the acquired original multispectral remote sensing images and corresponding RGB color label images based on preset quantiles, and perform overlapping grid cropping on the large-size images and label images to generate normalized image sub-sets and category index label sub-sets; S2: Based on the sliding window attention mechanism and the normalized image sub-set, multi-scale features are extracted through a hierarchical Transformer encoder, and contextual information is fused using a pyramid pooling decoder. At the same time, an auxiliary supervision branch based on mid-level features is introduced to construct an initial semantic segmentation model with dual-path supervision. S3: Input the normalized image sub-set and the category index label sub-set into the semantic segmentation initial model for iterative optimization. Calculate the composite loss of the main segmentation loss and the weighted auxiliary loss, and use an optimizer with weighted decay of location parameters and a segmented learning rate strategy to backpropagate and update the network parameters, outputting the trained land cover classification model. S4: Perform statistical normalization on the original remote sensing images to be classified based on preset quantiles and perform overlapping grid cropping, record the spatial starting coordinates of each cropped sub-image, and generate a test image sub-image set for model inference and a coordinate file that records the spatial location information of each sub-image; S5: Input the test image sub-map set into the land cover classification model for forward propagation calculation to obtain the category prediction matrix of each sub-map, and combine it with the spatial location information recorded in the coordinate file to stitch and align the prediction results of all sub-maps to generate a high-resolution pixel-level land cover classification map covering the entire original image range.

2. The method according to claim 1, characterized in that, S1 includes: S11: The pixel values ​​of each band of the acquired original multispectral remote sensing image are truncated and linearly stretched based on preset high and low quantiles. The preset high and low quantiles of the pixel values ​​of each band of the image are calculated, and the pixel values ​​are mapped to the target value range based on the actual distribution range of the pixel values ​​of the bands to generate normalized data for each band. S12: Perform color-to-category index mapping on the acquired RGB color label image. Based on the predefined RGB color values ​​of local object category standards, match the RGB vector of each pixel in the label image to the closest category and convert it into a single numerical index. Based on the preset unlabeled category identifier, assign an unlabeled identifier to the unmatched pixels to generate a single-channel category index image. S13: Synchronously segment the normalized data and the category index map for each band, and crop the original multispectral remote sensing image and label map with a set step size based on a sliding window of a preset size to generate a normalized image sub-atlas and a category index label sub-atlas.

3. The method according to claim 2, characterized in that, The formula for calculating the normalized data is: in, These are the normalized pixel values. For the original image in pixel coordinates The original pixel value at that location, and These represent the preset lower quantile and higher quantile cutoff boundaries to eliminate the effects of noise and extreme values, respectively. and These represent the minimum and maximum values ​​of the target numerical range after normalization, respectively.

4. The method according to claim 1, characterized in that, S2 includes: S21: Based on a non-overlapping window of a preset size, each image sub-image in the normalized image sub-image set is divided into regular image blocks, and each two-dimensional image sub-image is converted into a series of one-dimensional image block embedding vector sequences to generate an image block sequence to be encoded; S22: Based on the sliding window multi-head self-attention mechanism and the shift window multi-head self-attention mechanism, the image block sequence is hierarchically feature encoded, so that the image block sequence is sequentially passed through multiple hierarchical Transformer encoders composed of window multi-head self-attention layer, shift window multi-head self-attention layer and feedforward network layer to perform feature interaction and transformation, generating multi-scale encoded feature maps with decreasing resolution, increasing number of channels and containing long-distance contextual semantic information; S23: Based on the feature pyramid network technology, multi-scale context information is aggregated on the multi-scale encoded feature map. Adaptive pooling at different scales is performed in parallel on the deepest feature map of the multi-scale encoded feature map. The pooling results at different scales are then upsampled and concatenated with the original feature map to generate an enhanced feature map rich in multi-scale global context. S24: The enhanced feature map is progressively upsampled through a convolutional layer and then channel-wise concatenated and fused with the original encoded feature map of the multi-scale encoded feature map to gradually restore spatial details and generate a high-resolution decoded feature map for pixel-wise classification. S25: Based on the intermediate layer feature map of the multi-scale encoded feature map in the layer output of the hierarchical Transformer encoder, construct an auxiliary supervision branch, select feature output of a specific stage from the intermediate layer feature map, restore the spatial resolution of the intermediate layer feature map to the input image size through independent convolutional layers and upsampling operations, and generate an auxiliary prediction feature map for providing additional gradient supervision during the training stage. S26: Based on the high-resolution decoded feature map, perform main supervision branch construction processing, perform linear projection of the channel dimension on the high-resolution decoded feature map, map the number of feature channels to be consistent with the total number of land cover categories, and generate a main prediction feature map for outputting the final pixel-level classification result; S27: Based on the main predicted feature map and the auxiliary predicted feature map, a dual-path supervised model is integrated. The processing path corresponding to the main predicted feature map is defined as the main supervised branch, and the processing path corresponding to the auxiliary predicted feature map is defined as the auxiliary supervised branch. The two branches are structurally linked to form the initial semantic segmentation model with dual-path supervision.

