Intelligent bare soil identification method and system for high-resolution remote sensing image

Abstract

The invention discloses a bare soil intelligent identification method and system of a high-resolution remote sensing image, and relates to the technical field of ecological environment remote sensing monitoring, high-resolution remote sensing image data of a to-be-identified area is acquired and preprocessed, and preprocessed high-resolution remote sensing image data is obtained; performing filtering optimization on the preprocessed high-resolution remote sensing image data by adopting a dynamic adaptive multi-temporal filtering optimization method to obtain an optimized high-resolution remote sensing image data image; cutting the optimized high-resolution remote sensing image data, extracting contour position coordinates of a mask image, and constructing a training set; and inputting the training set into a bare soil recognition model for training to obtain a trained bare soil recognition model, and recognizing the bare soil plaques of the to-be-recognized region based on the trained bare soil recognition model. The multi-scale bare soil plaques under a complex background are accurately and automatically identified under a relatively low computing resource demand, and the distinction degree between the bare soil and other ground features is enhanced.

Description

Technical Field

[0001] This invention relates to the field of ecological environment remote sensing monitoring technology, and in particular to a method and system for intelligent identification of bare soil from high-resolution remote sensing images. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Bare soil is a significant source of atmospheric dust and one of the main sources of ambient air pollution. Exposure to bare soil increases the concentration of particulate matter in the air, especially PM10 and PM2.5, significantly impacting air quality and potentially leading to respiratory diseases, cardiovascular problems, and other health issues. The high reflectivity of bare soil also alters the Earth's surface energy balance, potentially affecting regional and global climate patterns, and the distribution and changes in bare soil have feedback effects on global and regional climate systems. Therefore, accurately identifying and monitoring bare soil areas and analyzing their distribution and changes can guide the development of relevant remediation measures, helping to control and mitigate air pollution. Furthermore, it helps clarify the energy exchange processes in the surface-atmosphere system, improving the accuracy of weather forecasts and climate predictions, and is of great significance to environmental governance and atmospheric environmental science.

[0004] In recent years, satellite remote sensing technology has developed rapidly, providing an efficient and objective data source for bare soil identification. Compared with traditional bare soil identification information sources such as on-site photos, video footage, and surveillance images, satellite remote sensing technology has the advantages of wide coverage, short revisit cycles, and large information volume. However, different remote sensing data sources and different technical solutions can significantly affect the efficiency, accuracy, and identification results of bare soil identification. Traditional remote sensing data sources include medium-resolution remote sensing satellites such as the Landsat series and Sentinel series. However, low-to-medium resolution remote sensing images based on bare soil indices can limit the accuracy of bare soil identification and affect the identification results. In addition, traditional methods for identifying ground features mainly rely on manual visual interpretation of remote sensing images, using expert experience to delineate and classify the boundaries of bare soil areas. However, traditional visual interpretation has significant limitations: low efficiency, high cost, and limited accuracy. It relies on personal experience and is prone to missed or incorrect identifications. Taking bare soil identification as an example, visual interpretation has limited ability to distinguish fragmented bare soil or ground features with similar spectral characteristics to bare soil. Therefore, it is necessary to utilize high-resolution satellite remote sensing imagery and leverage next-generation information technologies such as deep learning to achieve accurate and automatic identification and analysis of bare soil.

