Desert plant coverage monitoring method and system based on remote sensing technology

Abstract

The invention discloses a desert plant coverage monitoring method and system based on a remote sensing technology, and the method comprises the steps: obtaining original remote sensing image data, and carrying out the preliminary segmentation, and obtaining a preliminary segmentation image; extracting a multi-scale object set, enhancing vegetation features of the boundary fuzzy region, and obtaining an enhanced multi-scale object set; denoising the enhanced object set, and extracting a gray-level co-occurrence matrix by using a texture analysis method based on the denoised spectral feature set to obtain a texture feature enhanced set; extracting shape features from the texture feature set to obtain a standardized shape feature set; fusing the spectrum, texture and shape features to obtain a fused feature set; a training sample is extracted, and a vegetation coverage quantitative model is constructed and trained; and performing secondary correction on a model prediction result through a random forest algorithm to obtain a final vegetation coverage distribution map. According to the method, the precision and reliability of desert sparse vegetation coverage monitoring are effectively improved, and important technical support is provided for desert ecological environment monitoring and evaluation.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing technology and environmental monitoring technology, and in particular relates to a method and system for monitoring desert plant cover based on remote sensing technology. Background Art

[0002] Traditional methods for monitoring desert plant cover rely primarily on ground surveys and manual sampling. While these methods can provide relatively accurate data, they suffer from high costs, low efficiency, and limited coverage. These methods are particularly challenging to implement in large, complex desert areas, making them difficult to meet the demands of real-time, dynamic monitoring.

[0003] With the development of remote sensing technology, the use of remote sensing imagery for vegetation cover monitoring has gradually become an important tool. Remote sensing technology offers advantages such as wide coverage, short data acquisition cycles, and relatively low costs, effectively supporting large-scale, long-term vegetation monitoring. However, due to the sparse distribution of vegetation in desert areas, complex backgrounds, and significant influences from factors such as light and climate, vegetation information in remote sensing imagery often suffers from blurred boundaries and noise interference. This poses numerous challenges to traditional remote sensing analysis methods for desert plant cover monitoring.

[0004] Existing remote sensing vegetation monitoring methods typically analyze spectral features. However, because the spectral characteristics of desert vegetation are similar to those of background features (such as sand and bare soil), relying solely on spectral information makes it difficult to achieve high-precision vegetation extraction. Furthermore, desert vegetation is often fragmented and discontinuous in distribution. Traditional classification algorithms are prone to misclassification and omission when dealing with such complex scenes, affecting the accuracy of monitoring results.

[0005] To overcome these challenges, researchers have recently attempted to improve the accuracy of desert vegetation monitoring by combining multi-source remote sensing data with multi-feature analysis methods. However, these methods still have limitations in practical applications, such as low automation of feature extraction and insufficient model generalization, making them difficult to meet the requirements for high-precision monitoring in complex desert environments.

[0006] In summary, existing technologies for monitoring desert plant cover still face challenges such as insufficient feature extraction, insufficient model accuracy, and weak anti-interference capabilities. Therefore, there is an urgent need for an efficient and accurate desert plant cover monitoring method to meet monitoring needs in complex environments and provide a scientific basis for desertification prevention and ecological restoration. Summary of the Invention

[0007] In order to solve the above technical problems, the present invention proposes a method and system for monitoring desert plant cover based on remote sensing technology to solve the problems existing in the above-mentioned prior art.

[0008] To achieve the above object, the present invention provides a method for monitoring desert plant cover based on remote sensing technology, the method comprising:

[0009] Obtaining original remote sensing image data, performing preliminary segmentation on the original remote sensing image data through a preset adaptive scale filter to obtain a preliminary segmented image;

[0010] Extract a multi-scale object set from the preliminary segmented image, and use a gradient-based edge detection algorithm to enhance the vegetation features in the fuzzy boundary area to obtain the enhanced multi-scale object set.

[0011] Denoising and correcting the enhanced multi-scale object set to obtain a denoised spectral feature set;

[0012] Based on the denoised spectral feature set, the texture analysis method is used to obtain the gray-level co-occurrence matrix of each object, and the texture features are extracted according to the matrix results to obtain the texture feature enhancement set;

[0013] Extract shape features from the texture feature enhancement set, apply morphological operations to smooth the object boundary, calculate shape parameters based on the smoothed boundary contour, and obtain a standardized shape feature set;

[0014] Based on the standardized shape feature set, the spectrum, texture and shape features are fused by principal component analysis to obtain the fused feature set;

[0015] Extract training samples from the fusion feature set, use the support vector machine algorithm to build a vegetation cover quantification model, and train the vegetation cover quantification model based on the training samples to obtain a preliminary quantitative model.

[0016] The prediction results of the preliminary quantitative model were corrected twice using the random forest algorithm, and the model output was optimized based on the corrected prediction values to obtain the final vegetation coverage distribution map.

[0017] The sparse vegetation area data was extracted from the final vegetation coverage distribution map, and the spatial interpolation method was used to fill the missing information of the fragmented objects. The interpolation radius was adjusted according to the local noise interference, and the coverage distribution was updated according to the interpolation results to obtain the final monitoring results.

[0018] Optionally, extracting a multi-scale object set from the preliminary segmented image, and using a gradient-based edge detection algorithm to enhance vegetation features in fuzzy boundary areas to obtain an enhanced multi-scale object set includes:

[0019] Extract multi-scale object sets from the preliminary segmented image and use segmentation strategies at different scales to extract vegetation features in the image;

[0020] During the feature extraction process, the gradient threshold is dynamically adjusted according to the vegetation distribution density and image complexity to obtain a multi-scale object set;

[0021] Aiming at the vegetation features in the fuzzy boundary areas of the multi-scale object set, a gradient-based edge detection algorithm is used to detect the edges between vegetation and background by calculating the gradient amplitude and direction of pixels in the image, thereby obtaining the enhanced multi-scale object set.

[0022] Optionally, performing denoising and correction processing on the enhanced multi-scale object set to obtain a denoised spectral feature set includes:

[0023] The enhanced multi-scale object set is processed using a spectral feature separation algorithm to separate and remove the atmospheric scattering component to obtain a corrected spectral feature set.

[0024] Combined with the surface albedo correction model, the corrected spectral feature set is corrected for background interference to obtain the denoised spectral feature set.

[0025] Optionally, the method of obtaining a gray-level co-occurrence matrix of each object based on the denoised spectral feature set using a texture analysis method, extracting texture features according to the matrix results, and obtaining a texture feature enhancement set includes:

[0026] Convert the denoised spectral feature set into a grayscale image, preset the conversion direction and distance parameters, and calculate the gray-level co-occurrence matrix of each object;

[0027] For each object's gray level co-occurrence matrix, texture features are extracted to obtain a texture feature enhancement set; wherein the texture features include but are not limited to contrast, angular second moment, correlation, entropy and inverse difference moment.

