A remote sensing image ground feature classification method and device based on index and coverage characteristics guidance

Abstract

The application provides a remote sensing image ground object classification method and device based on index and coverage feature guidance. The method comprises the following steps: acquiring a remote sensing spectral image; inputting the remote sensing spectral image into a trained ground object classification model to classify the ground cover of the earth's surface; the ground object classification model comprises a backbone network, an index feature extraction module, a first splicing module, a first channel attention module, a multi-scale feature extraction module, a coverage feature extraction module, a second splicing module, a second channel attention module, a first convolution module, a second convolution module, a third splicing module and a third convolution module. The method provided by the application can significantly improve the accuracy of ground object classification.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, and in particular to a method and apparatus for classifying land features in remote sensing images based on index and coverage-guided features. Background Technology

[0002] Vegetation and wetlands play vital roles in global ecosystems, serving as important habitats for many plants and animals and significantly impacting ecological balance, water purification, climate regulation, and carbon storage. However, with intensifying human activities and climate change, vegetation and wetlands face increasingly severe threats. Understanding and monitoring the distribution and changes of vegetation and wetlands is crucial for protecting biodiversity and maintaining ecosystem health. In this context, land cover classification, especially the classification of vegetation and wetlands, has significant scientific and practical implications.

[0003] Land cover classification helps identify different types of land cover, monitor ecosystem changes, assess the effectiveness of environmental protection measures, and provide data support for land management and planning. With the development of remote sensing technology, multispectral remote sensing data has made large-scale, high-precision land cover classification possible. However, current land cover classification methods face many challenges. Due to the complexity and heterogeneity of land cover, phenomena such as "same spectral, different land cover" and "same land cover, different spectral characteristics" exist, increasing the difficulty of land cover classification. Although using deep learning methods to directly learn from multispectral data has improved classification accuracy to some extent, the accuracy is still not high enough. Summary of the Invention

[0004] This invention provides a method and apparatus for classifying land cover in remote sensing images based on index and coverage features, in order to address the shortcomings of insufficient accuracy in land cover classification in existing technologies.

[0005] A remote sensing image land cover classification method guided by index and cover features, comprising:

[0006] Acquire remote sensing spectral images;

[0007] The remote sensing spectral images are input into a trained land cover classification model to classify the land cover on the Earth's surface.

[0008] The land cover classification model includes:

[0009] The backbone network is used to extract global and local features from the remotely sensed spectral image.

[0010] The index feature extraction module is used to extract the index features corresponding to the spectral indices of the remote sensing spectral image; the spectral indices of the remote sensing spectral image include: Normalized Difference Vegetation Index (NDVI) and Dead Fuel Index (DFI).

[0011] The first splicing module is used to splice the exponential features with the global features to obtain a first spliced ​​feature map;

[0012] The first channel attention module is used to adjust the channel weights of the first spliced ​​feature map to obtain the first feature map after weight adjustment.

[0013] A multi-scale feature extraction module is used to extract multi-scale features from the weight-adjusted first feature map to obtain a multi-scale feature map.

[0014] The land cover feature extraction module is used to extract the land cover features corresponding to the land cover of the remote sensing spectral image; the land cover of the remote sensing spectral image includes: photosynthetic vegetation cover f. PV Non-photosynthetic vegetation cover f NPV With bare soil coverage f BS ;

[0015] The second stitching module is used to stitch the coverage feature with the multi-scale feature map to obtain a second stitched feature map;

[0016] The second channel attention module is used to adjust the channel weights of the second spliced ​​feature map to obtain the second feature map after weight adjustment.

[0017] The first convolution module is used to convolve the local features to obtain the first convolutional features;

[0018] The second convolution module is used to convolve the second feature map to obtain the second convolution feature;

[0019] The third concatenation module is used to concatenate the first convolutional feature with the second convolutional feature to obtain the third concatenated feature map;

[0020] The third convolutional module is used to classify the third stitched feature map to obtain the specific type of land cover.

[0021] Furthermore, in the remote sensing image land cover classification method based on index and coverage features as described above, the index of the remote sensing spectral image is extracted according to the following formula;

[0022]

[0023] Where R represents the red band, NIR represents the near-infrared band, and SWIR1 and SWIR2 represent two short-wave infrared bands.

[0024] Furthermore, in the remote sensing image land cover classification method based on index and coverage features as described above, the index feature extraction module includes a 7×7 convolutional layer, a max pooling layer, four different residual blocks, and a 1×1 convolutional layer connected in sequence.

[0025] The 7×7 convolutional layer is used to extract the basic features of the spectral indices of the remote sensing spectral image;

[0026] The max-pooling layer is used to reduce the spatial dimension of the basic features;

[0027] The four different sets of residual blocks are used to extract the depth features of the basic features after the spatial dimension reduction;

[0028] The 1×1 convolutional layer is used to enhance the feature representation of the depth features and adjust the number of channels so that the adjusted number of channels matches the number of channels in the first channel attention module;

[0029] The coverage feature extraction module includes: a 7×7 convolutional layer, a max pooling layer, four different sets of residual blocks, and a 1×1 convolutional layer connected in sequence; the four different sets of residual blocks include multiple convolutional layers and batch normalization layers; the multiple convolutional layers adopt dilated convolution.

[0030] Furthermore, in the remote sensing image land cover classification method based on index and coverage features as described above, the multi-scale feature extraction module uses a dense void spatial pyramid pooling module to extract the multi-scale feature map of the weighted first feature map.

[0031] Furthermore, in the remote sensing image land cover classification method guided by index and cover features as described above, the land cover of the remote sensing spectral image is determined according to the following method:

[0032] Feature pixel candidate areas and clean pixel index thresholds are determined based on the remote sensing spectral images, respectively; the feature pixel candidate areas include: photosynthetic vegetation feature pixel candidate areas, non-photosynthetic vegetation feature pixel candidate areas, and bare soil feature pixel candidate areas.

[0033] Pure pixels whose pure pixel index is greater than the pure pixel index threshold are extracted from the candidate feature pixel region as target feature pixels;

[0034] Determine the endmember feature value based on the target feature pixel;

[0035] The land cover of the remote sensing spectral image is determined based on the end-member feature values.

[0036] Furthermore, in the remote sensing image land cover classification method guided by index and cover features as described above, the step of determining the feature pixel candidate region based on the remote sensing spectral image includes:

[0037] The remote sensing spectral image is preprocessed to obtain a preprocessed remote sensing spectral image; and the NDVI and DFI corresponding to each pixel in the preprocessed remote sensing spectral image are calculated.

[0038] Random sampling is performed on the pixels containing the NDVI and DFI to obtain the sampling index feature;

[0039] The sampling index features are clustered using the DBSCAN method to obtain the largest cluster.

