Large-scale land cover remote sensing intelligent information extraction method and extraction system based on sequence mask prediction

Abstract

The invention provides a large-scale land cover remote sensing intelligent information extraction method and system based on sequence mask prediction, and aims to improve the precision and efficiency of land cover information extraction. The method comprises the steps of preprocessing and region division, sample set generation and labeling, U-ConvHDNet model training, sequence mask prediction, model precision evaluation and the like. Through preprocessing of geometric correction, radiation correction and the like, fine region division is constructed in combination with multiple data sources. A sample set is independently manufactured and labeled, a U-ConvHDNet model is used for binary classification training, ConvNeXt, Haar wavelet down-sampling and up-sampling modules DySample are fused in the model, and deep features and multi-scale information are effectively captured. In the sequence mask prediction stage, a double receptive field seamless sliding window is adopted, and step-by-step prediction is carried out in combination with the ground feature extraction difficulty. The method and the system have the characteristics of high automation degree, high extraction precision, strong adaptability and the like, and can be widely applied to the field of large-scale land utilization mapping and monitoring.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image processing, and in particular relates to a large-scale land cover remote sensing intelligent information extraction method and extraction system based on sequence mask prediction. Background Art

[0002] Land cover refers to the comprehensive manifestation of various types of materials on the earth's surface and their natural properties and characteristics. As an important part of the earth system, it plays a vital role in global change. Changes in land cover directly affect biogeochemical cycles, including energy, water and carbon cycles from land to atmosphere, which in turn have a profound impact on climate change. Therefore, scientifically and accurately measuring and monitoring the dynamic changes of global land cover is of immeasurable value for studying the energy balance, carbon cycle, other biogeochemical cycles of the earth system, and predicting and responding to climate change.

[0003] With the rapid development of remote sensing technology, especially the advancement of satellite remote sensing technology, we can obtain high-resolution, multi-band remote sensing images, which provides unprecedented opportunities for the rapid and accurate extraction of land cover information. Remote sensing technology, with its macroscopic, real-time, periodic and comprehensive characteristics, has become the core technical means for dynamic monitoring of ecological environment and international land cover change research. Through remote sensing technology, we can achieve global coverage and real-time monitoring of land cover information, and provide scientific basis for policy making, environmental protection, resource management and disaster warning. Deep learning technology, especially convolutional neural network (CNN), has shown powerful capabilities in image recognition, classification and segmentation tasks because it can automatically learn high-level feature representations of images. In remote sensing image processing, deep learning technology is widely used in intelligent extraction of land cover information. At present, deep learning technology can identify a variety of land object types including cities, forests, farmlands, etc., and can even further refine and identify specific land objects such as buildings and roads. This refined classification capability greatly improves the accuracy and efficiency of land cover information extraction, reduces the labor of manual classification, and provides more detailed data support for land use planning and environmental management.

[0004] Due to the vast territory of my country, the surface coverage scenes are complex and unevenly distributed. When using traditional classification strategies and multi-classification models, it is difficult to obtain samples of all target categories in the region, which will lead to the problem of uneven categories. This will not only reduce the classification accuracy of the model, but also make it difficult for the model to converge. In this regard, the present invention proposes a cluster of binary classification sequence models based on sample focus, and a partitioned and layered surface cover intelligent information extraction method. The method adopts the idea of ​​"partition control, layered extraction" in geological divide-and-conquer information extraction, and divides the area to be extracted into several areas according to the geographical scene area. In each homogeneous division within the region, high-precision surface cover information extraction is achieved according to the strategy sequence of easy first and difficult later, layer by layer extraction, and layer by layer control. Summary of the invention

[0005] In view of the shortcomings of the existing intelligent information extraction of land cover remote sensing, the purpose of the present invention is to provide a large-scale land cover remote sensing intelligent information extraction method based on sequence mask prediction, the method comprising the following steps:

[0006] Step 1, preprocessing and region division:

[0007] Step 1.1: Perform preprocessing such as geometric correction and radiation correction on the image to correct the geometric distortion and spectral error in the image; adjust the spatial resolution of the image to meet the subsequent analysis requirements.