[0005] Existing technologies disclose a method for identifying bare soil, which utilizes camera images and filters the images by setting a confidence threshold to extract bare soil features, obtaining a binarized image and bare soil judgment results. This method uses high-resolution camera photos for bare soil identification, making it difficult to achieve full urban coverage. Existing technologies also disclose a method for extracting surface water bodies based on an improved SegFormer network and remote sensing imagery. This method utilizes Landsat medium-resolution multispectral imagery, and through optimized combinations of different bands, employs a parallel approach of SegFormer and CNN to establish a water body identification model, enabling the identification of water bodies such as lakes, reservoirs, and relatively wide rivers. This method, based on medium-resolution multi-band spectral information, cannot meet the needs of fine-scale object identification in three-band high-resolution imagery. Finally, existing technologies disclose a remote sensing intelligent detection method for rural homesteads based on an improved SegFormer network. This method utilizes high-resolution remote sensing imagery, employing patch embedding blocks and patch merging to split and scale the image, achieving the identification and extraction of rural homesteads. This method can identify homesteads of similar size in rural areas with a single land use type (mainly forest land and grassland), but it is difficult to effectively identify various types of bare soil such as construction sites, storage yards, and farmland of different sizes in complex terrain backgrounds in urban areas, suburbs, and rural areas. Summary of the Invention

[0006] To overcome the shortcomings of the existing technologies, this invention provides a method and system for intelligent bare soil identification from high-resolution remote sensing images. Based on the improved SegFormer algorithm, by incorporating spatial attention and channel attention mechanisms and adopting an adaptive feature fusion strategy, the computational load and parameter load are significantly reduced. In addition, an innovative dynamic adaptive multi-temporal filtering optimization method is designed to further improve the accuracy and robustness of bare soil identification under different seasons and weather conditions. While maintaining high accuracy, the accuracy of bare soil identification in complex backgrounds is improved, and the bare soil patches and areas of different spatial distribution types are automatically extracted.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention provides a method for intelligent identification of bare soil in high-resolution remote sensing images, comprising: Acquire high-resolution remote sensing image data of the area to be identified and preprocess it to obtain preprocessed high-resolution remote sensing image data; A dynamic adaptive multi-temporal filtering optimization method is used to filter and optimize the preprocessed high-resolution remote sensing image data to obtain the optimized high-resolution remote sensing image data. The optimized high-resolution remote sensing image data is cropped and the contour position coordinates of the mask image are extracted to construct a training set; The training set is input into the bare soil recognition model for training to obtain a trained bare soil recognition model. Based on the trained bare soil recognition model, bare soil patches in the area to be identified are identified. The bare soil recognition model is an improved SegFormer model, which includes a hierarchical Transformer encoder and an All-MLP decoder. A spatial attention module is added to each stage of the Transformer encoder, and a channel attention module is added between the MLP layers of the decoder.

[0008] In a further technical solution, the preprocessing includes geometric fine correction and radiometric correction.

[0009] A further technical solution involves using a dynamic adaptive multi-temporal filtering optimization method to filter and optimize the preprocessed high-resolution remote sensing image data. Specifically: Acquire historical high-resolution satellite remote sensing image data and its corresponding meteorological and geographic data, preprocess them, and integrate the remote sensing images with the corresponding meteorological and geographic information to construct a historical image database. A dynamic model of bare soil spectral characteristics was constructed based on a historical image database; Acquire current environmental factors and integrate them with preprocessed current-temporal high-resolution satellite remote sensing image data; Based on the dynamic model of bare soil spectral characteristics and current environmental factors, the filtering weights of each band and pixel in the current temporal image data are calculated. Multi-temporal filtering is performed on the preprocessed high-resolution remote sensing image data based on the filtering weights to obtain the optimized high-resolution remote sensing image data.

[0010] A further technical solution is that the dynamic model of the bare soil spectral characteristics adopts an autoregressive integral moving average model, expressed as:

[0011] in, For at a certain point in time The observed values, For the constant term of the model, Let the autoregressive order of the model be denoted as . The moving average order of the model. and It is a non-zero coefficient. For the error term, , For summation index indicators.

[0012] A further technical solution is that the multi-temporal filtering operation is expressed as follows:

[0013] in, This is the filtered image. For the first Image For the number of images, For pixel coordinates, This is the weight matrix.