[0028] Optionally, the step of fusing spectrum, texture, and shape features based on the standardized shape feature set by principal component analysis to obtain a fused feature set includes:

[0029] Integrate the standardized shape feature set with spectral and texture features to form a multi-dimensional feature matrix;

[0030] The principal component analysis method is used to calculate the covariance matrix of the multi-dimensional feature matrix, and solve the eigenvalues and eigenvectors. Several eigenvectors with the largest eigenvalues are selected as the principal component directions to achieve dimensionality reduction of the multi-dimensional feature matrix.

[0031] Construct a fusion feature set based on the feature matrix after dimensionality reduction.

[0032] Optionally, extracting training samples from the fused feature set, constructing a vegetation cover quantization model using a support vector machine algorithm, and training the vegetation cover quantization model based on the training samples to obtain a preliminary quantization model includes:

[0033] Extracting training samples from the fused feature set and introducing a synthetic minority class oversampling technique to increase the number of training samples; the training samples include spectral, texture, and shape features related to vegetation coverage;

[0034] Use support vector machine algorithm to build vegetation cover quantification model;

[0035] The vegetation cover quantification model was trained based on the training samples, using radial basis function as the kernel function, and the model parameters were optimized by grid search method combined with cross validation method to obtain a preliminary quantitative model.

[0036] The present invention also provides a desert plant cover monitoring system based on remote sensing technology, which is used to implement the method described above, comprising: an image segmentation module, a feature enhancement module, a feature denoising module, a texture extraction module, a shape extraction module, a feature fusion module, a model construction module, a model correction module and a plant monitoring module;

[0037] The image segmentation module is used to obtain original remote sensing image data, and perform preliminary segmentation on the original remote sensing image data through a preset adaptive scale filter to obtain a preliminary segmented image;

[0038] The feature enhancement module is used to extract a multi-scale object set from the preliminary segmented image, and enhance the vegetation features in the fuzzy boundary area using a gradient-based edge detection algorithm to obtain an enhanced multi-scale object set;

[0039] The feature denoising module is used to perform denoising and correction processing on the enhanced multi-scale object set to obtain a denoised spectral feature set;

[0040] The texture extraction module is used to obtain the gray level co-occurrence matrix of each object based on the denoised spectral feature set using a texture analysis method, extract texture features according to the matrix results, and obtain a texture feature enhancement set;

[0041] The shape extraction module is used to extract shape features from the texture feature enhancement set, apply morphological operations to smooth the object boundary, calculate shape parameters based on the smoothed boundary contour, and obtain a standardized shape feature set;

[0042] The feature fusion module is used to fuse spectrum, texture and shape features based on the standardized shape feature set through principal component analysis to obtain a fused feature set;

[0043] The model building module is used to extract training samples from the fusion feature set, use the support vector machine algorithm to build a vegetation cover quantization model, and train the vegetation cover quantization model based on the training samples to obtain a preliminary quantitative model;

[0044] The model correction module is used to perform secondary correction on the prediction results of the preliminary quantitative model using the random forest algorithm, optimize the model output according to the corrected prediction value, and obtain the final vegetation coverage distribution map;

[0045] The plant monitoring module extracts sparse vegetation area data from the final vegetation coverage distribution map, uses spatial interpolation to fill in the missing information of fragmented objects, adjusts the interpolation radius according to local noise interference, and updates the coverage distribution according to the interpolation results to obtain the final monitoring results.

[0046] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0047] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method when executed by a processor.

[0048] The present invention also provides a computer program product, comprising a computer program, which implements the steps of the method when executed by a processor.

[0049] Compared with the prior art, the present invention has the following advantages and technical effects:

[0050] This invention discloses a high-precision monitoring method for sparse vegetation coverage in deserts. This method addresses the sparse distribution, weak spatial regularity, and irregular morphology of vegetation in desert areas. It extracts vegetation objects through adaptive scale filtering and multi-scale segmentation techniques, and constructs a comprehensive feature set combining spectral, texture, and shape features. To address the issue of scarce samples, a synthetic minority class oversampling technique is introduced to generate supplementary samples. A support vector machine and random forest algorithm are used to construct a vegetation coverage quantification model. Finally, spatial interpolation technology is used to fill in missing information for fragmented objects, effectively improving the accuracy and reliability of sparse vegetation coverage monitoring in deserts and providing important technical support for desert ecological environment monitoring and assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0052] Figure 1 This is a flow chart of the overall monitoring method according to an embodiment of the present invention;

[0053] Figure 2 This is a flow chart of a method for generating a vegetation coverage distribution map according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0055] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0056] Example 1

[0057] like Figure 1-2 As shown, this embodiment provides a method for monitoring desert plant cover based on remote sensing technology, the method comprising:

[0058] Obtaining original remote sensing image data, performing preliminary segmentation on the original remote sensing image data through a preset adaptive scale filter to obtain a preliminary segmented image;

[0059] Extract a multi-scale object set from the preliminary segmented image, and use a gradient-based edge detection algorithm to enhance the vegetation features in the fuzzy boundary area to obtain the enhanced multi-scale object set.

[0060] Denoising and correcting the enhanced multi-scale object set to obtain a denoised spectral feature set;

[0061] Based on the denoised spectral feature set, the texture analysis method is used to obtain the gray-level co-occurrence matrix of each object, and the texture features are extracted according to the matrix results to obtain the texture feature enhancement set;

[0062] Extract shape features from the texture feature enhancement set, apply morphological operations to smooth the object boundary, calculate shape parameters based on the smoothed boundary contour, and obtain a standardized shape feature set;

[0063] Based on the standardized shape feature set, the spectrum, texture and shape features are fused by principal component analysis to obtain the fused feature set;

[0064] Extract training samples from the fusion feature set, use the support vector machine algorithm to build a vegetation cover quantification model, and train the vegetation cover quantification model based on the training samples to obtain a preliminary quantitative model.

[0065] The prediction results of the preliminary quantitative model were corrected twice using the random forest algorithm, and the model output was optimized based on the corrected prediction values to obtain the final vegetation coverage distribution map.

[0066] The sparse vegetation area data was extracted from the final vegetation coverage distribution map, and the spatial interpolation method was used to fill the missing information of the fragmented objects. The interpolation radius was adjusted according to the local noise interference, and the coverage distribution was updated according to the interpolation results to obtain the final monitoring results.

[0067] It is feasible to obtain the original remote sensing image data, which is the basis of the entire processing flow. Subsequently, the image is preliminarily segmented using a preset adaptive scale filter. This filter can dynamically adjust the scale according to the characteristics of the image to adapt to the sparse distribution and significant spatial heterogeneity of vegetation in desert areas. In view of the special distribution characteristics of desert vegetation, the filter parameters are further adjusted to optimize the ability to recognize the spectral differences between vegetation and background. By analyzing the spectral differences between vegetation and background, the initial object boundary is determined, thereby achieving accurate division of vegetation areas. Finally, after the above processing, a preliminary segmented image is obtained, providing high-quality input data for subsequent vegetation feature enhancement and analysis. By combining the adaptive scale filter with spectral difference analysis, this method effectively solves the segmentation problem caused by the uneven distribution of desert vegetation and improves the accuracy and efficiency of remote sensing image processing.