[0008] Step 1.2, based on the third national land survey data, water system data and road network data as the base map, generate masks of land features such as water bodies, glaciers, deserts, cultivated land, woodlands, grasslands, etc. On this basis, construct the global features of the samples, spatial relationship features and essential features of the samples, and subdivide these areas into different types of layers, each layer represents a specific category sample. The distribution and characteristics of the internal features of each area after partitioning remain highly consistent.

[0009] Step 2: Sample set generation and annotation:

[0010] Step 2.1, make a sample set for each partition independently. All samples in the sample set are binary samples and evenly cover all scenes on the image. The sample types in this step are divided into two categories, namely focused samples and control samples. Focused samples are the main force for capturing and extracting template features. The foreground ratio of the extracted target in each sample must be greater than 50%. Control samples are mainly used to improve the model's resistance to misextraction problems caused by complex backgrounds, and do not interfere with the foreground extraction effect. The foreground ratio in the control sample is less than or equal to 5%, and the ratio to the focused sample is less than 1:5.

[0011] In step 2.2, label the sample using a vector format file with the same coordinate system as the image.

[0012] Step 2.3: Divide the finely annotated sample set into training set, validation set, and test set in a ratio of 7:1:2. The training set is sample enhanced by random cropping, random rotation, flipping, random color change, random occlusion, etc. The purpose is to enrich the content of the training set and improve the robustness and generalization ability of the model.

[0013] Step 3, U-ConvHDNet model training:

[0014] For each category of training set, the U-ConvHDNet model is used to independently train the binary classification model; the overall architecture of U-ConvHDNet consists of three main parts: convolutional neural network ConvNeXt, Haar wavelet downsampling module (Haar Wavelet Downsampling, HWD), and upsampling module DySample.

[0015] Step 3.1, convolutional neural network ConvNeXt. Based on ResNet, ConvNeXt increases the number of stacking of convolutional layers, optimizes the neck network, introduces deep convolution in combination with ordinary convolution, and enhances the model's ability to capture deep features. ConvNeXt uses GELU as the activation function and uses layer normalization instead of batch normalization. The module randomly discards some paths through regularization to prevent the model from overfitting. Considering that the sample size after random cropping is small, the four-fold downsampling in the initial stage is further replaced with two-fold downsampling, and deconvolution is applied to replace bilinear interpolation at the end, thereby enhancing the model's ability to express the spatial detail features of the objects.

[0016] Step 3.2, the downsampling module HWD, through multi-scale decomposition, effectively extracts various frequency components of the image, thereby accurately capturing the spatial detail characteristics of the object. HWD first uses Haar wavelet transform to perform multi-scale decomposition on the input image, and decomposes the input image into a low-frequency component yL and three high-frequency components y through discrete wavelet transform (DWT). L H.y H L.y H H. Among them, the low-frequency component retains the main information of the image, the high-frequency component captures the detailed features in the image, and then the components are concatenated, and finally the features are fused through convolution.

[0017] Step 3.3, upsampling module DySample. Dynamically generate convolution kernels at each position and adaptively adjust the filter. This technology not only maintains high-quality upsampling effects, but also reduces computational overhead. DySample transforms the input feature map through a linear layer, adjusts the feature dimension, and introduces position information to enhance the accuracy of convolution kernel generation, so that the network can better understand the spatial relationship in the image. The upsampling effect is further improved by fusing the transformed feature map with the original feature map, processing the feature map through grid sampling, and fusing the low-level features with the upsampled features through jump connections, effectively retaining multi-scale information.

[0018] Step 4, sequence mask prediction:

[0019] Step 4.1, based on the seamless sliding window of dual receptive fields, the receptive field is divided into a basic receptive field and an extended receptive field. The basic receptive field is the internal area of ​​the prediction window with complete context information. The extended receptive field is the peripheral area of ​​the prediction window with incomplete context information. When performing the model sliding window prediction, the window slides with a step size of (L-2αL), and only the prediction results within the basic receptive field are retained and finally spliced ​​into the final prediction results. The inference results within the basic receptive field have an extended perception of the background context within the extended receptive field.