[0014] Further technical solutions, specifically for constructing the training set, are as follows: Match the vector range of bare land and high-resolution imagery; Convert vector data into binary raster data; Cropping is performed based on edge repetition rate to obtain cropped high-resolution images and mask images, and the original image dataset is constructed. The actual pixel position of the bare land in the corresponding image is obtained based on the bare land pixel coordinates; Based on OpenCV, the contour coordinates of the mask image are extracted from the bare ground pixels; The original image dataset is divided into training and testing sets according to a set ratio.

[0015] A further technical solution involves optimizing the bare soil identification model by combining pixel contrast loss and cross-entropy loss functions.

[0016] Secondly, the present invention provides a bare soil intelligent identification system for high-resolution remote sensing images, comprising: The data processing module is configured to: acquire high-resolution remote sensing image data of the area to be identified, and preprocess it to obtain preprocessed high-resolution remote sensing image data; The filtering optimization module is configured to use a dynamic adaptive multi-temporal filtering optimization method to filter and optimize the preprocessed high-resolution remote sensing image data to obtain the optimized high-resolution remote sensing image data. The training set construction module is configured to: crop the optimized high-resolution remote sensing image data and extract the contour position coordinates of the mask image to construct the training set; The model building and recognition module is configured to: input the training set into the bare soil recognition model for training to obtain a trained bare soil recognition model; and identify bare soil patches in the area to be identified based on the trained bare soil recognition model. The bare soil recognition model is an improved SegFormer model, which includes a hierarchical Transformer encoder and an All-MLP decoder. A spatial attention module is added to each stage of the Transformer encoder, and a channel attention module is added between the MLP layers of the decoder.

[0017] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for intelligent identification of bare soil in high-resolution remote sensing images as described in the first aspect.

[0018] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the intelligent bare soil identification method for high-resolution remote sensing images as described in the first aspect.

[0019] The above one or more technical solutions have the following beneficial effects: This invention employs an improved SegFormer model for intelligent bare soil identification from high-resolution remote sensing images. It utilizes a novel position-encoding-free hierarchical transformer encoder and a lightweight All-MLP decoder design, enabling the output of multi-scale features without relying on position encoding. This avoids performance degradation when the test resolution differs from the training resolution. Furthermore, this design avoids complex decoder structures. By aggregating information from different layers and combining local and global attention, it renders powerful feature representations, effectively extracting bare soil patches of different spatial distribution types.

[0020] This invention, based on the existing SegFormer model, incorporates spatial and channel attention mechanisms to enhance the model's ability to focus on important spatial and channel information, improving segmentation accuracy for images with complex backgrounds and rich details. Simultaneously, a dynamic adaptive multi-temporal filtering optimization method constructs a dynamic model of bare soil spectral characteristics and adjusts filter weights in real time using meteorological and geographic information, achieving intelligent fusion of multi-temporal images. This method effectively suppresses the influence of vegetation, shadows, and other interference factors on bare soil spectra, demonstrating significant advantages over traditional single-temporal processing in areas with complex terrain and significant seasonal variations in vegetation cover. Furthermore, data-driven dynamic weight calculation reduces reliance on manual parameter tuning, improving the model's generalization ability across different environments.

[0021] High-resolution imagery data has the advantage of high spatial resolution. This invention adopts an adaptive feature fusion strategy to dynamically adjust the feature fusion method and parameters, which can capture ground feature details with superior accuracy, greatly improving the accuracy and precision of bare soil identification. This provides a scientific basis for policy and measure formulation, meets the accuracy and practicality requirements of specific applications, and provides an effective solution for fields such as remote sensing bare soil identification. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a flowchart of a method for intelligent identification of bare soil in high-resolution remote sensing images according to an embodiment of the present invention; Figure 2 These are the steps for constructing the original bare soil image dataset in this embodiment of the invention; Figure 3 This is a comparison between the bare soil identification results of the embodiments of the present invention and the actual label data. Detailed Implementation