[0040] The convex hull corresponding to the largest cluster class is determined using the QuickHull method to construct the boundary of the feature triangle;

[0041] The vertex corresponding to the largest area triangle formed by the boundary of the feature triangle is determined.

[0042] Each region within the endmember feature threshold θ, centered on a vertex of the triangle, is selected as a candidate region for feature pixels. The candidate regions for feature pixels include: candidate regions for photosynthetic vegetation, candidate regions for non-photosynthetic vegetation, and candidate regions for bare soil. The candidate regions for feature pixels are obtained based on the assumptions of the pixel triad model.

[0043] Furthermore, in the remote sensing image land cover classification method guided by index and cover features as described above, the step of determining the clean pixel index threshold based on the remote sensing spectral image includes:

[0044] The preprocessed remote sensing spectral image is subjected to MNF transformation to obtain an MNF image;

[0045] The PPI value of each pixel in the MNF image is determined using the PPI algorithm;

[0046] The clean pixel index threshold is determined based on the PPI values ​​of all pixels in the MNF image.

[0047] Furthermore, in the remote sensing image land cover classification method guided by index and cover features as described above, the step of determining the endmember feature value based on the target feature pixel includes:

[0048] Determine the NDVI and DFI corresponding to each pixel in the target feature pixels;

[0049] Based on the NDVI and DFI corresponding to each pixel, the average NDVI and average DFI of the target feature pixel are determined respectively; the average NDVI and average DFI of the target feature pixel are used as the endmember feature value; the endmember feature value includes NDVI. PV NDVI NPV NDVI BS DFI PV DFI NPV DFI BS .

[0050] Furthermore, in the remote sensing image land cover classification method guided by index and cover features as described above, determining the cover feature based on the endmember feature value includes:

[0051] The endmember feature value is calculated according to the following formula:

[0052] NDVI = ∑f i V i =f PV NDVI PV +f NPV NDVI NPV +f BS NDVI BS #(3)

[0053] DFI=∑f i D i =f PV DFI PV +f NPV DFI NPV +f BS DFI BS #(4)

[0054] ∑f i =f PV +f NPV +f BS =1#(5)

[0055]

[0056] NDVI and DFI are index values ​​calculated based on remote sensing spectral images, f PV f NPV and f BS NDVI represents the coverage of light and vegetation, non-light and vegetation, and bare soil, respectively. PV and DFI PV These represent the NDVI index and DFI index of photosynthetic vegetation cover, respectively; NDVI NPV and DFI NPVThese represent the NDVI and DFI indices of non-photosynthetic vegetation, respectively; NDVI BS and DFI BS C represents the NDVI index and DFI index of bare soil, respectively. x C y and C z This indicates the coverage of a certain end-member type where outliers occur.

[0057] A remote sensing image land cover classification device guided by index and cover features, comprising:

[0058] The acquisition unit allows users to acquire remote sensing spectral images;

[0059] The classification unit is used to input the remote sensing spectral image into the trained land cover classification model to classify the land cover on the Earth's surface.

[0060] The land cover classification model includes:

[0061] The backbone network is used to extract global and local features from the remotely sensed spectral image.

[0062] The index feature extraction module is used to extract the index features corresponding to the spectral indices of the remote sensing spectral image; the spectral indices of the remote sensing spectral image include: NDVI and DFI;

[0063] The first splicing module is used to splice the exponential features with the global features to obtain a first spliced ​​feature map;

[0064] The first channel attention module is used to adjust the channel weights of the first spliced ​​feature map to obtain the first feature map after weight adjustment.

[0065] A multi-scale feature extraction module is used to extract multi-scale features from the weight-adjusted first feature map to obtain a multi-scale feature map.

[0066] The land cover feature extraction module is used to extract the land cover features corresponding to the land cover of the remote sensing spectral image; the land cover of the remote sensing spectral image includes: photosynthetic vegetation cover f. PV Non-photosynthetic vegetation cover f NPV With bare soil coverage f BS ;

[0067] The second stitching module is used to stitch the coverage feature with the multi-scale feature map to obtain a second stitched feature map;

[0068] The second channel attention module is used to adjust the channel weights of the second spliced ​​feature map to obtain the second feature map after weight adjustment.

[0069] The first convolution module is used to convolve the local features to obtain the first convolutional features;

[0070] The second convolution module is used to convolve the second feature map to obtain the second convolution feature;

[0071] The third concatenation module is used to concatenate the first convolutional feature with the second convolutional feature to obtain the third concatenated feature map;

[0072] The third convolutional module is used to classify the third stitched feature map to obtain the specific type of land cover.

[0073] The remote sensing image land cover classification method provided by this invention, which introduces index and cover features into the model in stages, can classify land cover and significantly improve the accuracy of land cover classification. Attached Figure Description

[0074] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0075] Figure 1 This is a flowchart of the remote sensing image land cover classification method based on index and coverage features provided by the present invention;

[0076] Figure 2 The structural diagram of the land cover classification model provided by this invention;

[0077] Figure 3 The structural diagram of the exponential feature extraction module provided by this invention;

[0078] Figure 4 A schematic diagram of the channel attention module structure provided by the present invention;

[0079] Figure 5 A schematic diagram of the dense void space pyramid pooling module structure provided by the present invention;

[0080] Figure 6 This is a flowchart of the method for extracting land cover according to the present invention;

[0081] Figure 7 This is a feature space distribution diagram of the NDVI-DFI model under ideal conditions;

[0082] Figure 8 This is a comparison chart of the classification effects of the method of the present invention with other methods. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0084] Remote sensing technology is a technique for acquiring and recording information about the Earth's surface and atmosphere using long-range sensors. It aims to provide fundamental data on the Earth's surface, vegetation, water bodies, soil, and atmosphere for research across various disciplines. Remote sensing images can be categorized by data type into panchromatic images, spectral images, and radar images. Panchromatic images, acquired by panchromatic sensors, contain only grayscale information in a single band and typically possess higher spatial resolution, capturing more image details. Spectral images, acquired by multispectral sensors, contain information from multiple bands, providing rich spectral features and reflecting the characteristic differences of different land features. Radar images, acquired by radar sensors, are primarily used for microwave band data acquisition, possessing the advantage of penetrating clouds and fog, and are unaffected by atmospheric interference, making them suitable for topographic surveying, snow and ice monitoring, and military reconnaissance. The purpose of this invention is the identification of land cover; therefore, high-resolution multispectral remote sensing images are employed to more accurately extract vegetation feature details.

[0085] Figure 1 This is a flowchart of the remote sensing image land cover classification method based on index and cover features provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps:

[0086] Step 1: Acquire remote sensing spectral imagery;

[0087] Step 2: Input the remote sensing spectral image into the trained land cover classification model to classify the land cover on the Earth's surface.