[0020] Step 4.2, predict each image block in turn according to the difficulty of object extraction, use the previous prediction result as the mask for the subsequent prediction, and realize the information extraction of all feature categories based on the fast mask update mechanism of bit operation. The mask is initialized to a two-dimensional matrix with the same size as the image, 1 represents that there is no prediction result at this position, and 0 represents that the position has been filled with a prediction result. All models are binary classification models, and the prediction results record the non-target object pixel value as 0 and the target object pixel value as 1. After the prediction is completed, the prediction result is first multiplied by the mask to realize the pre-mask operation. Then the mask is updated, specifically, the prediction result is negated, and then a logical AND operation is performed with the mask (~predict_result&mask), and the result is written into the mask. Finally, the prediction result is multiplied by the model serial number and added to the two-dimensional matrix recording the final prediction result.

[0021] Step 5, model accuracy evaluation, quantitatively analyzes the accuracy of the proposed surface cover remote sensing intelligent information extraction method and system to ensure that the extraction results are highly consistent with the actual surface cover. The accuracy evaluation indicators used include precision, recall, and F1 score.

[0022] A large-scale land cover remote sensing intelligent information extraction system based on sequence mask prediction, including the following modules:

[0023] The first module is a preprocessing and area division module; the functions of this module include preprocessing such as geometric correction and radiation correction; area division and masking;

[0024] The second module is the sample set generation and annotation module; the functions of this module include image cropping; interactive annotation of image blocks; random division and generation of sample sets;

[0025] The third module is the U-ConvHDNet model training module; the functions of this module include building a U-ConvHDNet model and implementing network training.

[0026] The fourth module is the sequence mask prediction module. The function of this module is to realize seamless sliding window cropping based on dual receptive fields for a high-resolution remote sensing image to be extracted, and input the binary classification model for mask prediction according to the difficulty of ground object extraction, and obtain the full-element category results of the final surface cover information extraction through bit operations;

[0027] The fifth module is the model accuracy assessment module. This module is used to assess the accuracy of the prediction results output by module four. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a technical flow chart of the intelligent information extraction method of land cover remote sensing based on zoning and stratification;

[0029] Figure 2 The present invention is a flow chart of cultivated land extraction and change detection of a convolutional neural network fused with residual correction;

[0030] Figure 3 This is a diagram of the dual receptive field seamless sliding window reasoning scheme of the present invention;

[0031] Figure 4 This is a diagram showing the implementation effect of an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to better understand the technical solution of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0033] The environment used in the embodiments of the present invention: the hardware environment is configured as Intel(R) Core(TM) i9-13980HX CPU@2.2GHz, NVIDIA GeForce RTX 4080 Laptop 12GB, 32GB RAM.

[0034] The research area of ​​the embodiment of the present invention is Xinjiang Uygur Autonomous Region of my country. The specific process is as follows Figure 1 , a large-scale land cover remote sensing intelligent information extraction method based on sequence mask prediction, the specific steps are as follows:

[0035] Step 1, preprocessing and region division:

[0036] Step 1.1: Perform preprocessing such as geometric correction and radiation correction on the image to correct the geometric distortion and spectral error in the image. Adjust the spatial resolution of the image to meet the needs of subsequent analysis.

[0037] Step 1.2, based on the third national land survey data, water system data and road network data as the base map, generate masks of land features such as water bodies, glaciers, deserts, cultivated land, woodlands, grasslands, etc. In the process of regional division, vector format files are used as the basis, combined with land feature mask data, to divide the whole Xinjiang into multiple regions. On this basis, the global features, spatial relationship features and essential features of the samples are constructed, and these regions are subdivided into different types of layers, each layer representing a specific category sample. The distribution and characteristics of the internal features of each region after partitioning remain highly consistent.

[0038] In the embodiment, the global features of the sample include geographic information, image resolution, and band information; the spatial relationship features refer to topological relationships such as separation, inclusion, and adjacency between objects; and the essential features of the sample include spectrum, texture, and shape.