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, 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.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0027] Example 1 like Figure 1 As shown in the figure, this embodiment discloses a method for intelligent identification of bare soil from high-resolution remote sensing images. The method includes the following steps: S1: Acquire high-resolution remote sensing image data of the area to be identified and preprocess it to obtain preprocessed high-resolution remote sensing image data; In this embodiment, the high-resolution remote sensing image data required for intelligent bare soil identification comes from the GF-1, GF-2, GF-6 and other high-resolution series satellites. The spatial resolution of the high-resolution series remote sensing satellites is 0.6-2m for panchromatic space resolution and 2.6-8m for multispectral space resolution, which is better than the spatial resolution of most traditional remote sensing satellites.

[0028] High-resolution remote sensing image data (such as Gaofen-2 and Gaofen-5) required for intelligent bare soil identification are read and preprocessed, including geometric and radiometric correction. The bare soil data is then labeled. The distribution of bare soil in the sample data is identified through visual interpretation, and the vector ranges of the bare land and high-resolution images are matched. The vector data is then converted into binary raster data (mask images), resulting in approximately 10,000 training data labels corresponding to various land cover categories.

[0029] S2: The preprocessed high-resolution remote sensing image data is filtered and optimized using a dynamic adaptive multi-temporal filtering optimization method to obtain optimized high-resolution remote sensing image data. S201: Construct a historical image database based on historical remote sensing image data. Collect high-resolution satellite remote sensing image data of the target area covering different seasons and weather conditions over the past 5 years, totaling 1200 images. Preprocess these historical images, including radiometric correction and geometric fine correction, to ensure the accuracy and consistency of the data; use Geographic Information System (GIS) technology to integrate the image data with corresponding geographic coordinate information to construct the historical image database.

[0030] S202: Constructing a dynamic model of bare soil spectral characteristics based on a historical image database. Known bare soil sample areas are selected from preprocessed historical images, and their spectral characteristic values ​​at different bands and time phases are extracted. Time series analysis methods, such as the Autoregressive Integral Moving Average (ARIMA) model, are used to establish a dynamic model of the changes in bare soil spectral characteristics over time. The model needs to consider the influence of seasonal variations and weather conditions on the spectral characteristics of bare soil. Through learning and analysis of a large amount of historical data, the quantitative relationship between these factors and the spectral characteristics of bare soil is determined, enabling the model to accurately predict the spectral characteristics of bare soil under different environments.

[0031] Specifically, 500 known bare soil sample areas were selected from historical imagery, and their spectral characteristics were extracted across eight bands (red, green, blue, and near-infrared), four seasons (spring, summer, autumn, and winter), and three weather conditions (sunny, cloudy, and post-rain), resulting in 12,000 data sets. An ARIMA model was used to establish a dynamic model of bare soil spectral characteristics. After training, the model's root mean square error in predicting bare soil spectral characteristics under different environments was less than 0.05, accurately reflecting the influence of seasons and weather on bare soil spectra.

[0032] The dynamic model of bare soil spectral characteristics adopts the autoregressive integral moving average (ARIMA) model, which is expressed as:

[0033] in, For at a certain point in time The observed values, This is the constant term (intercept) of the model. Let the autoregressive order of the model be denoted as . The moving average order of the model. and It is a non-zero coefficient. For the error term, , For summation index indicators.

[0034] S203: Acquire and preprocess current time-phase data. Acquire current time-phase meteorological data (such as temperature, precipitation, cloud cover, etc.) and corresponding geographic information (latitude and longitude, topography, etc.). Integrate the preprocessed current time-phase remote sensing image data obtained in S1 with the meteorological data and geographic information. Specifically, obtain meteorological data from the meteorological department showing a temperature of 25℃, no precipitation, and 10% cloud cover, as well as geographic information indicating an average altitude of 50 meters and flat terrain in the area. Integrate the high-resolution remote sensing image data with the meteorological data and geographic information. Integration refers to matching remote sensing images with corresponding meteorological and geographic information to provide data support for subsequent model calculations.