[0088] Specifically, because different land cover materials have different spectral characteristics, even the same type of land cover may have different spectral characteristics. For example, vegetation includes woodlands and grasslands, as well as healthy vegetation and vegetation with pests and diseases; the leaf color of vegetation can be green or withered yellow. Due to differences in the water content and chlorophyll content of plants, their ability to reflect and absorb electromagnetic waves also varies, resulting in different land cover materials having different spectral characteristics. This invention aims to identify land cover materials using the spectral, spatial, and temporal characteristics of remote sensing imagery. The classification results can provide data support for wetland distribution mapping, ecological environment assessment, climate change analysis, and soil and water conservation.

[0089] like Figure 2As shown, the land cover classification model provided by this invention includes: a backbone network for extracting global and local features of the remote sensing spectral image; an index feature extraction module for extracting index features corresponding to the spectral indices of the remote sensing spectral image; the spectral indices of the remote sensing spectral image include NDVI and DFI; a first stitching module for stitching the index features with the global features to obtain a first stitched feature map; a first channel attention module for adjusting the channel weights of the first stitched feature map to obtain a weight-adjusted first feature map; a multi-scale feature extraction module for extracting multi-scale features of the weight-adjusted first feature map to obtain a multi-scale feature map; and a land cover feature extraction module for extracting land cover features corresponding to the land cover of the remote sensing spectral image; the land cover of the remote sensing spectral image includes photosynthetic vegetation cover f. PV Non-photosynthetic vegetation cover f NPV With bare soil coverage f BS The system comprises: a second stitching module for stitching the land cover feature with the multi-scale feature map to obtain a second stitched feature map; a second channel attention module for adjusting the channel weights of the second stitched feature map to obtain a weight-adjusted second feature map; a first convolution module for convolutionally processing the local features to obtain a first convolutional feature; a second convolution module for convolutionally processing the second feature map to obtain a second convolutional feature; a third stitching module for stitching the first convolutional feature with the second convolutional feature to obtain a third stitched feature map; and a third convolution module for classifying the third stitched feature map to obtain the specific type of land cover.

[0090] Specifically, the local features extracted by the backbone network are the spectral features extracted by the first few layers of the backbone network. These features focus more on local image details (edges, textures, shapes) and are primarily used to capture low-level visual information such as edges and textures. The global features extracted by the backbone network are the depth features obtained after the entire backbone network has passed through it. These features focus more on the contextual information of the image and the relationships between objects, such as more vegetation in wetlands, the presence of buildings around farmland, or the presence of rivers near cities. The first and second convolutional modules are mainly for feature integration and dimensionality compression, while the third convolutional module is for local feature extraction and spatial information processing.

[0091] The land cover classification model provided by this invention adopts an encoder-decoder architecture. First, spectral data from remote sensing imagery is input into the backbone network to extract local and global features. The local features are then processed by a 1×1 convolution and retained for use in the decoder. The global features are concatenated with the exponential features extracted by the Index Feature Extraction Module (IFEM). This concatenated feature map is then processed by the first channel attention module (CA) to obtain a channel-weighted feature map. Next, the DenseASPP module is applied to the obtained feature map to extract multi-scale features. These multi-scale features are then concatenated with the coverage features extracted by the Cover Feature Extraction Module (CFEM). The concatenated feature map is processed by a second attention module to obtain a channel-weighted feature map. Finally, the channel-weighted feature map is concatenated with the 1×1 convolutioned local features and then restored to the same size as the input image through a 3×3 convolution and bilinear interpolation, outputting the land cover classification result.

[0092] It should be noted that: when splicing feature information in this invention, since the dimensionality of local features may be high, in order to prevent the global features from being weakened, it is necessary to first perform 1×1 convolution dimensionality reduction on the local features, and then splice them with the feature map after 1×1 convolution and bilinear interpolation.

[0093] Different spectral bands possess different feature representation capabilities. Therefore, recombinating spectral information in a certain form can construct a functional relationship with vegetation characteristic parameters; this combination of relationships is called a vegetation index. Vegetation indices are closely related to vegetation growth status, not only eliminating the influence of other interfering factors to a certain extent but also fully reflecting the characteristic information of vegetation. Therefore, they are often used in vegetation type identification, crop growth trend forecasting, land cover classification, and vegetation cover evaluation. Index features can characterize the vegetation growth status; a higher NDVI index indicates better vegetation growth at the corresponding location, while a higher DFI index indicates that the vegetation at the corresponding location is in a dry state with less water supply. Therefore, the method provided in this invention extracts index features through an index feature extraction module, making these index features contain a certain degree of semantic information. This helps the model to understand the state and growth environment of vegetation in the image more deeply, contributing to improving the accuracy and robustness of vegetation type identification in remote sensing images.

[0094] Vegetation cover refers to the percentage of the vertically projected area of ​​vegetation within an observation area, representing the state of plant community coverage on the ground. As a comprehensive ecological indicator, ground cover can quantitatively assess vegetation cover status. Cover characteristics rely on spectral data and index features, encompassing not only the physical or ecological attributes reflected by the index but also more information about the types, density, and distribution of ground cover, thus possessing higher-level semantic information.

[0095] The method provided in this invention addresses the issue that the complexity and noise of remote sensing imagery often lead to unsatisfactory semantic segmentation accuracy. To improve the classification accuracy and robustness of land cover, this invention utilizes an index feature extraction module (IFEM) and a cover feature extraction module (CFEM) to add the extracted index and cover features to different parts of the model in stages. This approach fully leverages the semantic information of the index and cover features to guide the model, thereby more effectively extracting and integrating vegetation features and improving the classification accuracy of land cover.

[0096] Furthermore, the index of the remote sensing spectral image is extracted according to the following formula;

[0097]

[0098] Wherein, R represents the red band, NIR represents the near-infrared band, and SWIR1 and SWIR2 represent two short-wave infrared bands.

[0099] Furthermore, such as Figure 3 As shown, the exponential feature extraction module includes a 7×7 convolutional layer, a max pooling layer, four different residual blocks, and a 1×1 convolutional layer connected in sequence. The 7×7 convolutional layer is used to extract the basic features of the spectral index of the remote sensing spectral image. The max pooling layer is used to reduce the spatial dimension of the basic features. The four different residual blocks are used to extract the depth features of the basic features after the spatial dimension reduction. The 1×1 convolutional layer is used to enhance the feature representation of the depth features and adjust the number of channels so that the adjusted number of channels matches the number of channels in the first channel attention module.