[0039] Step 2: Sample set generation and annotation:

[0040] Step 2.1, make a sample set for each partition independently. All samples in the sample set are binary samples and evenly cover all scenes on the image. The sample types in this step are divided into two categories, namely focused samples and control samples. Focused samples are the main force for capturing and extracting template features. The foreground ratio of the extracted target in each sample must be greater than 50%. Control samples are mainly used to improve the model's resistance to misextraction problems caused by complex backgrounds, and do not interfere with the foreground extraction effect. The foreground ratio in the control sample is less than or equal to 5%, and the ratio to the focused sample is less than 1:5.

[0041] In step 2.2, label the sample using a vector format file with the same coordinate system as the image.

[0042] In the sample labeling process in the embodiment, the sample labeler needs to combine the background knowledge of the labeling scene to label the closed polygon along the outer contour of the target object. The labeling accuracy must reach pixel-level accuracy to minimize the cases of missed labels and mislabeling. For objects with fuzzy boundaries, the spectral threshold of the classification must be clarified first, and then the labeling is performed after pre-segmentation by the spectral threshold. For the non-target object area inside the labeled closed polygon, it is necessary to dig a hole with a digging tool.

[0043] Step 2.3: Divide the finely annotated sample set into training set, validation set, and test set in a ratio of 7:1:2. The training set is sample enhanced by random cropping, random rotation, flipping, random color change, random occlusion, etc. The purpose is to enrich the content of the training set and improve the robustness and generalization ability of the model.

[0044] In the embodiment, random cropping needs to apply different random cropping sizes according to the actual spatial scale of different land types. For example, the scale of buildings is smaller than that of grassland, and a smaller random cropping window is applied. Random color transformation is to perform random contrast and brightness adjustment on each band, specifically, to randomly add and subtract pixel values ​​of each band and to apply random linear stretching to each band. The amplitude of color transformation should not be too large, and the upper limit of the transformation amplitude should not exceed 5% of the original pixel value. Random occlusion applies multiple occlusion areas, and the interior of the occlusion area is filled with uniform random pixel values, and the total occlusion area does not exceed 20% of the sample area. The size of the enhanced training set is at least 64 times the size of the original training set.

[0045] Step 3, U-ConvHDNet model training:

[0046] For each category of training set, the U-ConvHDNet model is used to train the model independently. The U-ConvHDNet model architecture is as follows Figure 2 As shown in the figure. The U-ConvHDNet model combines multi-scale feature extraction and high-fidelity up-and-down sampling strategies for spatial details, improving the ability to recognize complex and fine structures of natural objects in remote sensing images with meter-level spatial resolution. The overall architecture of U-ConvHDNet consists of three main parts: convolutional neural network ConvNeXt, Haar wavelet downsampling module (HaarWavelet Downsampling, HWD), and upsampling module DySample.

[0047] Among them, ConvNeXt combines the advantages of ResNet and Swin Transformer, and enhances the feature extraction capability by introducing large convolution kernels and deep convolution. Based on ResNet, ConvNeXt increases the number of convolution layer stacking, optimizes the neck network, introduces deep convolution in combination with ordinary convolution, strengthens the model's ability to capture deep features, and enables the network to understand the image content more comprehensively by increasing the size of the convolution kernel and expanding the receptive field. ConvNeXt uses GELU as the activation function, replacing the ReLU in the traditional convolution network, and uses layer normalization instead of batch normalization. The module randomly discards some paths through regularization to prevent the model from overfitting. Considering that the sample size after random cropping is small, the four-fold downsampling in the initial stage is further replaced by two-fold downsampling, and deconvolution is applied to replace bilinear interpolation at the end, thereby enhancing the model's ability to express the spatial detail features of the ground object.

[0048] Among them, the downsampling module HWD can perform multi-scale decomposition and effectively extract various frequency components of the image, so as to accurately capture the spatial detail characteristics of the ground object. HWD first uses Haar wavelet transform to perform multi-scale decomposition on the input image, and decomposes the input image into a low-frequency component yL and three high-frequency components y through discrete wavelet transform (DWT). L H.y H L.y H H. Among them, the low-frequency component retains the main information of the image, the high-frequency component captures the detailed features in the image, and then the components are concatenated, and finally the features are fused through convolution.