[0035] S204: Based on the dynamic model of bare soil spectral characteristics and current environmental factors (time, weather, etc.), calculate the filtering weights of each band and pixel in the current temporal image data.

[0036]

[0037] in, For filtering weights, For band weights, Pixel weights.

[0038] For example, if it is currently the vegetation growing season and the weather is sunny, according to the dynamic model, the weights of bands or pixels with high correlation to the spectral characteristics of bare soil in this environment will be increased; conversely, the weights of those with low correlation will be decreased. For areas with complex terrain (such as mountains), the filtering weights are further adjusted by combining terrain slope and aspect data. On the shady slopes of mountains, due to poor lighting conditions, the spectral characteristics of bare soil may be affected by shadows. In this case, increasing the weights of bands sensitive to shadowed areas enhances the ability to identify bare soil under shadow. In open terrains such as plains, large-area continuous spectral features are utilized, and the weights of continuous areas that match the spectral characteristics of bare soil are appropriately increased to improve the identification efficiency of large-scale bare soil areas.

[0039] For example, under the current sunny summer conditions, the near-infrared band has a high correlation with the bare soil spectrum, and its weight is increased from the default 0.2 to 0.3; for continuous pixels in flat areas, if they meet the spectral characteristics of bare soil, their weight is increased by 0.1 on the original basis.

[0040] S205: Multi-temporal filtering is performed on the preprocessed high-resolution remote sensing image data based on filtering weights. Environmental feature vectors are constructed for each historical and current image. The cosine similarity between the current and historical images is calculated. Ten historical images similar to the current environment are selected. The calculated weights are applied to the current temporal image data and multiple historical temporal images for weighted fusion (multi-temporal filtering), highlighting the key features of bare soil under different environments while filtering out noise interference. Multi-temporal filtering enhances the distinguishability of bare soil from other land features. After fusion, the spectral characteristics of the bare soil area are significantly improved in distinguishing it from vegetation, buildings, and other land features. In the near-infrared band, the spectral difference between bare soil and vegetation increases from 0.15 to 0.25, providing higher-quality data for subsequent bare soil identification.

[0041] The formula for multi-phase filtering is expressed as follows:

[0042] in, Represented as the filtered image, For the first Image For the number of images, For pixel coordinates, This is the weight matrix.

[0043] S3: Cropping the optimized high-resolution remote sensing image data and extracting the contour coordinates of the mask image to construct a training set; The optimized high-resolution remote sensing image data was cropped and the coordinates were extracted to create a raw image bare soil dataset with label data. The raw dataset was then divided into a training set and a test set in an 8:2 ratio.

[0044] like Figure 2 As shown, the optimized original image and mask label data are both cropped to 640*640 pixels with an edge repetition rate of 20%. The actual pixel positions of the bare ground in the corresponding image are calculated based on the bare ground pixel coordinates. Finally, the contour position coordinates of the mask image (referring to the label image) are extracted based on the bare ground pixels using OpenCV. The extracted contour position coordinates are used for subsequent generation of thematic reports.

[0045] S301: Crop the optimized original image and mask image (labeled image) to 640*640 pixels with an edge repetition rate of 20%, and obtain the original image dataset that corresponds to the cropped image.

[0046] S302: Obtain the actual pixel position of the bare land in the corresponding image based on the bare land pixel coordinates.

[0047] S303: Extracts the contour coordinates of a mask image based on bare ground pixels using OpenCV.

[0048] S304: Divide the original image dataset into a training set and a test set according to the set ratio (8:2 in this embodiment).

[0049] S4: Input the training set into the bare soil recognition model for training to obtain a trained bare soil recognition model, and identify bare soil patches in the area to be identified based on the trained bare soil recognition model.