[0068] One feasible approach is to acquire multispectral imagery covering desert areas using satellite remote sensing platforms, such as Landsat or Sentinel-2. This data typically covers visible and near-infrared bands, reflecting the basic spectral characteristics of vegetation and the background. For example, in the desert regions of Inner Mongolia, imagery from summer can be selected, when the spectral signal of vegetation is relatively strong and more distinct from the sandy background. This not only provides a reliable foundation for subsequent segmentation but also effectively reduces data interference caused by seasonal variations.

[0069] Subsequently, the image is preliminarily segmented using a preset adaptive scale filter. This technology is key to processing desert vegetation distribution. In principle, the adaptive scale filter dynamically adjusts the size of the filter window based on the spatial and spectral characteristics of the pixels in the image. Specifically, in areas with sparse vegetation, the filter may use a smaller scale, such as 3×3 pixels, to capture small patches of vegetation; while in sandy areas with a more uniform background, the scale may be expanded to 7×7 pixels to smooth noise and highlight the main features. For example, in remote sensing images of the edge of the Taklamakan Desert, vegetation may be scattered in the form of bushes. The adaptive filter will automatically reduce the scale to 5×5 pixels based on local spatial heterogeneity, thereby retaining the boundary information of these tiny vegetation. The advantage of this method is that it improves the flexibility of segmentation and avoids the problem of insufficient recognition of sparse vegetation by traditional fixed-scale filters.

[0070] In view of the special distribution characteristics of desert vegetation, further adjusting the filter parameters is an important step to optimize the segmentation effect. Specifically, the weight parameters of the filter can be adjusted by analyzing the spectral reflectance characteristics of the vegetation, such as high reflectance in the near-infrared band and low reflectance in the visible light band. For example, when processing images of the Gobi region, the weight of the near-infrared band can be set to 0.6 and the visible light band to 0.4 to enhance the contrast between the vegetation and the background. It should be noted that this adjustment needs to be verified in combination with field sample data, such as vegetation distribution points collected by drones to determine the applicability of the parameters. Doing so can significantly improve segmentation accuracy, especially in extremely arid areas where vegetation coverage is less than 10%.

[0071] The method of extracting a multi-scale object set from the preliminary segmented image and using a gradient-based edge detection algorithm to enhance vegetation features in fuzzy boundary areas to obtain an enhanced multi-scale object set may include:

[0072] A multi-scale object set is extracted from the preliminary segmented image, and the vegetation features in the image are extracted using segmentation strategies at different scales. During the feature extraction process, the gradient threshold is dynamically adjusted according to the vegetation distribution density and image complexity to obtain a multi-scale object set. For the vegetation features in the fuzzy boundary areas of the multi-scale object set, a gradient-based edge detection algorithm is used to detect the edges between the vegetation and the background by calculating the gradient amplitude and direction of the pixels in the image, thereby obtaining an enhanced multi-scale object set.

[0073] As an implementable approach, extracting a multi-scale object set from the preliminarily segmented image is to make full use of the segmentation strategies at different scales and capture the geographical features in the image from micro to macro. The core of this method is that the distribution of desert vegetation has significant spatial heterogeneity, and it is difficult for a single scale to fully reflect its characteristics. As an implementation method, multiple scale windows, such as 5 meters, 15 meters, and 30 meters, can be set in the preliminarily segmented image of the Alxa region in Inner Mongolia to extract the outlines of small shrubs, medium vegetation patches, and larger oases, respectively. Specifically, at the 5-meter scale, scattered camel thorn distribution can be captured, and its area is usually between 20 and 50 square meters, while the 30-meter scale can outline the edge of the oasis, which may cover an area of several hectares. The advantage of this multi-scale extraction is that it provides rich structural information for subsequent classification and analysis, avoiding the defects of detail loss or over-generalization at a single scale.

[0074] To address the difficulty of achieving uniform gradient thresholds in multi-scale segmentation, dynamic adjustment of the gradient threshold becomes a key step. For example, in the area surrounding Dunhuang, Gansu, vegetation density is low, and image complexity primarily stems from sand dune texture. In contrast, in the Yanchi region of Ningxia, vegetation density is slightly higher and accompanied by microtopographic variations. As an implementation, the threshold can be adaptively adjusted based on local vegetation density and image texture characteristics. Specifically, in the Dunhuang region, the gradient threshold can be set to 0.06 in sparsely vegetated areas to capture subtle boundaries, while in the densely vegetated Yanchi region, it can be adjusted to 0.1 to avoid misjudging sand texture. It should be noted that this dynamic adjustment requires the use of NDVI values to aid judgment. If the NDVI in a particular area exceeds 0.25, the threshold is appropriately raised to highlight vegetation boundaries. The benefit of this strategy is that it ensures consistent boundary extraction at different scales while adapting to the spatially varying characteristics of desert vegetation. From multiple perspectives, the extraction of multi-scale object sets lays the foundation for subsequent analysis.

[0075] For vegetation features in areas with blurred boundaries, a gradient-based edge detection algorithm is used to significantly enhance segmentation accuracy. For example, at the edge of the Taklimakan Desert in Xinjiang, the boundary between vegetation and sand is often blurred due to dust interference. When using the Sobel algorithm, areas with significant changes in reflectivity can be detected by calculating the grayscale gradient amplitude and direction of pixels in the image. Specifically, assuming that the near-infrared reflectivity of a strip of vegetation is 0.3 and that of sand is 0.15, the gradient amplitude may reach above 0.1 at the boundary, while the internal uniform area is close to 0. By setting an amplitude threshold, such as marking areas greater than 0.08 as edges, the outline of vegetation can be clearly outlined.

[0076] The denoising and correction processing of the enhanced multi-scale object set to obtain the denoised spectral feature set may be performed, including:

[0077] The enhanced multi-scale object set is processed using a spectral feature separation algorithm to separate and remove the atmospheric scattering components to obtain a corrected spectral feature set. Combined with the surface albedo correction model, the corrected spectral feature set is corrected for background interference to obtain a denoised spectral feature set.

[0078] As a specific implementation, a spectral feature separation algorithm is used to process the enhanced multi-scale object set, effectively separating atmospheric scattering components and improving data purity. As an implementation, the FLAASH model simulates atmospheric transmission processes and combines image radiometric values with sensor parameters to estimate the interference of atmospheric scattering on spectral signals. For example, in 30-meter resolution imagery from the desert region of Inner Mongolia, high atmospheric water vapor content in summer can cause significant deviations in the shortwave infrared band. The FLAASH model calculates the contribution of the scattering component based on local meteorological data, such as water vapor content and aerosol optical depth, and removes this effect from the spectral data. Furthermore, when processing vegetation imagery from the Yanchi region of Ningxia, atmospheric scattering can obscure the spectral characteristics of vegetation due to frequent sandstorms. After FLAASH processing, the spectral reflectance peak of vegetation in the near-infrared band increases from 0.3 to 0.45, more closely resembling the true characteristics of the surface. This method has significant technical benefits and can provide more reliable basic data for subsequent spectral analysis.