[0100] Specifically, the basic features of spectral indices from remotely sensed spectral images are first extracted using 7×7 convolutions, and max pooling is used to reduce the spatial dimensionality of the feature maps. Then, four different sets of residual blocks are applied to extract depth features, with 3, 4, 6, and 3 blocks respectively. Each residual block contains multiple convolutional layers and batch normalization, and features are reused through skip connections to enhance network stability. Finally, 1×1 convolutions are used to enhance the feature representation and adjust the number of channels to accommodate the subsequent first-channel attention module.

[0101] Furthermore, the coverage feature extraction module includes: a 7×7 convolutional layer, a max pooling layer, four different sets of residual blocks, and a 1×1 convolutional layer connected in sequence; the four different sets of residual blocks include multiple convolutional layers and batch normalization layers; the multiple convolutional layers employ dilated convolution.

[0102] Specifically, the main difference between the coverage feature extraction module CFEM and the exponential feature extraction module IFEM lies in the design structure of the ResBlock. Since IFEM primarily focuses on the details of exponential features, larger-scale correlation information can be learned by guiding spectral features in the DenseASPP module. Therefore, the ResBlock in the exponential feature extraction module uses classic convolutional layers. In contrast, the ResBlock in CFEM, as a higher-level coverage feature, possesses extensive spatial semantic information. Therefore, the ResBlock in the coverage feature extraction module uses dilated convolutions to expand the receptive field before concatenating it with multi-scale features to maximize the preservation of different feature types.

[0103] Figure 4 This is a schematic diagram of the channel attention module structure provided by the present invention, as shown below. Figure 4 As shown, after introducing exponential and coverage features in stages, the CA module needs to be applied to constrain the model to focus on important features and ignore redundant information. Therefore, a channel attention module (CA) is added after the exponential and coverage features are integrated into the network. The CA module uses a channel attention mechanism to adjust the weights of different channels in the feature map. First, a global average pooling operation is performed to compress global spatial information into the channel descriptors. Then, two fully connected layers and the ReLU activation function are used to process the channel descriptors. Finally, the Softmax activation function is applied to convert the feature scores of the fully connected layers into a probability distribution, which is the weight matrix representing the importance of each channel. Multiplying the feature map and the weight matrix element by element yields a weighted feature map with importance preferences.

[0104] Furthermore, the multi-scale feature extraction module employs a dense, hollow spatial pyramid pooling module to extract the multi-scale feature map of the weight-adjusted first feature map.

[0105] Specifically, Figure 5This is a schematic diagram of the dense void space pyramid pooling module structure provided by the present invention, as shown below. Figure 5 As shown, features at different scales in remote sensing images can reflect important information containing more vegetation details. However, obtaining a large receptive field through ordinary convolution significantly increases the computational load, leading to excessively high network training time costs. Dilated convolution is a feasible solution, effectively expanding the receptive field by losing neighboring pixel feature information while reducing computational load. However, when the dilation rate is high, excessive loss of neighboring pixel information can cause dilated convolution to perform poorly or even fail. To address this issue, this invention introduces DenseASPP to extract multi-scale feature information. DenseASPP combines the advantages of ASPP and DenseNet, enabling efficient extraction of multi-scale feature information. ASPP is a core module in the DeepLabv3+ model, using dilated convolutions with different dilation rates to process the feature maps output by the backbone network to obtain multi-scale features. Inspired by the dense connection design in the DenseNet network, DenseASPP uses only a reasonable dilation rate for convolution, densely connecting the convolutional feature information to retain more effective information. This results in extracted features that not only possess multi-scale feature information but also retain a wide range of semantic information.

[0106] Furthermore, Figure 6 This is a flowchart of the method for extracting land cover according to the present invention, as shown below. Figure 6 As shown, the present invention determines the land cover of the remote sensing spectral image according to the following method:

[0107] Step 1: Determine the candidate regions for feature pixels and the pure pixel index threshold based on the remote sensing spectral image. The candidate regions for feature pixels contain land cover that needs to be classified, specifically: candidate regions for photosynthetic vegetation feature pixels, candidate regions for non-photosynthetic vegetation feature pixels, and candidate regions for bare soil feature pixels.

[0108] Step 2: Extract clean pixels with a clean pixel index greater than the clean pixel index threshold from the candidate feature pixel region as target feature pixels.

[0109] Specifically, let's take a net pixel index threshold of 5 as an example. Assume that pixels with a net pixel index of 5 are pure pixels containing only a single endmember type; these pixels can numerically reflect endmember features. Pixels with a net pixel index less than 5 are mixed pixels and cannot reflect endmember features, therefore no feature extraction is performed on them. Pure pixels with a net pixel index greater than 5 are extracted using the feature pixel candidate region. The arithmetic mean of the NDVI and DFI indices of these pixels is used as the endmember feature value. Substituting this into the pixel triad model allows us to calculate the f-value for all pixels. PV f NPV with fBS .

[0110] Step 3: Determine the endmember feature values ​​based on the target feature pixels.

[0111] Specifically, firstly, the NDVI and DFI corresponding to each pixel in the target feature pixels are determined; then, based on the NDVI and DFI corresponding to each pixel, the average NDVI and average DFI of the target feature pixels are determined; the average NDVI and average DFI of the target feature pixels are used as the endmember feature values; the endmember feature values ​​include NDVI. PV NDVI NPV NDVI BS DFI PV DFIN PV DFI BS Among them, NDVI PV and DFI PV These represent the NDVI index and DFI index of photosynthetic vegetation cover, respectively; NDVI NPV and DFI NPV These represent the NDVI and DFI indices of non-photosynthetic vegetation, respectively; NDVI BS and DFI BS These represent the NDVI index and DFI index of bare soil, respectively.

[0112] Step 4: Determine the land cover of the remote sensing spectral image based on the end-member feature values.

[0113] Specifically, the endmember feature value is calculated according to the following formula:

[0114] NDVI = ∑f i V i =f PV NDVI PV +f NPV NDVI NPV +f BS NDVI BS #(3)

[0115] DFI=∑f i D i =f PV DFI PV +f NPV DFI NPV +f BS DFI BS #(4)

[0116] ∑f i =f PV +f NPV +f BS=1#(5)

[0117]

[0118] NDVI and DFI are index values ​​calculated based on remote sensing spectral images, f PV f NPV and f BS NDVI represents the coverage of light and vegetation, non-light and vegetation, and bare soil, respectively. PV and DFI PV These represent the NDVI index and DFI index of photosynthetic vegetation cover, respectively; NDVI NPV and DFI NPV These represent the NDVI and DFI indices of non-photosynthetic vegetation, respectively; NDVI BS and DFI BS C represents the NDVI index and DFI index of bare soil, respectively. x C y and C z This indicates the coverage of a certain endmember type where outliers occur. Formula (6) represents the range of outliers in the coverage, C. x C y and C z In essence, it refers to the abnormal range of the three coverage levels, which represents the coverage level at which a certain abnormal situation occurs. For example, if f PV If it's an outlier, then C x Then it is f PV C y and C z Then it means f NPV and f BS If f NPV If it's an outlier, then C x It refers to f NPV Because C x If the other two coverage values ​​are within the normal range, then the values ​​of the other two coverage values ​​need to be recalculated according to the formula to satisfy formula (5). If the other two coverage values ​​are also outliers, then they are set to 0, indicating that the pixel is abnormal.