[0049] Among them, the upsampling module DySample breaks through the limitations of traditional fixed convolution kernels, can dynamically generate convolution kernels at each position, and adaptively adjust filters. This technology not only maintains high-quality upsampling effects, but also reduces computational overhead. DySample transforms the input feature map through a linear layer, adjusts the feature dimension, and introduces position information to enhance the accuracy of convolution kernel generation, enabling the network to better understand the spatial relationship in the image. The upsampling effect is further improved by fusing the transformed feature map with the original feature map, processing the feature map through grid sampling, and fusing the low-level features with the upsampled features through jump connections, effectively retaining multi-scale information.

[0050] In the embodiment, the model training uses the Adam optimizer and the FocalLoss loss function. During the training process, the current model loss and its accuracy on the validation set are calculated after each Epoch to ensure that the learning rate is adjusted after the model converges. After the model training is completed, the accuracy needs to be verified on the test set first, and the model network weights are stored after the accuracy meets the standard. If the accuracy does not meet the standard, it is necessary to adjust the training hyperparameters or update the sample set.

[0051] Step 4, sequence mask prediction:

[0052] Step 4.1, a dual receptive field seamless sliding window solution is adopted. The schematic diagram of the solution is as follows Figure 3 As shown in the figure. The receptive field is divided into a basic receptive field and an extended receptive field. The basic receptive field is the inner area of ​​the prediction window with complete context information. The extended receptive field is the outer area of ​​the prediction window with incomplete context information. When executing the model sliding window prediction, the window slides with a step size of (L-2αL), and only the prediction results within the basic receptive field are retained and spliced ​​into the final prediction results at the end. The inference results within the basic receptive field have an extended perception of the background context within the extended receptive field, which can greatly improve the stability of the spliced ​​prediction results and effectively avoid the instability near the sliding window splicing line.

[0053] Step 4.2, predict each image block in turn according to the difficulty of object extraction, and use the previous prediction results as masks for subsequent predictions to achieve overall high-precision extraction. The invention proposes a fast mask application and update mechanism based on bit operations. The mask is initialized to a two-dimensional matrix of the same size as the image, 1 represents that there is no prediction result at this position, and 0 represents that the position has been filled with a prediction result. All models are binary classification models, and the prediction results record the non-target object pixel value as 0 and the target object pixel value as 1. After the prediction is completed, the prediction result is first multiplied by the mask to achieve the pre-mask operation. Then the mask is updated, specifically, the prediction result is inverted, and then a logical AND operation is performed with the mask (~predict_result&mask), and the result is written into the mask. Finally, the prediction result is multiplied by the model serial number and added to the two-dimensional matrix recording the final prediction result.

[0054] In this embodiment, the above operation can achieve the effect of fast updating of the billion-pixel mask within 100 milliseconds.

[0055] Step 5: Model accuracy evaluation

[0056] Model accuracy evaluation can quantitatively analyze the accuracy of the proposed land cover remote sensing intelligent information extraction method and system, ensuring that the extraction results are highly consistent with the actual land cover situation. Through accuracy evaluation, the reliability, stability and applicability of the method can be evaluated, providing a scientific basis for subsequent application and promotion.

[0057] The accuracy evaluation indicators used include precision, recall, and F1 score. The calculation formula is as follows:

[0058]

[0059] Among them, TP, TN, FP, and FN are correctly predicted positive pixels, correctly predicted negative pixels, incorrectly predicted positive pixels, and incorrectly predicted negative pixels, respectively. Precision focuses on how many pixels predicted to be of a certain category are actually correct. Recall focuses on how many pixels that actually belong to a certain category are successfully identified. The F1 score provides a balanced evaluation indicator between precision and recall, which is particularly important in the case of imbalanced categories. It avoids the shortcomings of relying solely on precision or recall.

[0060] The implementation effect of this embodiment is shown in the figure, and the accuracy evaluation results are shown in Table 1.