[0079] Subsequently, combining the surface albedo correction model to correct background interference in the spectral data is the key to further improving data quality. Specifically, albedo correction requires parameter adjustment based on the surface type and environmental conditions. For example, in the area around Dunhuang, Gansu, the surface is mainly sandy with a high albedo, usually around 0.25, while the albedo of the vegetated area is lower, around 0.1. During correction, the albedo parameters can be dynamically adjusted based on the surface classification results to optimize the spectral signals of the sand and vegetation respectively.

[0080] After the FLAASH model separates the scattered components, albedo correction further removes surface background interference, while the recalculation of spectral features integrates the results of the first two steps to form high-quality, denoised data. In the desert areas of Inner Mongolia, the spectral feature set processed through the complete process can reduce the error in vegetation cover estimation from 15% to 8%, providing more accurate support for ecological monitoring. In the Yanchi area of Ningxia, the recognition rate of vegetation boundaries has increased from 75% to 90%, contributing to the refined management of desertification control. The progressive logic and diversity of this methodology ensure the applicability of the solution in different environments, while significantly improving the robustness and practical value of the data.

[0081] The method can be implemented by using a texture analysis method to obtain a gray level co-occurrence matrix of each object based on the denoised spectral feature set, extracting texture features according to the matrix results, and obtaining a texture feature enhancement set, including:

[0082] The denoised spectral feature set is converted into a grayscale image, the conversion direction and distance parameters are preset, and the grayscale co-occurrence matrix of each object is calculated; the texture features are extracted from the grayscale co-occurrence matrix of each object to obtain a texture feature enhancement set; wherein the texture features include but are not limited to contrast, angular second moment, correlation, entropy and inverse difference moment.

[0083] As a specific implementation, the denoised spectral feature set lays a solid foundation for subsequent analysis. This dataset, through a spectral denoising algorithm, removes the effects of atmospheric scattering and background interference, preserving the true spectral characteristics of the object. Texture analysis becomes a key tool for further mining spatial information from such data, and the gray-level co-occurrence matrix (GLCM), a classic statistical tool, effectively captures the spatial distribution patterns of grayscale values within an image. As an implementation, the spectral feature data is first converted into a grayscale image. This process can be based on the reflectance of a specific band, such as the near-infrared band. Assuming that the corrected near-infrared reflectance in vegetation images from the desert region of Inner Mongolia is stable at around 0.42, it is converted to a grayscale value range of 0-255 using a linear mapping to generate a grayscale image. This conversion preserves the spatial details of the spectral information and provides reliable input for GLCM calculations. Specifically, GLCM calculations require the determination of window size, direction, and distance parameters. Because vegetation in desertified areas is sparsely distributed and has weak spatial regularity, a fixed window size may not be able to adapt to object features of varying scales. For example, in vegetation patches in the Yanchi region of Ningxia, using a small 3×3 pixel window may miss large-scale texture information; while using a window that is too large, such as 15×15 pixels, may introduce excessive background noise. To this end, the window size can be dynamically adjusted, for example, using a 5×5 window for small vegetation patches and a 9×9 window for larger sand textures. At the same time, the choice of directional parameters is also crucial, usually including four directions: 0°, 45°, 90°, and 135°, to fully capture the spatial anisotropy of the texture. In sandy land images around Dunhuang, Gansu, the GLCM in the 0° direction shows the horizontal regularity of the dune texture, while the 45° direction highlights the oblique features formed by wind erosion. This multi-directional analysis significantly improves the comprehensiveness of texture description.

[0084] Furthermore, texture features extracted from the GLCM include contrast, angular second moment, correlation, entropy, and inverse difference moment. Each feature reflects different properties of texture. For example, contrast measures the magnitude of local variations in grayscale values. In vegetation strips at the edge of the Taklamakan Desert in Xinjiang, the contrast of unprocessed images can be low, only around 0.8. However, after denoising and GLCM processing, the contrast increases to 1.2, indicating a clearer boundary between vegetation and sand. The angular second moment (energy) reflects texture uniformity. In vegetation images of Yanchi, Ningxia, the energy value increases from 0.25 to 0.35, indicating an enhanced regularity in texture distribution after correction. Correlation describes the linear dependence between grayscale values. In the desert regions of Inner Mongolia, the correlation increases from 0.6 to 0.75, indicating greater consistency in the denoised data. Entropy reflects texture complexity. In images severely disturbed by dust, the entropy value decreases from 3.5 to 3.2, indicating that the texture complexity decreases after noise reduction, closer to the true state. The inverse moment measures local uniformity and increases from 0.4 to 0.55 in small vegetation patches, highlighting the smooth nature of the vegetation area.

[0085] It's important to note that the extraction of these features relies not only on GLCM calculation but also on prior denoising. For example, in images of sandy land in Dunhuang, Gansu, directly calculating the GLCM without using the FLAASH model to remove atmospheric scattering can result in low contrast and blurred vegetation boundaries. However, when combined with a denoised spectral feature set, the signal-to-noise ratio of the texture features is significantly improved, boosting the discrimination between vegetation and sandy land from 70% to 85%.

[0086] As an extension, multi-band data can be combined to further enrich texture features. For example, by generating grayscale images using visible and shortwave infrared bands, calculating GLCMs, and fusing the feature sets, the fused texture feature set improved the accuracy of vegetation cover estimation from 80% to 90% on the desert edge of Xinjiang, providing more precise support for ecological monitoring. This approach, through the synergistic effect of spectral denoising and texture analysis, produces a high-quality enhanced texture feature set. For example, in the desert regions of Inner Mongolia, texture feature enhancement reduced the vegetation classification error rate from 12% to 5%, significantly improving the reliability of subsequent analysis. In the Yanchi region of Ningxia, the boundary recognition rate increased from 75% to 90%, providing a more refined data foundation for desertification control. This progressive logic ensures the solution's applicability to sparse vegetation and complex surface environments, while also providing greater robustness and practical value.

[0087] It is feasible to extract shape features from the texture feature enhancement set, apply morphological operations to smooth the object boundaries, adjust the operation kernel size according to the irregularity of the growth morphology, calculate the shape parameters according to the smoothed boundary contour, and obtain a standardized shape feature set.

[0088] As a feasible approach, when extracting shape information for each object from a texture feature enhancement set, directly extracted boundaries often exhibit noise and discontinuities due to the sparse distribution and irregular growth patterns of vegetation in desertified areas. To address this issue, morphological operations have become a key tool for smoothing boundaries. Morphological operations adjust image shape through the sliding operation of structuring elements. For example, dilation fills boundary gaps, while erosion removes small protrusions. As an implementation, in vegetation images from the desert regions of Inner Mongolia, if the original boundary appears jagged due to dust interference, a 3×3 pixel circular structuring element can be used to open the boundary to remove noise points, followed by a 5×5 structuring element to close the broken edges. This processing improves boundary smoothness from 60% to 85%, laying the foundation for subsequent shape analysis. Specifically, the size of the structuring element needs to be dynamically adjusted based on the object's morphology. In small vegetation patches in Yanchi, Ningxia, where the boundary is rich in detail but noisy, a small 3×3 structuring element can preserve detail while removing isolated noise points, increasing boundary continuity from 70% to 90%. In the large-scale sand dune textures of Dunhuang, Gansu, the boundaries vary gently but the irregularities span a wide range. Using a 7×7 structuring element effectively smooths the edges, reduces redundant protrusions, and increases boundary regularity from 65% to 88%. This adaptive adjustment ensures that the shape characteristics of objects at different scales are accurately captured.