[0119] The following section describes how to determine candidate regions for feature pixels based on remote sensing spectral images:

[0120] like Figure 7As shown, the NDVI-DFI pixel triad model assumes that a pixel consists of three components: PV, NPV, and BS. Ideally, the NDVI-DFI feature space map is represented as a triangle, where PV has the largest NDVI value and the smallest DFI value, located in the middle right side of the triangle. The green circle indicates the distribution location of pure PV pixels. NPV has the smallest NDVI value and the largest DFI value, located in the upper left corner of the triangle. The red circle indicates the distribution location of pure NPV pixels. BS has the smallest NDVI and DFI values, located in the lower left corner of the triangle. The blue circle indicates the distribution location of pure BS pixels.

[0121] The NDVI and DFI indices conform to the linear relationship of formulas (3)-(5), where NDVI and DFI are the index values ​​calculated from a given Landsat remote sensing image, f PV f NPV with f BS These represent the coverage of light and vegetation, non-light and vegetation, and bare soil, respectively. NDVI PV NDVI NPV NDVI BS and DFI PV DFI NPV DFI BS C represents two eigenvalues ​​representing three types of endmembers. x C represents the coverage of a certain endmember type outside the triangle after the mixed pixel decomposition. y and C z This indicates two other endmember coverage methods. The NDVI-DFI pixel triad model limits f. PV f NPV with f BS If the sum of the values ​​is equal to 1, and the estimated coverage is not within the range of 0 to 1, then it is processed according to formula (6), which means that the corresponding pixel is covered and no unmixing is performed.

[0122] Therefore, determining the PV, NPV, and BS endmember eigenvalues ​​is a crucial step in the successful decomposition of the NDVI-DFI pixel triad model. One relatively mature method currently is to use the PPI-2DSP algorithm to determine clean pixels and use the arithmetic mean of the clean pixels as the endmember eigenvalues. PPI is a widely used, semi-automatic, interactive method for finding clean pixels in an image. It iteratively maps random unit vectors onto an N-dimensional scatter plot, projecting each pixel onto the unit vector, recording pixels within a certain threshold range of extreme values, and using the total number of times these pixels appear within the extreme value range as the purity. Higher pixel purity indicates a cleaner pixel; pixels with purity exceeding a certain threshold can be considered clean pixels containing only a single endmember. According to the assumptions of the pixel triad model, clean pixels near the vertices of the feature space triangle can reflect the characteristics of the corresponding endmembers. Therefore, the arithmetic mean of the NDVI and DFI indices in these clean pixels can be considered as the eigenvalues ​​of the NDVI and DFI indices of the three endmembers, respectively.

[0123] 2DSP is a feature space map constructed based on NDVI and DFI indices. Combined with the clean pixel index extracted by the PPI algorithm, clean pixels located at the vertices of feature triangles can be manually selected. The arithmetic mean of these indices is then used as the feature values ​​for the three types of endmembers. Figure 7 As shown, the blue circle represents the BS feature pixel candidate region, the green circle represents the PV feature pixel candidate region, and the red circle represents the NPV feature pixel candidate region. However, since the 2DSP algorithm relies on manual selection of feature pixel candidate regions, the calculation of endmember feature values ​​has significant uncertainty and low efficiency, making it difficult to extend to practical applications. Therefore, this invention provides an efficient adaptive endmember feature value extraction method by introducing the DBSCAN clustering algorithm and the QuickHull convex hull detection algorithm to reduce the dependence of the endmember extraction process on manual selection of feature pixel candidate regions. The method includes the following steps:

[0124] Step 1: Preprocess the remote sensing spectral image to obtain a preprocessed remote sensing spectral image; and calculate the NDVI and DFI corresponding to each pixel in the preprocessed remote sensing spectral image.

[0125] Specifically, the preprocessing measures for the remote sensing spectral image of the present invention are as follows: First, data with the smallest possible cloud cover ratio is obtained using quality assessment bands. Then, interference factors such as clouds, shadows, water bodies, and snow are removed. Next, the coordinate system of the remote sensing image data is converted to WGS-84 using the arcpy and gdal toolkits, and the multispectral bands are combined. Finally, the background value of the image is set, and the preprocessed remote sensing spectral image is obtained. Then, based on the band data of the preprocessed remote sensing spectral image, the NDVI and DFI corresponding to each pixel are calculated using the band calculation tools of ArcGIS software according to calculation formulas (1) and (2).

[0126] Step 2: Randomly sample the pixels containing the NDVI and DFI to obtain the sampling index feature.

[0127] Specifically, since a remote sensing image contains a large amount of pixel data, directly putting the remote sensing image into the model would result in a large amount of computation. Therefore, this invention uses a sampling rate η to randomly select pixels containing the NDVI and DFI as a subset to determine the land cover.

[0128] Step 3: Cluster the sampling index features using the DBSCAN method to obtain the largest cluster.

[0129] Specifically, since the NDVI-DFI feature space is represented by irregular triangles and contains a large number of outliers, and the DBSCAN algorithm can adapt to irregularly shaped clusters, it has good performance in handling noise and outliers. Furthermore, considering that this algorithm is not affected by the choice of initial points and has low sensitivity to parameters, exhibiting relatively stable characteristics, the DBSCAN algorithm is chosen to remove discrete points and construct the largest cluster containing the feature triangles. In the DBSCAN algorithm, for each point in a given dataset, its ε-neighborhood is defined as a circle centered at that point with radius ε. If a point's ε-neighborhood contains at least min... pts If there are ε-neighborhoods of points p and q in the dataset, then that point is called a core point. For two points p and q in the dataset, if q lies within the ε-neighborhood of p, and p is a core point, then q is said to be directly density-reachable from p. If there exists a sequence of points p1, p2, ..., pq in the dataset... n Where p1 = p, p n =q, and p i+1 By p iIf q is directly density-reachable, then it is said that q is density-reachable by p. The DBSCAN algorithm can then be described in the following four steps: 1. Initialization: Mark all data points as unvisited; 2. Traverse the dataset: For each unvisited data point p, if p is a core point, then starting from p, mark the cluster to which p belongs based on the density reachability relation; 3. Iteration: Repeat step 2 until all data points have been visited; 4. Output: Take the resulting cluster as the clustering result and mark the remaining points as noise.