[0050] In this embodiment, the bare soil recognition model is constructed using an improved SegFormer algorithm, and the improved bare soil recognition model (SegFormer) is obtained through training. The improved bare soil recognition model includes a hierarchical Transformer encoder and a lightweight All-MLP decoder. For the Transformer encoder, a spatial attention module is added at each stage of the encoder (e.g., after each Transformer layer). The input of the spatial attention module is set as the output feature map of that stage, and the output is used as the input feature map of the next stage. The similarity between each position in the feature map and other positions is calculated to generate a similarity matrix. The Softmax function is applied to the similarity matrix to obtain an attention weight map. The attention weight map is applied to the original feature map to obtain a weighted feature map, which helps the model focus on important spatial regions in the feature map, reduces the interference of complex backgrounds, and improves the accuracy of segmentation. Pyramid pooling technology is used to process the deep semantic feature map extracted by the Transformer encoder to extract contextual information at different scales, and the enhanced features are adaptively fused with features at other scales.

[0051] In the spatial attention module, the calculation formula for the spatial attention mechanism is as follows:

[0052] in, For the input feature map, Here is the spatial attention weight matrix. It is a 7×7 convolutional layer. For global average pooling, For max pooling, and These represent the global average pooling feature and the max pooling feature, respectively. This is the Sigmoid activation function.

[0053] For the All-MLP decoder, a channel attention module is added between the MLP layers of the decoder. The input of the channel attention module is set as the input feature map of the MLP layer, and the output is used as the output feature map of the MLP layer. Global average pooling is performed on each channel of the feature map to obtain a channel descriptor. Channel attention weights are generated by passing through a fully connected layer and a sigmoid activation function. The channel attention weights are applied to each channel of the original feature map to obtain a weighted feature map, in which important channels are enhanced and unimportant channels are suppressed. In addition, a hybrid loss function (combining pixel contrast loss and cross-entropy loss) is used to optimize the process to better handle the problem of class imbalance.

[0054] In the channel attention module, the formula for calculating channel attention weights is as follows:

[0055] in, Here is the channel attention weight matrix. and These represent the global average pooling feature and the max pooling feature, respectively. It is the Sigmoid activation function. , This represents the weights of the two fully connected layers.

[0056] The All-MLP decoder unifies feature map resolution through bilinear interpolation and fuses multi-scale features to generate the final segmentation result. During multi-scale feature fusion, an adaptive feature fusion strategy is employed, dynamically adjusting the feature fusion method and parameters based on the specific content of the input image. Scene judgment is achieved through the model's adaptive capabilities, dynamically determining weights based on image content. In complex scenes, the weights of larger feature maps are increased to better capture detailed information; conversely, in simple scenes, the weights of smaller feature maps are increased to better capture global contextual information.

[0057] The training set is input into an improved bare soil recognition model for intelligent bare soil extraction. This model extracts information such as bare soil patch vectors, latitude and longitude, area, and location. By adding spatial and channel attention mechanisms to the existing model, its performance in image segmentation tasks can be further improved. This improvement not only helps the model better focus on important spatial and channel information but also enhances the model's segmentation accuracy for images with complex backgrounds and rich details.

[0058] In this embodiment, accuracy, precision, recall, and F1 score are used as the metrics. ScoreThe accuracy of the bare soil information identified by the SegFormer model was verified using the mean Intersection over Union (mIoU) and mean accuracy per unit area (mAP) as evaluation metrics. The specific verification is as follows: The accuracy of the SegFormer bare soil recognition model was verified using the ADE20K public dataset. One hundred test scenes were selected from the test set, and evaluation metrics such as accuracy, precision, recall, and mean precision were calculated by comparing them with corresponding labeled data. In this embodiment, the SegFormer bare soil recognition model achieved an accuracy of 0.99, a precision of 0.91, a recall of 0.89, an F1 score of 0.90, and a mean precision (mAP) of 0.90, indicating that the SegFormer model has reliable accuracy and good performance in intelligent bare soil recognition. The accuracy of the algorithm's identification results was verified using the S3 test set label data. The actual bare soil identification and label image data were compared. The comparison data included the original high-resolution satellite remote sensing image, the original high-resolution image visual interpretation binary result of bare land, the original high-resolution image bare soil identification result, and the original high-resolution image bare land identification binary result.