[0089] Furthermore, the smoothed boundary contour is used to calculate shape parameters. For example, area reflects the size of the object, perimeter measures the length of the boundary, compactness describes the regularity of the shape through the ratio of area to perimeter, and the shape factor further quantifies the complexity of the morphology. In the vegetation strip on the edge of the Taklimakan Desert in Xinjiang, the area before smoothing was 1200 pixels, the perimeter was 180 pixels, the compactness was 0.46, and the shape factor was low at 0.62, indicating an irregular boundary. After smoothing, the area increased slightly to 1250 pixels, the perimeter decreased to 160 pixels, the compactness increased to 0.61, and the shape factor increased to 0.75, indicating a more regular morphology. This quantitative description provides a reliable basis for subsequent classification.

[0090] It should be noted that scale differences between objects can lead to insufficient contrast in shape parameters, making standardization essential. As an implementation method, Z-score standardization transforms parameters into a distribution with a mean of 0 and a standard deviation of 1. For example, in the desert regions of Inner Mongolia, the areas of multiple vegetation patches range from 800 to 2000 pixels, and their perimeters from 140 to 220 pixels. After standardization, the relative differences between these parameters are more distinct, allowing the shape characteristics of small and large patches to be directly compared. In Yanchi, Ningxia, min-max standardization maps compactness to a range of 0-1. Standardization results in a more even distribution of compactness values from 0.4-0.7, improving the accuracy of the classification model from 82% to 91%. For example, in sandy land images from Dunhuang, Gansu, unstandardized shape parameters lead to a 15% confusion rate due to scale differences. After standardization, the distinction between the compactness of small vegetation and sand dunes increased from 0.2 to 0.5, reducing the confusion rate to 4%, significantly improving classification reliability. Specifically, for smaller vegetation patches, the standardized shape factor increased from 0.6 to 0.8, reflecting their regularity after smoothing, while the shape factor of sand dunes remained around 0.5, highlighting their morphological differences. This processing not only improves data consistency but also provides more accurate shape features for subsequent ecological monitoring.

[0091] The method of fusing spectrum, texture and shape features based on the standardized shape feature set by principal component analysis to obtain a fused feature set may include:

[0092] The standardized shape feature set is integrated with the spectral and texture features to form a multi-dimensional feature matrix. The principal component analysis method is used to calculate the covariance matrix of the multi-dimensional feature matrix, and the eigenvalues and eigenvectors are solved. Several eigenvectors with the largest eigenvalues are selected as the principal component directions to achieve dimensionality reduction of the multi-dimensional feature matrix. The fusion feature set is constructed based on the feature matrix after dimensionality reduction.

[0093] As a feasible approach, a standardized shape feature set is integrated with spectral and texture features. The resulting multidimensional feature matrix can have dimensions of up to 20 or more, encompassing shape, spectral, and texture information. Specifically, the high dimensionality of a multidimensional feature matrix increases computational complexity, necessitating dimensionality reduction. Principal component analysis (PCA) is an effective dimensionality reduction method. It calculates the covariance matrix of the feature matrix and identifies the direction of maximum data variance. The eigenvalues reflect the contribution of each principal component, while the eigenvectors define the projection direction. For example, in an image analysis of vegetation and sand in Yanchi, Ningxia, the original feature matrix contained 15-dimensional data, including shape parameters, spectral reflectance, and texture contrast. After applying PCA, the eigenvalues of the first three principal components were 8.5, 3.2, and 1.8, respectively, with a cumulative contribution rate of 90%, indicating that these three directions retain the vast majority of the original information. Projecting the data into a low-dimensional space reduces the data from 15 dimensions to 3, significantly reducing computational complexity while preserving key features. For example, the compactness and near-infrared reflectance of vegetation patches have higher weights in the first principal component, reflecting the strong correlation between their morphology and spectrum, while the texture contrast of sand is more prominent in the second principal component. This dimensionality reduction method not only reduces complexity but also highlights the characteristic differences between different landforms.

[0094] Furthermore, when selecting the number of principal components, it is necessary to balance information retention and dimensionality reduction effects based on the size of the eigenvalues. As an implementation method, in the dune images of Dunhuang, Gansu, the eigenvalues are 9.0, 2.5, 1.0, and 0.5 from large to small, with a total variance of 13. If the first two principal components are selected, the contribution rate is about 85%, which is suitable for fast classification tasks; if the first three are retained, the contribution rate rises to 95%, which is more suitable for fine ecological monitoring. Specifically, the first two principal components may miss the shape factor details of small vegetation, while the third principal component supplements this information, reducing the classification confusion rate from 10% to 5%. This flexible adjustment ensures the representativeness of the eigenvector after dimensionality reduction. For example, at the edge of the Taklimakan Desert in Xinjiang, the 3D eigenvector after dimensionality reduction clearly distinguishes between vegetation strips and sandy land. The comprehensive eigenvalues of vegetation are concentrated around 1.2, while those of sandy land are around -0.8, and the discrimination is improved from 0.3 to 1.0.

[0095] It's important to note that the reduced-dimensionality integrated feature set integrates information from multiple sources, improving the efficiency of subsequent tasks. For example, in estimating vegetation cover in Inner Mongolia, the unreduced feature matrix took 15 seconds to process and achieved a classification accuracy of 88%. After dimensionality reduction, processing time was reduced to 5 seconds, and accuracy rose to 92%. Specifically, the integrated feature set combines shape compactness with spectral reflectance, resulting in a more concentrated distribution of feature values in vegetation areas, greater dispersion of sandy areas, and increased boundary recognition from 85% to 94%. As an extension, incorporating texture contrast to further enrich feature dimensions, for example, in Yanchi, Ningxia, the integrated feature set increased the discrimination between small vegetation patches and sandy areas from 0.4 to 0.7, and the F1 score of the classification model from 0.85 to 0.93. This multi-dimensional synergy significantly enhances the robustness of the solution in complex environments. From multiple perspectives, spectral features highlight feature type, shape features quantify morphological regularity, and texture features characterize surface roughness. Fusion of these three features through PCA dimensionality reduction reduces redundancy while preserving core information. For example, in Dunhuang, Gansu, the overlap between small vegetation and sand dunes reached 12% before dimensionality reduction, but dropped to 3% afterward, providing more accurate data support for desertification monitoring. This method significantly improves the practical value and technical effectiveness of image analysis through a progressive logic of feature integration, dimensionality reduction, and optimization.

[0096] The method of extracting training samples from the fusion feature set, constructing a vegetation cover quantization model using a support vector machine algorithm, and training the vegetation cover quantization model based on the training samples to obtain a preliminary quantization model may include:

[0097] Training samples are extracted from the fused feature set, and a synthetic minority class oversampling technique is introduced to increase the number of training samples; the training samples include spectral, texture, and shape features related to vegetation cover; a vegetation cover quantization model is constructed using a support vector machine algorithm; the vegetation cover quantization model is trained based on the training samples, using radial basis functions as kernel functions, and the model parameters are optimized through a grid search method combined with a cross-validation method to obtain a preliminary quantitative model.