[0130] Step 4: Determine the convex hull corresponding to the largest cluster class using the QuickHull method to construct the boundary of the feature triangle.

[0131] Specifically, the QuickHull algorithm is an efficient algorithm for calculating convex hulls. Its core idea is to construct the convex hull through recursive divide-and-conquer and elimination methods. Due to the massive scale of remote sensing image data, and the QuickHull algorithm's fast processing speed, it can efficiently handle large-scale point sets. Therefore, this invention chooses this algorithm to detect convex hulls, preparing for subsequent determination of feature triangles. The QuickHull algorithm principle is relatively simple and can be described in the following four steps: 1. Initialization: Select two points A and B in the data point set with the smallest and largest x-coordinates, respectively, as the two endpoints of the convex hull; 2. Partition: Divide the data point set into two parts according to the line AB, each located on one side of the line; 3. Recursion: Recursively apply the QuickHull algorithm to each part to construct the convex hull; 4. Merge the convex hulls obtained from the two parts into the overall convex hull.

[0132] Step 5: Determine the vertex corresponding to the largest area triangle formed by the boundary.

[0133] Step 6: Taking the vertex of the triangle as the center, select the region within the endmember feature threshold θ as the feature pixel candidate region; the feature pixel candidate region includes: photosynthetic vegetation feature pixel candidate region, non-photosynthetic vegetation feature pixel candidate region and bare soil feature pixel candidate region; the feature pixel candidate region is obtained according to the assumption of the pixel triad model.

[0134] The method provided by this invention mainly includes a random sampling rate η, a neighborhood radius ε, and a minimum number of neighbors min. pts The four parameters are η and the endpoint threshold θ. η is used in the random sampling module to reduce the data size of NDVI and DFI; ε and min pts In the DBSCAN clustering module, the former is used to define the neighborhood range of a data point, and the latter is used to represent the minimum number of neighbors required for a core point; θ in the feature triangle vertex extraction module is used to determine the size of the feature cell candidate region.

[0135] The following section describes how to obtain the clean pixel index threshold for remotely sensed spectral images:

[0136] Step 1: Perform MNF transformation on the preprocessed remote sensing spectral image to obtain an MNF image;

[0137] Step 2: Determine the PPI value of each pixel in the MNF image using the PPI algorithm;

[0138] Step 3: Determine the clean pixel index threshold based on the PPI values ​​of all pixels in the MNF image.

[0139] Specifically, the higher the pixel purity of a remote sensing spectral image, the purer the pixel. Pixels with purity exceeding a certain threshold can be considered pure pixels containing only a single endmember. Based on the assumption of the pixel triad model, pure pixels located near the vertices of the feature space triangle can reflect the characteristics of the corresponding endmember. Therefore, this invention selects pure pixels corresponding to photosynthetic vegetation, non-photosynthetic vegetation, and bare soil from the feature pixel candidate area by determining a pure pixel index threshold.

[0140] The training method for the land cover classification model provided by this invention is described below:

[0141] This invention uses Python as the programming language and relies on the PyTorch development framework for development. All experiments are run on a Linux server, and the specific configuration is shown in Table 1. Experimental data are divided into training, validation, and test sets in a 7:2:1 ratio. The model is trained on the training set, the hyperparameters of the model are adjusted on the validation set, and the segmentation performance of the model is evaluated on the test set.

[0142] Table 1 Experimental Environment

[0143]

[0144] To balance computational overhead and efficiency, the backbone network for the feature extraction module uses the widely adopted ResNet-50. The loss function employed is the classic multi-class cross-entropy loss function, a core feature in image segmentation tasks, as shown in Equation 7. During training, the multi-class cross-entropy loss function tends to converge quickly, its gradient calculation is relatively simple, and it directly outputs the probability distribution of each class, making it easy to understand and interpret.

[0145]

[0146] Where x and y represent the sample and label, respectively, h and w represent the height and width of the sample, respectively, c represents the class, and F represents the output of the remote sensing image land cover classification method of this invention. To address the potential imbalance problem in the dataset, in actual training, using only the multi-class cross-entropy loss function can lead to the model favoring samples with higher frequency occurrences, ultimately resulting in poor segmentation performance. Therefore, the Dice loss is introduced. This loss, widely used in medical image segmentation, aims to measure the similarity between the model's output segmentation mask and the true segmentation mask, making the model more robust to imbalanced datasets. By minimizing the Dice loss, the model is encouraged to generate more accurate segmentation results, as shown in Equation 8:

[0147]

[0148] Where Intersection represents the intersection of the model predictions and the labels, Union represents the union of the model predictions and the labels, and ∈ is a very small smoothing coefficient used to prevent the denominator from being zero and to smooth the loss function; in the experiment, it was set to a value of 10. -10 The cross-entropy loss is combined with the Dice loss to comprehensively consider the spatial location and category information of the remote sensing image, thereby achieving better training results and segmentation performance. The overall loss is shown in Equation 9.

[0149] Loss=crossEntropyLoss+DiceLoss#(9)

[0150] The learning rate optimizer uses the Adam adaptive learning rate algorithm, with an initial learning rate set to 10. -4 The learning rate scheduler uses StepLR, the decay factor (γ) is set to 0.9, the batch size is set to 32, and each experiment is trained for 200 rounds. As shown in Equations 4 to 6, the evaluation metrics for each category are accuracy (Accuracy, Acc), intersection over union (IoU), and F1 score. As shown in Equations (13) to (15), the overall evaluation metric is the arithmetic mean of the above three metrics, namely AA (Average Accuracy, AA), mIoU (Mean Intersection over Union, mIoU), and mF1 (Mean F1 Score, mF1). Overall accuracy (OA) is not used because this metric measures the ratio of the number of pixels correctly classified by the model to the total number of pixels in the entire image, which is not sensitive enough in the case of imbalanced class distribution.

[0151]

[0152] Where N represents the number of categories, which is 9 in total in the FVC-subSEN12MS dataset since the ice and snow category has been removed; TP represents the number of True Positives (TP); FP represents the number of False Positives (FP); and FN represents the number of False Negatives (FN).

[0153] The effectiveness of the remote sensing image land cover classification method of the present invention is verified by experiments below:

[0154] The experiment consisted of an ablation experiment and a comparative experiment. Unless otherwise specified, the experimental parameters were consistent with those described above. The ablation experiment explored the improvement effects of different guiding features and the CA and DenseASPP modules on vegetation type identification tasks. The comparative experiment compared the remote sensing image land cover classification method of this invention with commonly used segmentation models to evaluate the classification performance of the method of this invention.