[0059] The comparison results of actual bare soil identification and labeled image data as described in S304 are as follows: Figure 3 As shown, the first column is the original high-resolution image, the second column is the binary image of bare soil in the original image (white represents bare soil patches), the third column is the bare soil recognition result, and the fourth column is the binary image of the bare soil recognition result. It can be seen that the algorithm model can segment bare soil patches very well, which is basically consistent with the spatial distribution of bare soil in the real label data. Moreover, the model can also have a good recognition effect on small bare soil patches.

[0060] Accuracy, Precision, Recall, F1 score The expressions for calculating the mean intersection-union ratio (mIoU) and mean accuracy (mAP) are as follows:

[0061] Wherein, TP stands for True Positive, meaning that positive classes are predicted as positive; TN stands for True Negative, meaning that negative classes are predicted as negative; FP stands for False Positive, meaning that negative classes are predicted as positive; and FN stands for False Negative, meaning that positive classes are predicted as negative.

[0062] Based on the bare soil identification results output by the bare soil identification model, a special report on bare soil monitoring is compiled. The report includes the spatial distribution of bare soil in the overall region, the proportion of different area intervals, and statistical information such as the total area, ranking, and proportion of bare soil in each district (county), which intuitively reflects the detailed information on the distribution of bare soil in the target area.

[0063] This invention combines high-resolution satellite remote sensing imagery, multi-temporal filtering methods, and advanced deep learning algorithms to accurately and automatically identify multi-scale bare soil patches in complex backgrounds with low computational resource requirements, enhance the distinction between bare soil and other land features, and generate a visualized bare soil monitoring report.

[0064] Example 2 This embodiment discloses a bare soil intelligent identification system for high-resolution remote sensing images, including: The data processing module is configured to: acquire high-resolution remote sensing image data of the area to be identified, and preprocess it to obtain preprocessed high-resolution remote sensing image data; The filtering optimization module is configured to use a dynamic adaptive multi-temporal filtering optimization method to filter and optimize the preprocessed high-resolution remote sensing image data to obtain the optimized high-resolution remote sensing image data. The training set construction module is configured to: crop the optimized high-resolution remote sensing image data and extract the contour position coordinates of the mask image to construct the training set; The model building and recognition module is configured to: input the training set into the bare soil recognition model for training to obtain a trained bare soil recognition model; and identify bare soil patches in the area to be identified based on the trained bare soil recognition model. The bare soil recognition model is an improved SegFormer model, which includes a hierarchical Transformer encoder and an All-MLP decoder. A spatial attention module is added to each stage of the Transformer encoder, and a channel attention module is added between the MLP layers of the decoder.

[0065] Example 3 The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of Embodiment 1.

[0066] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the method of Embodiment 1.

[0067] The steps and methods involved in the apparatuses of Embodiments 3 and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0068] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0070] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for intelligent recognition of bare soil in high-resolution remote sensing images, characterized in that, The method comprises the following steps: acquire high-resolution remote sensing image data of a to-be-identified region, and pre-process the high-resolution remote sensing image data to obtain pre-processed high-resolution remote sensing image data; perform filter optimization on the pre-processed high-resolution remote sensing image data by using a dynamic adaptive multi-temporal filter optimization method to obtain optimized high-resolution remote sensing image data images; crop the optimized high-resolution remote sensing image data and extract the contour position coordinates of a mask image to construct a training set; input the training set into a bare soil identification model for training to obtain a trained bare soil identification model, and identify bare soil patches in the to-be-identified region based on the trained bare soil identification model; the bare soil identification model is an improved SegFormer model, which comprises a hierarchical Transformer encoder and an All-MLP decoder, and a spatial attention module is added at each stage of the Transformer encoder and a channel attention module is added between the MLP layers of the decoder.