[0098] As an implementable approach, extracting training samples from a fused feature set is the basis for building a quantitative model of vegetation cover. The fused feature set contains spectral, texture, and shape features, which can comprehensively characterize the characteristics of vegetation-covered areas. For example, in remote sensing imagery of desertified areas in Inner Mongolia, training samples may include vegetation areas with a near-infrared reflectance of 0.4, a texture contrast of 0.8, and a shape compactness of 0.61, and corresponding values of 0.1, 0.3, and 0.45 for sandy areas. These samples are converted into three-dimensional feature vectors through the aforementioned dimensionality reduction process, retaining key information.

[0099] However, due to the low vegetation cover in desertified areas, the number of positive samples is often far less than the number of negative samples. For example, in a 1000×1000 pixel image, vegetation accounts for only 10%, or approximately 100 valid samples, while sand samples account for as many as 900. This data imbalance causes the model to tend to predict the majority class and ignore vegetation areas. As an implementation, the synthetic minority oversampling technique (SMOTE) effectively alleviates this problem by generating new samples through interpolation. It works by finding k nearest neighbors between minority class samples and randomly interpolating along the connecting lines to generate synthetic samples. For example, in a vegetation image of Yanchi, Ningxia, the eigenvector of a positive sample is [1.2, 0.5, 0.3], and its nearest neighbors are [1.1, 0.6, 0.4]. SMOTE can generate new samples such as [1.15, 0.55, 0.35]. Specifically, if the original number of positive samples is 100, SMOTE can increase that to 300. The new samples retain the characteristic distribution characteristics of vegetation, such as the near-infrared reflectance concentration around 0.4. It is important to note that SMOTE does not simply copy samples, but rather enhances data diversity through linear interpolation, thereby improving the model's ability to identify sparse vegetation. For example, in Dunhuang sand dune imagery in Gansu Province, the application of SMOTE increased the recall rate of vegetation classification from 70% to 85%, significantly reducing false negatives.

[0100] Furthermore, using support vector machines (SVMs) to construct a quantitative model is the subsequent core step. The advantage of SVM is that it finds the optimal hyperplane that distinguishes vegetation from non-vegetation by maximizing the classification interval, and is particularly suitable for nonlinear data. The radial basis function (RBF) kernel function processes the complex relationship between features by mapping the data into a high-dimensional space. For example, on the edge of the Taklimakan Desert in Xinjiang, the 3D feature vector distributions of vegetation and sand have a high degree of overlap, making it difficult to distinguish them by linear classification. The RBF kernel function introduces nonlinear mapping, enabling the model to capture the coordinated changes in compactness and reflectivity. As an implementation method, grid search is combined with cross-validation to optimize SVM parameters to ensure model performance. For example, in the estimation of vegetation cover in Inner Mongolia, the penalty factor C controls the complexity of the model and is set in the range of 0.1 to 10. The kernel parameter γ affects the smoothness of the decision boundary and is set in the range of 0.01 to 1. Through 5-fold cross-validation, it was determined that when C = 1 and γ = 0.1, the model had the highest F1 score of 0.92 on the validation set. Specifically, when C is too high, such as 10, the model overfits, and the proportion of sand misclassified as vegetation increases to 8%. When γ is too small, such as 0.01, the boundary is too smooth, and the misclassification rate increases to 12%. The optimized parameters balance accuracy and generalization, and the processing time is only 3 seconds, making it suitable for real-time monitoring.

[0101] It's important to note that the combination of SMOTE and SVM significantly improves quantization accuracy in scenarios with few positive samples. For example, in Yanchi, Ningxia, an SVM model trained on raw data achieved an 80% accuracy rate and a 15% confusion rate for estimating vegetation cover. After introducing SMOTE, the number of positive samples increased to a balanced level, raising the model's accuracy to 93% and reducing the confusion rate to 5%.

[0102] Furthermore, if texture contrast extension features are combined, such as incorporating contrast ratios from 0.8 for vegetation and 0.3 for sand into training, the model's recognition rate for small vegetation patches increases from 75% to 88%. This multi-dimensional synergy makes the model more robust in complex desertification environments. From multiple perspectives, SMOTE solves data imbalance, SVM provides strong classification capabilities, and parameter optimization ensures practicality. For example, in Dunhuang, Gansu, the fusion feature set after dimensionality reduction combined with this solution reduced the processing time for vegetation cover estimation from 10 seconds to 4 seconds, and reduced the coverage estimation error from 8% to 3%. This method provides efficient support for ecological monitoring, especially in desertified areas where positive samples are scarce, significantly improving technical effectiveness and application value.

[0103] It is feasible to obtain a preliminary quantitative model, perform secondary correction on the model prediction results through the random forest algorithm, adjust the feature weights according to the background feature dependency problem, optimize the model output according to the corrected prediction value, and obtain the final vegetation cover distribution map.

[0104] As an implementable approach, the preliminary quantitative model is based on a support vector machine (SVM), whose prediction results depend to a certain extent on the distribution of input features. For example, in remote sensing imagery of desertified areas in Inner Mongolia, vegetation cover predictions can be affected by the presence of sandy background features, leading the model to misclassify edge areas as non-vegetation. Specifically, when the near-infrared reflectance of a region is 0.35 and the texture contrast is 0.5, the SVM may predict it as a negative sample due to the proximity of the feature values of the background sand, resulting in an accuracy of only 75%. This background dependence stems from the limitations of the SVM's linear or nonlinear partitioning of the feature space in complex scenarios. To address this issue, a random forest algorithm is introduced for secondary correction. By integrating multiple decision trees, random forests can capture complex nonlinear relationships between features and exhibit strong robustness to noise. For example, in images of sand dunes in Dunhuang, Gansu, random forests, through a voting mechanism, reduced the preliminary model's misclassification rate from 15% to 6%, significantly improving prediction stability. As an implementation method, random forests specifically optimize for background feature dependence during correction. Feature importance ranking is a key step. By calculating the contribution of each feature to the classification, the weights are dynamically adjusted. For example, in vegetation images of Yanchi, Ningxia, the importance of near-infrared reflectance for vegetation identification is 0.6, while the shape compactness of the background sand is only 0.2. After adjustment, the model reduces its reliance on compactness and increases the weight of reflectance, increasing the prediction recall rate of vegetation areas from 78% to 90%. Specifically, in one experiment, the original feature vector [0.4, 0.8, 0.61] was misclassified as sand by the preliminary model. After adjusting the weights, the random forest correctly identified it as vegetation. This method effectively reduces background interference and improves the model's sensitivity to key features.

[0105] Furthermore, the corrected predictions from the random forest model can be used to optimize model output. This correction process not only corrects biases in the initial results but also smooths out prediction noise through a multi-tree ensemble. For example, on the edge of the Taklamakan Desert in Xinjiang, the initial model predicted a vegetation cover of 25%, but the actual value was 35%. The random forest model, based on an ensemble of 50 decision trees, corrected the prediction to 33%, reducing the bias from 10% to 2%.