[0061] Table 1 Precision evaluation table of the results of this embodiment

[0062]

[0063] This patent provides a corresponding large-scale land cover remote sensing intelligent information extraction system based on sequence mask prediction, which includes the following submodules:

[0064] The first module is the preprocessing and region division module. The functions of this module include preprocessing such as geometric correction and radiation correction; region division and masking.

[0065] The second module is the sample set generation and annotation module. The functions of this module include image cropping, interactive annotation of image blocks, and random division and generation of sample sets.

[0066] The third module is the U-ConvHDNet model training module. The functions of this module include building a U-ConvHDNet model and implementing network training.

[0067] The fourth module is the sequence mask prediction module. The function of this module is to realize seamless sliding window cropping based on dual receptive fields for a high-resolution remote sensing image to be extracted, and input the binary classification model for mask prediction according to the difficulty of ground object extraction, and obtain the full-element category results of the final surface cover information extraction through bit operations.

[0068] The fifth module is the model accuracy assessment module. This module is used to assess the accuracy of the prediction results output by module four.

[0069] The implementation of each module can refer to the corresponding steps, which will not be described in detail in the present invention.

[0070] Compared with the existing methods, the advantages and positive effects of the present invention are as follows: First, the disassembly process is performed by using a binary classification recognition strategy, which greatly simplifies the complexity of sample selection, enables high-quality training samples to be generated independently for each category, and ensures that each sub-model can achieve the best recognition effect in its specific field. By focusing on feature learning of a single category, the common problem of inter-category interference in multi-category classification is avoided, and the accuracy and robustness of the model are further improved. In addition, in the sequence mask prediction stage, the prediction order is cleverly designed to follow the principle of "easy first, difficult later". This strategy not only helps the model to establish a stable cognition of simple categories in the early stage of prediction, but also can use the obtained contextual information in subsequent stages to assist in solving more complex classification problems. Thirdly, the proposed U-ConvHDNet model combines multi-scale feature extraction and high-fidelity up- and down-sampling strategies for spatial details, which improves the recognition ability of complex and fine structures of natural objects in remote sensing images.

[0071] In this way, the accuracy of the overall prediction process is effectively improved, making the entire system more efficient and accurate in the intelligent extraction of surface cover information.

Claims

1. A large-scale land cover remote sensing intelligent information extraction method based on sequence mask prediction, which is characterized by the following steps: Step 1, preprocessing and region division; Step 2: Sample set generation and annotation; Step 3, U-ConvHDNet model training; for each category of training set, the U-ConvHDNet model is used to independently train the binary classification model; the overall architecture of U-ConvHDNet consists of three main parts: convolutional neural network ConvNeXt, Haar wavelet downsampling module (Haar Wavelet Downsampling, HWD), and upsampling module DySample; Step 4, sequence mask prediction; Step 5: Model accuracy evaluation. The accuracy evaluation indicators used include precision, recall, and F1 score.

2. The large-scale land cover remote sensing intelligent information extraction method based on sequence mask prediction according to claim 1 is characterized by: Step 1 includes the following sub-steps: Step 1.1, pre-process the image by geometric correction and radiation correction to correct the geometric distortion and spectral error in the image; adjust the spatial resolution of the image to meet the needs of subsequent analysis; Step 1.2, based on the third national land survey data, water system data and road network data as the base map, generate masks of land features such as water bodies, glaciers, deserts, cultivated land, woodlands, grasslands, etc.; on this basis, construct sample global features, spatial relationship features and sample essential features, and subdivide these areas into different types of layers, each layer representing a specific category sample.

3. The large-scale land cover remote sensing intelligent information extraction method based on sequence mask prediction according to claim 1 is characterized by: Step 2 includes the following sub-steps: Step 2.1, independently make a sample set for each partition; all samples in the sample set are binary samples and evenly cover all scenes on the image; the sample types in this step are divided into two categories, namely focused samples and control samples; focused samples are the main force for capturing and extracting template features; Step 2.2, label the sample using a vector format file with the same coordinate system as the image; In step 2.3, the finely annotated sample set is divided into a training set, a validation set, and a test set in a ratio of 7:1:2, where the training set needs to be sample enhanced.