2. The method of claim 1, wherein the method further comprises: The pre-processing comprises geometric precise correction and radiation correction.

3. The method of claim 1, wherein the method further comprises: determining a number of pixels in the image that are classified as bare soil; and determining a percentage of the image that is classified as bare soil based on the number of pixels classified as bare soil. The filter optimization on the pre-processed high-resolution remote sensing image data by using the dynamic adaptive multi-temporal filter optimization method specifically comprises the following steps: acquire historical high-resolution satellite remote sensing image data and corresponding meteorological and geographical data, pre-process the historical high-resolution satellite remote sensing image data, integrate the remote sensing image data and the corresponding meteorological and geographical information, and construct a historical image database; construct a bare soil spectral feature dynamic model based on the historical image database; acquire current environmental factors, and integrate the current environmental factors with the pre-processed current temporal high-resolution satellite remote sensing image data; calculate filter weights of each band and pixel point in the current temporal image data based on the bare soil spectral feature dynamic model and the current environmental factors; perform multi-temporal filtering operation on the pre-processed high-resolution remote sensing image data based on the filter weights to obtain the optimized high-resolution remote sensing image data images.

4. The method of claim 3, wherein the method further comprises: The bare soil spectral feature dynamic model adopts an autoregressive integrated moving average model and is expressed as: wherein, is the observation at time point , is a constant term of the model, is the autoregressive order of the model, is the moving average order of the model, and is a coefficient not equal to zero, is an error term, , is a summation index indicator.

5. The method of claim 3, wherein the method further comprises: The multi-temporal filtering operation is expressed as: wherein, is the filtered image, is the first is the image, is the number of images, is the pixel coordinate, is the weight matrix.

6. The method of claim 1, wherein the method further comprises: The construction of the training set specifically comprises the following steps: match bare land and high-resolution image vector ranges; convert the vector data into binary raster data; crop the high-resolution image and the mask image according to an edge repetition rate to obtain cropped high-resolution images and mask images, and construct an original image data set; obtain actual pixel positions of the bare land in the corresponding images according to the bare land pixel coordinates; extract the contour position coordinates of the mask image based on OpenCV according to the bare land pixels; divide the original image data set into a training set and a test set according to a set proportion.

7. The method of claim 1, wherein the method further comprises: determining a number of pixels in the image that are classified as bare soil; and determining a percentage of the image that is classified as bare soil based on the number of pixels classified as bare soil. The bare soil identification model is optimized in combination with a pixel contrast loss and a cross-entropy loss function.

8. A system for intelligent bare soil identification of high-resolution remote sensing images, characterized in that, The method comprises the following steps: a data processing module configured to acquire high-resolution remote sensing image data of a to-be-identified region, and pre-process the high-resolution remote sensing image data to obtain pre-processed high-resolution remote sensing image data; a filter optimization module configured to perform filter optimization on the pre-processed high-resolution remote sensing image data by using a dynamic adaptive multi-temporal filter optimization method to obtain optimized high-resolution remote sensing image data images; A training set construction module configured to crop the optimized high-resolution remote sensing image data and extract the contour position coordinates of the mask image to construct a training set; A model construction and identification module configured to input the training set into a bare soil identification model for training to obtain a trained bare soil identification model, and identify the bare soil patches in the to-be-identified region based on the trained bare soil identification model; the bare soil identification model is an improved SegFormer model, including a hierarchical Transformer encoder and an All-MLP decoder, a spatial attention module is added at each stage of the Transformer encoder, and a channel attention module is added between the MLP layers of the decoder.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the bare soil intelligent identification method of high-resolution remote sensing image in any one of claims 1-7.

10. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the bare soil intelligent identification method of high-resolution remote sensing image in any one of claims 1-7.

Images