[0106] It should be noted that this correction is not a simple averaging; rather, it incorporates feature weights and inter-tree variance to ensure that the results are closer to the true distribution. The corrected predictions also exhibit improved spatial continuity, avoiding the patchy discontinuities common in initial models and laying the foundation for subsequent distribution map generation. Generating the final vegetation cover distribution map is the core objective of the entire process. Based on the corrected predictions, the map more accurately reflects spatial heterogeneity. For example, in a 500×500 pixel image in Inner Mongolia, the initial model-generated distribution map indicated a vegetation cover of 12%, but this map contained numerous undercounts. After correction, the coverage was adjusted to 15%, highly consistent with the field measurement of 14.8%. Specifically, the random forest algorithm enhances the recognition of small vegetation patches, making vegetation boundaries clearer in the edge areas of the distribution map and reducing the obfuscation of the sandy background. This high-precision distribution map provides a reliable basis for ecological monitoring. For example, in vegetation resource management, managers can use the distribution map to precisely locate degraded areas and optimize vegetation restoration strategies. The introduction of random forests significantly improves the robustness and practicality of the model in many ways.

[0107] It is feasible to extract sparse vegetation area data from the final vegetation coverage distribution map, use spatial interpolation technology to fill the missing information of fragmented objects, adjust the interpolation radius according to local noise interference, and update the coverage distribution according to the interpolation results to obtain high-precision monitoring results.

[0108] As a feasible approach, extracting data on sparse vegetation areas from the final vegetation cover distribution map serves as the basis for subsequent analysis. Sparse vegetation areas often appear missing or discontinuous in the data due to their low coverage and fragmented distribution. For example, in a 500×500 pixel image of a desertified region in Inner Mongolia, areas with vegetation cover below 10% may account for only 5% of the total area, and the distribution is spotty, with some pixels being missed due to interference from the sandy background.

[0109] To solve this problem, spatial interpolation technology has become a key means. Spatial interpolation fills in missing information and improves data integrity by utilizing the spatial correlation of existing data. As an implementation method, Kriging interpolation is often used to deal with such scenarios because of its optimal unbiased estimation characteristics based on statistics. Specifically, on the edge of the Taklimakan Desert in Xinjiang, the coverage data points of a sparsely vegetated area are sparsely distributed. The original distribution map shows a coverage rate of 8%, but the field measurement is 11%. Through Kriging interpolation, combined with the near-infrared reflectivity of 0.3 and the texture contrast of 0.4 of the neighboring pixels, the coverage rate is adjusted to 10.5% after interpolation, which is closer to the actual value. This method takes advantage of spatial autocorrelation to ensure that the interpolation results reflect the true distribution trend.

[0110] Furthermore, inverse distance weighted interpolation is another optional solution that is suitable for quickly processing large-scale data. For example, in the dune images of Dunhuang, Gansu, the pixel spacing in the sparse vegetation area is large, and some areas are missing coverage data due to noise interference. Inverse distance weighted interpolation fills these gaps by giving higher weights to close points. Specifically, the coverage of the four known points in the vicinity of a pixel point is 5%, 7%, 6%, and 8%, respectively, and the distances are 50 meters, 70 meters, 100 meters, and 120 meters, respectively. After interpolation, the coverage of the point is estimated to be 6.8%, which is close to the field verification value of 7%.

[0111] It's important to note that the choice of interpolation radius directly impacts the accuracy of the results. If the radius is too large, the data may be oversmoothed, obscuring the fragmented characteristics of sparse vegetation; if it is too small, it may fail to effectively fill in missing areas. Dynamic adjustment of the interpolation radius has become an optimization tool to address local noise interference. For example, in vegetation imagery of Yanchi, Ningxia, texture noise from the sandy background caused some sparse vegetation to be misclassified as non-vegetation, resulting in an underestimation of the predicted coverage. By analyzing local characteristics, such as the standard deviation of near-infrared reflectance of 0.1 and the mean texture contrast of 0.3, the interpolation radius was dynamically adjusted from the default 100 meters to 50 meters, reducing the impact of noisy pixels. After this adjustment, the coverage of one area increased from 6% to 9%, consistent with the field measurement value of 8.8%. This dynamic adjustment balances accuracy and noise suppression, improving data reliability.

[0112] Furthermore, high-precision distribution maps are of great value in ecological and environmental monitoring. For example, in vegetation resource management, managers can use updated distribution maps to identify degraded areas with coverage below 10% and prioritize restoration measures. In a monitoring project in Xinjiang, the interpolated distribution map increased the identification rate of sparse vegetation from 75% to 88%, providing a basis for developing precise management plans.

[0113] It should be noted that the application of spatial interpolation can also bring additional technical effects. For example, when processing large-scale images, the interpolated data can be used as the training basis for subsequent machine learning models to improve prediction stability. In a 5000×5000 pixel image in Inner Mongolia, the interpolated coverage data was used to train a random forest model, and the model's recall rate for sparse vegetation increased from 80% to 93%. This multi-dimensional collaborative optimization not only improves data quality, but also provides more efficient support for ecological monitoring. From multiple aspects, whether it is the statistical accuracy of Kriging interpolation or the computational efficiency of inverse distance weighting, combined with dynamic radius adjustment, they all support the generation of high-precision distribution maps, providing a reliable tool for vegetation management in desertified areas.

[0114] Example 2

[0115] This embodiment also provides a remote sensing-based desert plant cover monitoring system for implementing the method, comprising: an image segmentation module, a feature enhancement module, a feature denoising module, a texture extraction module, a shape extraction module, a feature fusion module, a model building module, a model correction module, and a plant monitoring module;

[0116] The image segmentation module is used to obtain original remote sensing image data, and perform preliminary segmentation on the original remote sensing image data through a preset adaptive scale filter to obtain a preliminary segmented image;

[0117] The feature enhancement module is used to extract a multi-scale object set from the preliminary segmented image, and enhance the vegetation features in the fuzzy boundary area using a gradient-based edge detection algorithm to obtain an enhanced multi-scale object set;

[0118] The feature denoising module is used to perform denoising and correction processing on the enhanced multi-scale object set to obtain a denoised spectral feature set;

[0119] The texture extraction module is used to obtain the gray level co-occurrence matrix of each object based on the denoised spectral feature set using a texture analysis method, extract texture features according to the matrix results, and obtain a texture feature enhancement set;

[0120] The shape extraction module is used to extract shape features from the texture feature enhancement set, apply morphological operations to smooth the object boundary, calculate shape parameters based on the smoothed boundary contour, and obtain a standardized shape feature set;

[0121] The feature fusion module is used to fuse spectrum, texture and shape features based on the standardized shape feature set through principal component analysis to obtain a fused feature set;

[0122] The model building module is used to extract training samples from the fusion feature set, use the support vector machine algorithm to build a vegetation cover quantization model, and train the vegetation cover quantization model based on the training samples to obtain a preliminary quantitative model;

[0123] The model correction module is used to perform secondary correction on the prediction results of the preliminary quantitative model using the random forest algorithm, optimize the model output according to the corrected prediction value, and obtain the final vegetation coverage distribution map;

[0124] The plant monitoring module extracts sparse vegetation area data from the final vegetation coverage distribution map, uses spatial interpolation to fill in the missing information of fragmented objects, adjusts the interpolation radius according to local noise interference, and updates the coverage distribution according to the interpolation results to obtain the final monitoring results.