5. The method according to claim 1, characterized in that, S3 includes: S31: Input the normalized image sub-map set into the semantic segmentation initial model, process it through the encoder-decoder backbone network path, output the main segmentation prediction map, and process it through the auxiliary supervision branch path to output the auxiliary prediction map, outputting the corresponding main prediction feature map and auxiliary prediction feature map. S32: Based on the cross-entropy loss function, perform loss calculation processing on the main predicted feature map, the auxiliary predicted feature map and the category index label sub-map set. Calculate the difference between the main predicted feature map and the corresponding real label for each sample as the main segmentation loss, and the difference between the auxiliary predicted feature map and the corresponding real label as the auxiliary loss. Combine the preset positive decimal weights to perform superposition calculation to generate the composite loss value for the current training batch. S33: Based on the adaptive moment estimation algorithm, the composite loss value is backpropagated and updated to process the model parameters. The gradient of the composite loss value with respect to all trainable parameters of the model is calculated. The position encoding parameters and layer normalization parameters in the model are exempted from weight decay penalty. The gradient information is used to update the model weights to minimize the loss and generate a semantic segmentation model with updated parameters. S34: Based on the polynomial decay strategy combined with linear preheating, the semantic segmentation model is dynamically processed by learning rate scheduling. In the initial stage of training, the learning rate of the optimizer is linearly increased from zero to a preset peak value, and gradually decayed from the peak value according to the power rule during the training stage. The step size of parameter update is dynamically adjusted to control the model convergence process and generate a trained land cover classification model.

6. The method according to claim 5, characterized in that, The formula for calculating the composite loss value is as follows: in, This is the composite loss value. and Both represent the cross-entropy loss function; and These represent the prediction results output by the main supervision branch and the prediction results output by the auxiliary supervision branch of the model, respectively. A subgraph representing the actual category index labels; It is a preset auxiliary loss weight coefficient between 0 and 1, used to adjust the contribution of auxiliary supervision to the total loss.

7. The method according to any one of claims 1-6, characterized in that, S5 includes: S51: The test image sub-map set is input into the land cover classification model in batches for forward inference processing, so that each image sub-map is processed by the encoder-decoder structure of the model to output the category prediction matrix of each sub-map, and each element of the category prediction matrix represents the predicted land cover category index of the corresponding pixel. S52: Spatial positioning of the category prediction matrix of each sub-image according to the coordinate file; based on the coordinates of each sub-image recorded in the original image, the category prediction matrix is ​​placed at the corresponding position of a blank canvas with all zeros and the same size as the original image to generate a complete prediction map. S53: Arbitrate the pixel categories of the overlapping stitched areas in the complete prediction map. For pixel locations that are covered by multiple sub-image prediction results in the final stitched map due to overlapping cropping, calculate the mode of multiple prediction category indices corresponding to the pixel location, determine the final category of the pixels in the overlapping stitched areas, and generate a high-resolution pixel-level land cover classification map covering the entire original image range.

8. A remote sensing image land cover classification system based on deep learning, characterized in that, The system includes: The remote sensing image preprocessing and cropping module is used to perform statistical normalization on the acquired original multispectral remote sensing images and corresponding RGB color label images based on preset quantiles, and to perform overlapping grid cropping on the large-size images and label images to generate normalized image sub-atlases and category index label sub-atlases. The dual-path supervised segmentation model construction module is used to extract multi-scale features through a hierarchical Transformer encoder based on the sliding window attention mechanism and the normalized image sub-set, and to fuse contextual information using a pyramid pooling decoder. At the same time, an auxiliary supervision branch based on mid-level features is introduced to construct an initial semantic segmentation model with dual-path supervision. The segmentation model iterative training optimization module is used to input the normalized image sub-map set and the category index label sub-map set into the semantic segmentation initial model for iterative optimization. It calculates the composite loss of the main segmentation loss and the weighted auxiliary loss, and uses an optimizer with weighted decay of location parameters and a segmented change learning rate strategy to backpropagate and update the network parameters, and outputs the trained land cover classification model. The test image preprocessing coordinate recording module is used to perform statistical normalization on the original remote sensing images to be classified based on preset quantiles and to perform overlapping grid cropping, record the spatial starting coordinates of each cropped sub-image, and generate a test image sub-image set for model inference and a coordinate file that records the spatial location information of each sub-image. The land cover classification result inference and stitching module is used to input the test image sub-map set into the land cover classification model for forward propagation calculation to obtain the category prediction matrix of each sub-map, and stitch and align all the sub-map prediction results together with the spatial location information recorded in the coordinate file to generate a high-resolution pixel-level land cover classification map covering the entire original image range.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images