[0155] Based on the above settings, ablation experiments were conducted on the FVC-subSEN12MS dataset. The results of each evaluation index are shown in Table 2, with the best results indicated in bold. It can be observed that compared to the first set of experiments without using exponential and cover features, and without employing the DenseASPP and CA modules, using the DenseASPP module improved the overall AA, mIoU, and mF1 metrics by 0.4%, 0.4%, and 0.3%, respectively. Using the CA module improved the overall AA, mIoU, and mF1 metrics by 1.1%, 0.9%, and 0.6%, respectively. Furthermore, using exponential and cover features to guide network learning also achieved the best results in multiple categories, improving the overall AA, mIoU, and mF1 metrics by 2.3%, 2.0%, and 1.5%, respectively, compared to the first set of experiments. From the perspective of identifying specific vegetation categories, using exponential and cover features and each module achieved optimal results in almost all indicators, fully demonstrating the effectiveness of each module in this method.

[0156] Table 2 Ablation Experiment Results for Each Module (%)

[0157]

[0158] The results of the comparative experiments with other commonly used segmentation methods are shown in Table 3, with bold text indicating the best results. The accuracy of the method of this invention is optimal in the four categories of shrubs, savanna, grassland, and wetlands. The accuracy in forest land is 2.4% lower than the best-performing RefineNet, the accuracy in urban areas is 4.0% lower than the best-performing UNet, the accuracy in bare land is 0.8% lower than the best-performing FCN, and the accuracy in water bodies is 1.0% lower than the best-performing DeepLabv3+. The IoU and F1 scores of the method of this invention are optimal in the seven categories of forest land, shrubs, savanna, grassland, wetlands, cultivated land, and urban areas. The IoU and F1 scores in the bare land category are 0.8% and 0.4% lower than the best-performing DeepLabv3+, respectively, and the IoU and F1 scores in the water body category are 0.2% and 0.1% lower than the best-performing DeepLabv3+, respectively. Overall, the method of this invention achieves the best results in terms of AA, mIoU, and mF1, improving upon the most classic semantic segmentation model FCN by 4.6%, 6.9%, and 5.7%, respectively. The segmentation performance of each method is shown in Table 3. Figure 8 As shown, (a) represents RGB composite, (b) represents the real label, (c) represents FCN, (d) represents UNet, (e) represents FarSeg, (f) represents PSPNet, (g) represents DeepLabv3+, (h) represents HRNet, (i) represents RefineNet, (j) represents SegNet, and (k) represents the method of this invention. Compared with other methods, the classification method provided by this invention can better distinguish wetlands from water bodies, only causing some confusion in the edge and transition areas. It is more accurate in predicting cultivated land, especially in distinguishing it from grassland. In addition, although the transition area between woodland and grassland is identified as savanna, the misclassified area is smaller and the reliability is higher compared with other methods.

[0159] Table 3. Comparison of experimental results with commonly used segmentation methods (%)

[0160]

[0161] The present invention also provides a remote sensing image land cover classification device based on index and cover features, comprising:

[0162] The acquisition unit allows users to acquire remote sensing spectral images;

[0163] The classification unit is used to input the remote sensing spectral image into the trained land cover classification model to classify the land cover on the Earth's surface.

[0164] The land cover classification model includes:

[0165] The backbone network is used to extract global and local features from the remotely sensed spectral image.

[0166] The index feature extraction module is used to extract the index features corresponding to the spectral indices of the remote sensing spectral image; the spectral indices of the remote sensing spectral image include: NDVI and DFI;

[0167] The first splicing module is used to splice the exponential features with the global features to obtain a first spliced ​​feature map;

[0168] The first channel attention module is used to adjust the channel weights of the first spliced ​​feature map to obtain the first feature map after weight adjustment.

[0169] A multi-scale feature extraction module is used to extract multi-scale features from the weight-adjusted first feature map to obtain a multi-scale feature map.

[0170] The land cover feature extraction module is used to extract the land cover features corresponding to the land cover of the remote sensing spectral image; the land cover of the remote sensing spectral image includes: photosynthetic vegetation cover f. PV Non-photosynthetic vegetation cover f NPV With bare soil coverage f BS ;

[0171] The second stitching module is used to stitch the coverage feature with the multi-scale feature map to obtain a second stitched feature map;

[0172] The second channel attention module is used to adjust the channel weights of the second spliced ​​feature map to obtain the second feature map after weight adjustment.

[0173] The first convolution module is used to convolve the local features to obtain the first convolutional features;

[0174] The second convolution module is used to convolve the second feature map to obtain the second convolution feature;

[0175] The third concatenation module is used to concatenate the first convolutional feature with the second convolutional feature to obtain the third concatenated feature map;

[0176] The third convolutional module is used to classify the third stitched feature map to obtain the specific type of land cover.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for classifying land cover in remote sensing images based on index and coverage features, characterized in that, include: Acquire remote sensing spectral images; The remote sensing spectral images are input into a trained land cover classification model to classify the land cover on the Earth's surface. The land cover classification model includes: The backbone network is used to extract global and local features from the remotely sensed spectral image. The index feature extraction module is used to extract the index features corresponding to the spectral indices of the remote sensing spectral image; the spectral indices of the remote sensing spectral image include: NDVI and DFI; The first splicing module is used to splice the exponential features with the global features to obtain a first spliced ​​feature map; The first channel attention module is used to adjust the channel weights of the first spliced ​​feature map to obtain the first feature map after weight adjustment. A multi-scale feature extraction module is used to extract multi-scale features from the weight-adjusted first feature map to obtain a multi-scale feature map. The land cover feature extraction module is used to extract the land cover features corresponding to the land cover of the remote sensing spectral image; the land cover of the remote sensing spectral image includes: photosynthetic vegetation cover f. PV Non-photosynthetic vegetation cover f NPV With bare soil coverage f BS ; The second stitching module is used to stitch the coverage feature with the multi-scale feature map to obtain a second stitched feature map; The second channel attention module is used to adjust the channel weights of the second spliced ​​feature map to obtain the second feature map after weight adjustment. The first convolution module is used to convolve the local features to obtain the first convolutional features; The second convolution module is used to convolve the second feature map to obtain the second convolution feature; The third concatenation module is used to concatenate the first convolutional feature with the second convolutional feature to obtain the third concatenated feature map; The third convolution module is used to classify the third stitched feature map to obtain the specific type of land cover.

2. The remote sensing image land cover classification method based on index and coverage features as described in claim 1, characterized in that, The index of the remote sensing spectral image is extracted according to the following formula: Where R represents the red band, NIR represents the near-infrared band, and SWIR1 and SWIR2 represent two short-wave infrared bands.