[0125] Example 3

[0126] This embodiment further provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0127] Example 4

[0128] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method when executed by a processor.

[0129] Example 5

[0130] This embodiment also provides a computer program product, including a computer program, which implements the steps of the method when executed by a processor.

[0131] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for monitoring desert plant cover based on remote sensing technology, characterized in that: The method comprises: Obtaining original remote sensing image data, performing preliminary segmentation on the original remote sensing image data through a preset adaptive scale filter to obtain a preliminary segmented image; Extract a multi-scale object set from the preliminary segmented image, and use a gradient-based edge detection algorithm to enhance the vegetation features in the fuzzy boundary area to obtain the enhanced multi-scale object set. Denoising and correcting the enhanced multi-scale object set to obtain a denoised spectral feature set; Based on the denoised spectral feature set, the texture analysis method is used to obtain the gray-level co-occurrence matrix of each object, and the texture features are extracted according to the matrix results to obtain the texture feature enhancement set; Extract shape features from the texture feature enhancement set, apply morphological operations to smooth the object boundary, calculate shape parameters based on the smoothed boundary contour, and obtain a standardized shape feature set; Based on the standardized shape feature set, the spectrum, texture and shape features are fused by principal component analysis to obtain the fused feature set; Extract training samples from the fusion feature set, use the support vector machine algorithm to build a vegetation cover quantification model, and train the vegetation cover quantification model based on the training samples to obtain a preliminary quantitative model. The prediction results of the preliminary quantitative model were corrected twice using the random forest algorithm, and the model output was optimized based on the corrected prediction values to obtain the final vegetation coverage distribution map. The sparse vegetation area data was extracted from the final vegetation coverage distribution map, and the spatial interpolation method was used to fill the missing information of the fragmented objects. The interpolation radius was adjusted according to the local noise interference, and the coverage distribution was updated according to the interpolation results to obtain the final monitoring results.

2. The method according to claim 1, characterized in that The method of extracting a multi-scale object set from the preliminary segmented image and using a gradient-based edge detection algorithm to enhance vegetation features in fuzzy boundary areas to obtain an enhanced multi-scale object set includes: Extract multi-scale object sets from the preliminary segmented image and use segmentation strategies at different scales to extract vegetation features in the image; During the feature extraction process, the gradient threshold is dynamically adjusted according to the vegetation distribution density and image complexity to obtain a multi-scale object set; Aiming at the vegetation features in the fuzzy boundary areas of the multi-scale object set, a gradient-based edge detection algorithm is used to detect the edges between vegetation and background by calculating the gradient amplitude and direction of pixels in the image, thereby obtaining the enhanced multi-scale object set.

3. The method according to claim 1, characterized in that The denoising and correction processing is performed on the enhanced multi-scale object set to obtain a denoised spectral feature set, including: The enhanced multi-scale object set is processed using a spectral feature separation algorithm to separate and remove the atmospheric scattering component to obtain a corrected spectral feature set. Combined with the surface albedo correction model, the corrected spectral feature set is corrected for background interference to obtain the denoised spectral feature set.

4. The method according to claim 1, wherein The method of using a texture analysis method to obtain a gray-level co-occurrence matrix of each object based on the denoised spectral feature set, extracting texture features according to the matrix results, and obtaining a texture feature enhancement set includes: Convert the denoised spectral feature set into a grayscale image, preset the conversion direction and distance parameters, and calculate the gray-level co-occurrence matrix of each object; For each object's gray level co-occurrence matrix, texture features are extracted to obtain a texture feature enhancement set; wherein the texture features include but are not limited to contrast, angular second moment, correlation, entropy and inverse difference moment.

5. The method according to claim 1, wherein The method of fusing spectrum, texture and shape features based on the standardized shape feature set by principal component analysis to obtain a fused feature set includes: Integrate the standardized shape feature set with spectral and texture features to form a multi-dimensional feature matrix; The principal component analysis method is used to calculate the covariance matrix of the multi-dimensional feature matrix, and solve the eigenvalues and eigenvectors. Several eigenvectors with the largest eigenvalues are selected as the principal component directions to achieve dimensionality reduction of the multi-dimensional feature matrix. Construct a fusion feature set based on the feature matrix after dimensionality reduction.

6. The method according to claim 1, wherein The method extracts training samples from the fusion feature set, constructs a vegetation cover quantization model using a support vector machine algorithm, and trains the vegetation cover quantization model based on the training samples to obtain a preliminary quantization model, including: Extracting training samples from the fused feature set and introducing a synthetic minority class oversampling technique to increase the number of training samples; the training samples include spectral, texture, and shape features related to vegetation coverage; Use support vector machine algorithm to build vegetation cover quantification model; The vegetation cover quantification model was trained based on the training samples. The radial basis function was used as the kernel function. The model parameters were optimized by the grid search method combined with the cross-validation method to obtain a preliminary quantitative model.

7. A desert plant cover monitoring system based on remote sensing technology, characterized in that: Used to implement the method according to any one of claims 1 to 6, comprising: an image segmentation module, a feature enhancement module, a feature denoising module, a texture extraction module, a shape extraction module, a feature fusion module, a model building module, a model correction module and a plant monitoring module; The image segmentation module is used to obtain original remote sensing image data, and perform preliminary segmentation on the original remote sensing image data through a preset adaptive scale filter to obtain a preliminary segmented image; The feature enhancement module is used to extract a multi-scale object set from the preliminary segmented image, and enhance the vegetation features in the fuzzy boundary area using a gradient-based edge detection algorithm to obtain an enhanced multi-scale object set; The feature denoising module is used to perform denoising and correction processing on the enhanced multi-scale object set to obtain a denoised spectral feature set; The texture extraction module is used to obtain the gray level co-occurrence matrix of each object based on the denoised spectral feature set using a texture analysis method, extract texture features according to the matrix results, and obtain a texture feature enhancement set; The shape extraction module is used to extract shape features from the texture feature enhancement set, apply morphological operations to smooth the object boundary, calculate shape parameters based on the smoothed boundary contour, and obtain a standardized shape feature set; The feature fusion module is used to fuse spectrum, texture and shape features based on the standardized shape feature set through principal component analysis to obtain a fused feature set; The model building module is used to extract training samples from the fusion feature set, use the support vector machine algorithm to build a vegetation cover quantization model, and train the vegetation cover quantization model based on the training samples to obtain a preliminary quantitative model; The model correction module is used to perform secondary correction on the prediction results of the preliminary quantitative model using the random forest algorithm, optimize the model output according to the corrected prediction value, and obtain the final vegetation coverage distribution map; The plant monitoring module extracts sparse vegetation area data from the final vegetation coverage distribution map, uses spatial interpolation to fill in the missing information of fragmented objects, adjusts the interpolation radius according to local noise interference, and updates the coverage distribution according to the interpolation results to obtain the final monitoring results.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images