4. The large-scale land cover remote sensing intelligent information extraction method based on sequence mask prediction according to claim 1 is characterized by: Step 3 includes the following sub-steps: Step 3.1, convolutional neural network ConvNeXt; based on ResNet, ConvNeXt increases the number of convolutional layer stacking, optimizes the neck network, and introduces deep convolution in combination with ordinary convolution; ConvNeXt uses GELU as the activation function and layer normalization instead of batch normalization. The module randomly discards some paths through regularization to prevent overfitting of the model. It replaces the four-fold downsampling in the initial stage with two-fold downsampling, and applies deconvolution to replace bilinear interpolation at the end. Step 3.2, downsampling module HWD; through multi-scale decomposition, various frequency components of the image are extracted to accurately capture the spatial detail characteristics of the ground object; HWD first uses Haar wavelet transform to perform multi-scale decomposition on the input image, and decomposes the input image into a low-frequency component yL and three high-frequency components y through discrete wavelet transform (DWT). L H.y H L.y H H; Among them, the low-frequency component retains the main information of the image, the high-frequency component captures the detailed features in the image, and then the components are concatenated, and finally the features are fused through convolution; Step 3.3, upsampling module DySample; dynamically generate convolution kernels at each position and adaptively adjust the filter to maintain high-quality upsampling effect and reduce computational overhead; DySample transforms the input feature map through a linear layer, adjusts the feature dimension, and introduces position information; the upsampling effect is further improved by fusing the transformed feature map with the original feature map, processing the feature map through grid sampling, and fusing the low-level features with the upsampled features through jump connections to retain multi-scale information.

5. The large-scale land cover remote sensing intelligent information extraction method based on sequence mask prediction according to claim 1 is characterized by: Step 4 includes the following sub-steps: Step 4.1, based on the seamless sliding window of dual receptive fields, the receptive field is divided into a basic receptive field and an extended receptive field; the basic receptive field is the internal area of ​​the prediction sliding window and has complete context information; The extended receptive field is the peripheral area of ​​the prediction sliding window, and the context information is incomplete. When executing the model sliding window prediction, the sliding window slides with a step size of (L-2aL), and only the prediction results within the basic receptive field are retained and finally spliced ​​into the final prediction results. The inference results within the basic receptive field have an extended perception of the background context within the extended receptive field. Step 4.2: predict each image block in turn according to the difficulty of object extraction, use the previous prediction result as the mask for the subsequent prediction, and realize the information extraction of all feature categories based on the fast mask update mechanism based on bit operation; The mask is initialized to a two-dimensional matrix with the same size as the image. 1 represents that there is no prediction result at this position, and 0 represents that the position has been filled with a prediction result. All models are binary classification models, and the prediction results record the non-target object pixel value as 0 and the target object pixel value as 1. After the prediction is completed, the prediction result is first multiplied by the mask to achieve the pre-mask operation. Then the mask is updated, specifically by inverting the prediction result, and then performing a logical AND operation with the mask (~predict_result&mask), and the result is written into the mask. Finally, the prediction result is multiplied by the model serial number and added to the two-dimensional matrix recording the final prediction result.

6. A large-scale land cover remote sensing intelligent information extraction system based on sequence mask prediction, characterized in that: Includes the following modules: The first module is a preprocessing and area division module; the functions of this module include preprocessing such as geometric correction and radiation correction; area division and masking; The second module is the sample set generation and annotation module; the functions of this module include image cropping; interactive annotation of image blocks; random division and generation of sample sets; The third module is the U-ConvHDNet model training module; the functions of this module include building the U-ConvHDNet model and implementing network training; The fourth module is the sequence mask prediction module. The function of this module is to realize seamless sliding window cropping based on dual receptive fields for a high-resolution remote sensing image to be extracted, and input the binary classification model for mask prediction according to the difficulty of ground object extraction, and obtain the full-factor category results of the final surface cover information extraction through bit operations. The fifth module is the model accuracy assessment module. This module is used to assess the accuracy of the prediction results output by module four.

Images