3. The remote sensing image land cover classification method based on index and coverage features as described in claim 1, characterized in that, The exponential feature extraction module includes a 7×7 convolutional layer, a max pooling layer, four different residual blocks, and a 1×1 convolutional layer connected in sequence. The 7×7 convolutional layer is used to extract the basic features of the spectral indices of the remote sensing spectral image; The max-pooling layer is used to reduce the spatial dimension of the basic features; The four different sets of residual blocks are used to extract the depth features of the basic features after the spatial dimension reduction; The 1×1 convolutional layer is used to enhance the feature representation of the depth features and adjust the number of channels so that the adjusted number of channels matches the number of channels in the first channel attention module; The coverage feature extraction module includes: a 7×7 convolutional layer, a max pooling layer, four different residual blocks, and a 1×1 convolutional layer connected in sequence; The four different sets of residual blocks include multiple convolutional layers and batch normalization layers; the multiple convolutional layers employ dilated convolution.

4. The remote sensing image land cover classification method based on index and coverage features as described in claim 1, characterized in that, The multi-scale feature extraction module uses a dense, hollow spatial pyramid pooling module to extract the multi-scale feature map of the first feature map after weight adjustment.

5. The remote sensing image land cover classification method based on index and coverage features as described in claim 2, characterized in that, The land cover of the remote sensing spectral image is determined according to the following method: Feature pixel candidate areas and clean pixel index thresholds are determined based on the remote sensing spectral images, respectively; the feature pixel candidate areas include: photosynthetic vegetation feature pixel candidate areas, non-photosynthetic vegetation feature pixel candidate areas, and bare soil feature pixel candidate areas. Pure pixels whose pure pixel index is greater than the pure pixel index threshold are extracted from the candidate feature pixel region as target feature pixels; Determine the endmember feature value based on the target feature pixel; The land cover of the remote sensing spectral image is determined based on the end-member feature values.

6. The remote sensing image land cover classification method based on index and coverage features as described in claim 5, characterized in that, The step of determining the candidate region of feature pixels based on the remote sensing spectral image includes: The remote sensing spectral image is preprocessed to obtain a preprocessed remote sensing spectral image; and the NDVI and DFI corresponding to each pixel in the preprocessed remote sensing spectral image are calculated. Random sampling is performed on the pixels containing the NDVI and DFI to obtain the sampling index feature; The sampling index features are clustered using the DBSCAN method to obtain the largest cluster. The convex hull corresponding to the largest cluster class is determined using the QuickHull method to construct the boundary of the feature triangle; The vertex corresponding to the largest area triangle formed by the boundary of the feature triangle is determined. Each region within the endmember feature threshold θ, centered on a vertex of the triangle, is selected as a candidate region for feature pixels. The candidate regions for feature pixels include: candidate regions for photosynthetic vegetation, candidate regions for non-photosynthetic vegetation, and candidate regions for bare soil. The candidate regions for feature pixels are obtained based on the assumptions of the pixel triad model.

7. The remote sensing image land cover classification method based on index and coverage features as described in claim 6, characterized in that, Determining the clean pixel index threshold based on the remote sensing spectral image includes: The preprocessed remote sensing spectral image is subjected to MNF transformation to obtain an MNF image; The PPI value of each pixel in the MNF image is determined using the PPI algorithm; The clean pixel index threshold is determined based on the PPI values ​​of all pixels in the MNF image.

8. The remote sensing image land cover classification method based on index and coverage features as described in claim 5, characterized in that, Determining the endmember feature value based on the target feature pixel includes: Determine the NDVI and DFI corresponding to each pixel in the target feature pixels; Based on the NDVI and DFI corresponding to each pixel, the average NDVI and average DFI of the target feature pixel are determined respectively; the average NDVI and average DFI of the target feature pixel are used as the endmember feature value; the endmember feature value includes NDVI. PV NDVI NPV NDVI BS DFI PV DFI NPV DFI BS .

9. The remote sensing image land cover classification method based on index and coverage features as described in claim 8, characterized in that, Determining the coverage feature based on the endmember feature value includes: The endmember feature value is calculated according to the following formula: NDVI=∑f i V i =f PV NDVI PV +f NPV NDVI NPV +f BS NDVI BS (3) DFI=∑f i D i =f PV DFI PV +f NPV DFI NPV +f BS DFI BS (4) ∑f i =f PV +f NPV +f BS =1 (5) NDVI and DFI are index values ​​calculated based on remote sensing spectral images, f PV f NPV and f BS NDVI represents the coverage of light and vegetation, non-light and vegetation, and bare soil, respectively. PV and DFI PV These represent the NDVI index and DFI index of photosynthetic vegetation cover, respectively; NDVI NPV and DFI NPV These represent the NDVI and DFI indices of non-photosynthetic vegetation, respectively; NDVI BS and DFI BS C represents the NDVI index and DFI index of bare soil, respectively. x C y and C z This indicates the coverage of a certain end-member type where outliers occur.

10. A remote sensing image land cover classification device based on index and coverage features, characterized in that, include: The acquisition unit allows users to acquire remote sensing spectral images; The classification unit is used to input the remote sensing spectral image into the trained land cover classification model to classify the land cover on the Earth's surface. The land cover classification model includes: The backbone network is used to extract global and local features from the remotely sensed spectral image. The index feature extraction module is used to extract the index features corresponding to the spectral indices of the remote sensing spectral image; the spectral indices of the remote sensing spectral image include: NDVI and DFI; The first splicing module is used to splice the exponential features with the global features to obtain a first spliced ​​feature map; The first channel attention module is used to adjust the channel weights of the first spliced ​​feature map to obtain the first feature map after weight adjustment. A multi-scale feature extraction module is used to extract multi-scale features from the weight-adjusted first feature map to obtain a multi-scale feature map. The land cover feature extraction module is used to extract the land cover features corresponding to the land cover of the remote sensing spectral image; the land cover of the remote sensing spectral image includes: photosynthetic vegetation cover f. PV Non-photosynthetic vegetation cover f NPV With bare soil coverage f BS ; The second stitching module is used to stitch the coverage feature with the multi-scale feature map to obtain a second stitched feature map; The second channel attention module is used to adjust the channel weights of the second spliced ​​feature map to obtain the second feature map after weight adjustment. The first convolution module is used to convolve the local features to obtain the first convolutional features; The second convolution module is used to convolve the second feature map to obtain the second convolution feature; The third concatenation module is used to concatenate the first convolutional feature with the second convolutional feature to obtain the third concatenated feature map; The third convolution module is used to classify the third stitched feature map to obtain the specific type of land cover.

Images